RAG using LangChain : Part 1-Document Loaders

Photo by Paul Frenzel on Unsplash

Welcome to a new series of articles on LangChain and LLMs. In this series, we will be learning about RAG in LLMs. Specifically in this article, we will be looking into Document Loaders in RAG. Let’s begin!

Beyond the Hype: Addressing the Challenges of LLMs

We know that Large Language Models are often trained on huge corpus of data. This type of training makes them very efficient in terms of generating content very effectively. But there are also some challenges that exists with these models.

• Limited Knowledge Base- The data that the model is trained upon might not be the latest data. It can contain outdated information, factual errors or lack of context. There can also be some potential biases in the data.

The trained model itself may have some potential issues like:

• Black Box Nature -The output generated by a LLM is sometimes difficult to understand which makes it difficult to verify the source or accuracy of the generated information.
• Hallucinations/Generation Bias -LLMs can sometimes fabricate information i.e. generate information which is incorrect or misleading. This can be due to their limited understanding or incomplete or biased data.

Now, for a domain specific task, we can fine tune any pre-trained LLM and make it suitable for our task. But that itself is a computationally expensive task. Imagine the time and resource needed to retrain a massive LLM!

This is where RAG comes into picture!

RAG(Retrieval Augmented Generation)

RAG is a technique that improves the capabilities of LLMs by combining them with external data sources. There are three key functionalities of RAG.

1. Retrieval: This is where the most relevant information is identified(from the external data). In this article, we will mostly be covering this series.
2. Augmentation: This step involves preprocessing and analyzing the retrieved information to make it suitable for the LLM using Summarization, Fact Checking and Formatting.
3. Generation: With the retrieved and augmented information at end, the LLM generates better in context responses.

The Retrieval part has multiple components. We will be looking into the load part.

Document Loaders

Document loaders are tools that play a crucial role in data ingestion. They take in raw data from different sources and convert them into a structured format called “Documents”. These documents contain the document content as well as the associated metadata like source and timestamps. Let’s look into the different types of document loaders.

CSV Loader

The CSV loader loads a CSV file with a single row per document. The output contains both page content as well as the metadata associated.

from langchain.document_loaders import CSVLoader
loader = CSVLoader(file_path="../datasets/sns_datasets/titanic.csv")
data = loader.load()
print(data[0])

page_content='survived: 0\npclass: 3\nsex: male\nage: 22.0\nsibsp: 1\nparch: 0\nfare: 7.25\nembarked: S\nclass: Third\nwho: man\nadult_male: True\ndeck: \nembark_town: Southampton\nalive: no\nalone: False' metadata={'source': '../datasets/sns_datasets/titanic.csv', 'row': 0}

HTML Loader

We can load HTML documents in a document format that we can use for further downstream tasks. We have similar syntax.

from langchain.document_loaders import UnstructuredHTMLLoader
loader = UnstructuredHTMLLoader('../datasets/harry_potter_html/001.htm')
data = loader.load()
data

[Document(page_content='A Day of Very Low Probability\n\nBeneath the moonlight glints a tiny fragment of silver, a fraction of a line…\n\n(black robes, falling)\n\n…blood spills out in litres, and someone screams a word.\n\nEvery inch of wall space is covered by a bookcase. Each bookcase has six shelves, going almost to the ceiling. Some bookshelves are stacked to the brim with hardback books: science, maths, history, and everything else. Other shelves have two layers of paperback science fiction, with the back layer of books propped up on old tissue boxes or lengths of wood, so that you can see the back layer of books above the books in front. And it still isn’t enough. Books are overflowing onto the tables and the sofas and making little heaps under the windows.\n\nThis is the living-room of the house occupied by the eminent Professor Michael Verres-Evans, and his wife, Mrs. Petunia Evans-Verres, and their adopted son, Harry James Potter-Evans-Verres.\n\nThere is a letter lying on the living-room table, and an unstamped envelope of yellowish parchment, addressed to Mr. H. Potter in emerald-green ink.\n\nThe Professor and his wife are speaking sharply at each other, but they are not shouting. The Professor considers shouting to be uncivilised.\n\n“You’re joking,” Michael said to Petunia. His tone indicated that he was very much afraid that she was serious.\n\n“My sister was a witch,” Petunia repeated. She looked frightened, but stood her ground. “Her husband was a wizard.”\n\n“This is absurd!” Michael said sharply. “They were at our wedding—they visited for Christmas—”\n\n“I told them you weren’t to know,” Petunia whispered. “But it’s true. I’ve seen things—”\n\nThe Professor rolled his eyes. “Dear, I understand that you’re not familiar with the sceptical literature. You may not realise how easy it is for a trained magician to fake the seemingly impossible. Remember how I taught Harry to bend spoons? If it seemed like they could always guess what you were thinking, that’s called cold reading—”\n\n“It wasn’t bending spoons—”\n\n“What was it, then?”\n\nPetunia bit her lip. “I can’t just tell you. You’ll think I’m—” She swallowed. “Listen. Michael. I wasn’t—always like this—” She gestured at herself, as though to indicate her lithe form. “Lily did this. Because I—because I begged her. For years, I begged her. Lily had always been prettier than me, and I’d… been mean to her, because of that, and then she got magic, can you imagine how I felt? And I begged her to use some of that magic on me so that I could be pretty too, even if I couldn’t have her magic, at least I could be pretty.”\n\nTears were gathering in Petunia’s eyes.\n\n“And Lily would tell me no, and make up the most ridiculous excuses, like the world would end if she were nice to her sister, or a centaur told her not to—the most ridiculous things, and I hated her for it. And when I had just graduated from university, I was going out with this boy, Vernon Dursley, he was fat and he was the only boy who would talk to me. And he said he wanted children, and that his first son would be named Dudley. And I thought to myself, what kind of parent names their child Dudley Dursley? It was like I saw my whole future life stretching out in front of me, and I couldn’t stand it. And I wrote to my sister and told her that if she didn’t help me I’d rather just—”\n\nPetunia stopped.\n\n“Anyway,” Petunia said, her voice small, “she gave in. She told me it was dangerous, and I said I didn’t care any more, and I drank this potion and I was sick for weeks, but when I got better my skin cleared up and I finally filled out and… I was beautiful, people were nice to me,” her voice broke, “and after that I couldn’t hate my sister any more, especially when I learned what her magic brought her in the end—”\n\n“Darling,” Michael said gently, “you got sick, you gained some weight while resting in bed, and your skin cleared up on its own. Or being sick made you change your diet—”\n\n“She was a witch,” Petunia repeated. “I saw it.”\n\n“Petunia,” Michael said. The annoyance was creeping into his voice. “You know that can’t be true. Do I really have to explain why?”\n\nPetunia wrung her hands. She seemed to be on the verge of tears. “My love, I know I can’t win arguments with you, but please, you have to trust me on this—”\n\n“Dad! Mum!”\n\nThe two of them stopped and looked at Harry as though they’d forgotten there was a third person in the room.\n\nHarry took a deep breath. “Mum, your parents didn’t have magic, did they?”\n\n“No,” Petunia said, looking puzzled.\n\n“Then no one in your family knew about magic when Lily got her letter. How did they get convinced?”\n\n“Ah…” Petunia said. “They didn’t just send a letter. They sent a professor from Hogwarts. He—” Petunia’s eyes flicked to Michael. “He showed us some magic.”\n\n“Then you don’t have to fight over this,” Harry said firmly. Hoping against hope that this time, just this once, they would listen to him. “If it’s true, we can just get a Hogwarts professor here and see the magic for ourselves, and Dad will admit that it’s true. And if not, then Mum will admit that it’s false. That’s what the experimental method is for, so that we don’t have to resolve things just by arguing.”\n\nThe Professor turned and looked down at him, dismissive as usual. “Oh, come now, Harry. Really, magic? I thought you’d know better than to take this seriously, son, even if you’re only ten. Magic is just about the most unscientific thing there is!”\n\nHarry’s mouth twisted bitterly. He was treated well, probably better than most genetic fathers treated their own children. Harry had been sent to the best primary schools—and when that didn’t work out, he was provided with tutors from the endless pool of starving students. Always Harry had been encouraged to study whatever caught his attention, bought all the books that caught his fancy, sponsored in whatever maths or science competitions he entered. He was given anything reasonable that he wanted, except, maybe, the slightest shred of respect. A Doctor teaching biochemistry at Oxford could hardly be expected to listen to the advice of a little boy. You would listen to Show Interest, of course; that’s what a Good Parent would do, and so, if you conceived of yourself as a Good Parent, you would do it. But take a ten-year-old seriously? Hardly.\n\nSometimes Harry wanted to scream at his father.\n\n“Mum,” Harry said. “If you want to win this argument with Dad, look in chapter two of the first book of the Feynman Lectures on Physics. There’s a quote there about how philosophers say a great deal about what science absolutely requires, and it is all wrong, because the only rule in science is that the final arbiter is observation—that you just have to look at the world and report what you see. Um… off the top of my head I can’t think of where to find something about how it’s an ideal of science to settle things by experiment instead of arguments—”\n\nHis mother looked down at him and smiled. “Thank you, Harry. But—” her head rose back up to stare at her husband. “I don’t want to win an argument with your father. I want my husband to, to listen to his wife who loves him, and trust her just this once—”\n\nHarry closed his eyes briefly. Hopeless. Both of his parents were just hopeless.\n\nNow his parents were getting into one of those arguments again, one where his mother tried to make his father feel guilty, and his father tried to make his mother feel stupid.\n\n“I’m going to go to my room,” Harry announced. His voice trembled a little. “Please try not to fight too much about this, Mum, Dad, we’ll know soon enough how it comes out, right?”\n\n“Of course, Harry,” said his father, and his mother gave him a reassuring kiss, and then they went on fighting while Harry climbed the stairs to his bedroom.\n\nHe shut the door behind him and tried to think.\n\nThe funny thing was, he should have agreed with Dad. No one had ever seen any evidence of magic, and according to Mum, there was a whole magical world out there. How could anyone keep something like that a secret? More magic? That seemed like a rather suspicious sort of excuse.\n\nIt should have been a clean case for Mum joking, lying or being insane, in ascending order of awfulness. If Mum had sent the letter herself, that would explain how it arrived at the letterbox without a stamp. A little insanity was far, far less improbable than the universe really working like that.\n\nExcept that some part of Harry was utterly convinced that magic was real, and had been since the instant he saw the putative letter from the Hogwarts School of Witchcraft and Wizardry.\n\nHarry rubbed his forehead, grimacing. Don’t believe everything you think, one of his books had said.\n\nBut this bizarre certainty… Harry was finding himself just expecting that, yes, a Hogwarts professor would show up and wave a wand and magic would come out. The strange certainty was making no effort to guard itself against falsification—wasn’t making excuses in advance for why there wouldn’t be a professor, or the professor would only be able to bend spoons.\n\nWhere do you come from, strange little prediction? Harry directed the thought at his brain. Why do I believe what I believe?\n\nUsually Harry was pretty good at answering that question, but in this particular case, he had no clue what his brain was thinking.\n\nHarry mentally shrugged. A flat metal plate on a door affords pushing, and a handle on a door affords pulling, and the thing to do with a testable hypothesis is to go and test it.\n\nHe took a piece of lined paper from his desk, and started writing.\n\nDear Deputy Headmistress\n\nHarry paused, reflecting; then discarded the paper for another, tapping another millimetre of graphite from his mechanical pencil. This called for careful calligraphy.\n\nDear Deputy Headmistress Minerva McGonagall,\n\nOr Whomsoever It May Concern:\n\nI recently received your letter of acceptance to Hogwarts, addressed to Mr. H. Potter. You may not be aware that my genetic parents, James Potter and Lily Potter (formerly Lily Evans) are dead. I was adopted by Lily’s sister, Petunia Evans-Verres, and her husband, Michael Verres-Evans.\n\nI am extremely interested in attending Hogwarts, conditional on such a place actually existing. Only my mother Petunia says she knows about magic, and she can’t use it herself. My father is highly sceptical. I myself am uncertain. I also don’t know where to obtain any of the books or equipment listed in your acceptance letter.\n\nMother mentioned that you sent a Hogwarts representative to Lily Potter (then Lily Evans) in order to demonstrate to her family that magic was real, and, I presume, help Lily obtain her school materials. If you could do this for my own family it would be extremely helpful.\n\nSincerely,\n\nHarry James Potter-Evans-Verres.\n\nHarry added their current address, then folded up the letter and put it in an envelope, which he addressed to Hogwarts. Further consideration led him to obtain a candle and drip wax onto the flap of the envelope, into which, using a penknife’s tip, he impressed the initials H.J.P.E.V. If he was going to descend into this madness, he was going to do it with style.\n\nThen he opened his door and went back downstairs. His father was sitting in the living-room and reading a book of higher maths to show how smart he was; and his mother was in the kitchen preparing one of his father’s favourite meals to show how loving she was. It didn’t look like they were talking to one another at all. As scary as arguments could be, not arguing was somehow much worse.\n\n“Mum,” Harry said into the unnerving silence, “I’m going to test the hypothesis. According to your theory, how do I send an owl to Hogwarts?”\n\nHis mother turned from the kitchen sink to stare at him, looking shocked. “I—I don’t know, I think you just have to own a magic owl.”\n\nThat should’ve sounded highly suspicious, oh, so there’s no way to test your theory then, but the peculiar certainty in Harry seemed willing to stick its neck out even further.\n\n“Well, the letter got here somehow,” Harry said, “so I’ll just wave it around outside and call ‘letter for Hogwarts!’ and see if an owl picks it up. Dad, do you want to come and watch?”\n\nHis father shook his head minutely and kept on reading. Of course, Harry thought to himself. Magic was a disgraceful thing that only stupid people believed in; if his father went so far as to test the hypothesis, or even watch it being tested, that would feel like associating himself with that…\n\nOnly as Harry stumped out the back door, into the back garden, did it occur to him that if an owl did come down and snatch the letter, he was going to have some trouble telling Dad about it.\n\nBut—well—that can’t really happen, can it? No matter what my brain seems to believe. If an owl really comes down and grabs this envelope, I’m going to have worries a lot more important than what Dad thinks.\n\nHarry took a deep breath, and raised the envelope into the air.\n\nHe swallowed.\n\nCalling out Letter for Hogwarts! while holding an envelope high in the air in the middle of your own back garden was… actually pretty embarrassing, now that he thought about it.\n\nNo. I’m better than Dad. I will use the scientific method even if it makes me feel stupid.\n\n“Letter—” Harry said, but it actually came out as more of a whispered croak.\n\nHarry steeled his will, and shouted into the empty sky, “Letter for Hogwarts! Can I get an owl?”\n\n“Harry?” asked a bemused woman’s voice, one of the neighbours.\n\nHarry pulled down his hand like it was on fire and hid the envelope behind his back like it was drug money. His whole face was hot with shame.\n\nAn old woman’s face peered out from above the neighbouring fence, grizzled grey hair escaping from her hairnet. Mrs. Figg, the occasional babysitter. “What are you doing, Harry?”\n\n“Nothing,” Harry said in a strangled voice. “Just—testing a really silly theory—”\n\n“Did you get your acceptance letter from Hogwarts?”\n\nHarry froze in place.\n\n“Yes,” Harry’s lips said a little while later. “I got a letter from Hogwarts. They say they want my owl by the 31st of July, but—”\n\n“But you don’t have an owl. Poor dear! I can’t imagine what someone must have been thinking, sending you just the standard letter.”\n\nA wrinkled arm stretched out over the fence, and opened an expectant hand. Hardly even thinking at this point, Harry gave over his envelope.\n\n“Just leave it to me, dear,” said Mrs. Figg, “and in a jiffy or two I’ll have someone over.”\n\nAnd her face disappeared from over the fence.\n\nThere was a long silence in the garden.\n\nThen a boy’s voice said, calmly and quietly, “What.”', metadata={'source': '../datasets/harry_potter_html/001.htm'})]

We can also load HTML documents with BeautifulSoup using theBSHTMLLoader.

from langchain.document_loaders import BSHTMLLoader
loader = BSHTMLLoader('../datasets/harry_potter_html/001.htm')
data = loader.load()
data

Markdown Loader

from langchain.document_loaders import UnstructuredMarkdownLoader
loader = UnstructuredMarkdownLoader(file_path='../datasets/harry_potter_md/001.md')
data = loader.load()
print(data[0].page_content)

A Day of Very Low Probability

Beneath the moonlight glints a tiny fragment of silver, a fraction of a line…

(black robes, falling)

…blood spills out in litres, and someone screams a word.

Every inch of wall space is covered by a bookcase. Each bookcase has six shelves, going almost to the ceiling. Some bookshelves are stacked to the brim with hardback books: science, maths, history, and everything else. Other shelves have two layers of paperback science fiction, with the back layer of books propped up on old tissue boxes or lengths of wood, so that you can see the back layer of books above the books in front. And it still isn’t enough. Books are overflowing onto the tables and the sofas and making little heaps under the windows....

PDF Loader

Here we use PyPDF load the PDF documents. Each document contains the page content and metadata with page numbers.

from langchain.document_loaders import PyPDFLoader
loader = PyPDFLoader('../datasets/harry_potter_pdf/hpmor-trade-classic.pdf')

data = loader.load()

print(data[0].page_content)

Harry Potter and the Methods of Rationality

Wikipedia Loader

We use WikipediaLoader to fetch the documents. It mainly has 3 arguments: query: which is used to find the topic in Wikipedia, optional lang: default=“en”, and load_max_docs: default=100, which is used to limit the number of documents downloaded.

from langchain.document_loaders import WikipediaLoader
loader = WikipediaLoader(query='LangChain', load_max_docs=1)
data = loader.load()

data

[Document(page_content='LangChain is a framework designed to simplify the creation of applications using large language models (LLMs). As a language model integration framework, LangChain\'s use-cases largely overlap with those of language models in general, including document analysis and summarization, chatbots, and code analysis.\n\n\n== History ==\nLangChain was launched in October 2022 as an open source project by Harrison Chase, while working at machine learning startup Robust Intelligence. The project quickly garnered popularity, with improvements from hundreds of contributors on GitHub, trending discussions on Twitter, lively activity on the project\'s Discord server, many YouTube tutorials, and meetups in San Francisco and London. In April 2023, LangChain had incorporated and the new startup raised over $20 million in funding at a valuation of at least $200 million from venture firm Sequoia Capital, a week after announcing a $10 million seed investment from Benchmark.In October 2023 LangChain introduced LangServe, a deployment tool designed to facilitate the transition from LCEL (LangChain Expression Language) prototypes to production-ready applications.\n\n\n== Capabilities ==\nLangChain\'s developers highlight the framework\'s applicability to use-cases including chatbots, retrieval-augmented generation,  document summarization, and synthetic data generation.As of March 2023, LangChain included integrations with systems including Amazon, Google, and Microsoft Azure cloud storage; API wrappers for news, movie information, and weather; Bash for summarization, syntax and semantics checking, and execution of shell scripts; multiple web scraping subsystems and templates; few-shot learning prompt generation support; finding and summarizing "todo" tasks in code; Google Drive documents, spreadsheets, and presentations summarization, extraction, and creation; Google Search and Microsoft Bing web search; OpenAI, Anthropic, and Hugging Face language models; iFixit repair guides and wikis search and summarization; MapReduce for question answering, combining documents, and question generation; N-gram overlap scoring; PyPDF, pdfminer, fitz, and pymupdf for PDF file text extraction and manipulation; Python and JavaScript code generation, analysis, and debugging; Milvus vector database to store and retrieve vector embeddings; Weaviate vector database to cache embedding and data objects; Redis cache database storage; Python RequestsWrapper and other methods for API requests; SQL and NoSQL databases including JSON support; Streamlit, including for logging; text mapping for k-nearest neighbors search; time zone conversion and calendar operations; tracing and recording stack symbols in threaded and asynchronous subprocess runs; and the Wolfram Alpha website and SDK. As of April 2023, it can read from more than 50 document types and data sources.\n\n\n== References ==\n\n\n== External links ==\n\nOfficial website\nDiscord server support hub\nLangchain-ai on GitHub', metadata={'title': 'LangChain', 'summary': "LangChain is a framework designed to simplify the creation of applications using large language models (LLMs). As a language model integration framework, LangChain's use-cases largely overlap with those of language models in general, including document analysis and summarization, chatbots, and code analysis.\n\n", 'source': 'https://en.wikipedia.org/wiki/LangChain'})]

To extract the metadata.

data[0].metadata

{'title': 'LangChain',
 'summary': "LangChain is a framework designed to simplify the creation of applications using large language models (LLMs). As a language model integration framework, LangChain's use-cases largely overlap with those of language models in general, including document analysis and summarization, chatbots, and code analysis.\n\n",
 'source': 'https://en.wikipedia.org/wiki/LangChain'}

We can use this dictionary to extract specific information from the data.

ArXiv Loader

arXiv is an open-access archive for 2 million scholarly articles. We can use ArxivLoader to extract information of any paper. We need the article id that would be available in the URL of the paper to use the loader.

from langchain_community.document_loaders import ArxivLoader
loader = ArxivLoader(query='1706.03762', load_max_docs=1) 
data = loader.load()
data

[Document(page_content='Provided proper attribution is provided, Google hereby grants permission to\nreproduce the tables and figures in this paper solely for use in journalistic or\nscholarly works.\nAttention Is All You Need\nAshish Vaswani∗\nGoogle Brain\navaswani@google.com\nNoam Shazeer∗\nGoogle Brain\nnoam@google.com\nNiki Parmar∗\nGoogle Research\nnikip@google.com\nJakob Uszkoreit∗\nGoogle Research\nusz@google.com\nLlion Jones∗\nGoogle Research\nllion@google.com\nAidan N. Gomez∗ †\nUniversity of Toronto\naidan@cs.toronto.edu\nŁukasz Kaiser∗\nGoogle Brain\nlukaszkaiser@google.com\nIllia Polosukhin∗ ‡\nillia.polosukhin@gmail.com\nAbstract\nThe dominant sequence transduction models are based on complex recurrent or\nconvolutional neural networks that include an encoder and a decoder. The best\nperforming models also connect the encoder and decoder through an attention\nmechanism. We propose a new simple network architecture, the Transformer,\nbased solely on attention mechanisms, dispensing with recurrence and convolutions\nentirely. Experiments on two machine translation tasks show these models to\nbe superior in quality while being more parallelizable and requiring significantly\nless time to train. Our model achieves 28.4 BLEU on the WMT 2014 English-\nto-German translation task, improving over the existing best results, including\nensembles, by over 2 BLEU. On the WMT 2014 English-to-French translation task,\nour model establishes a new single-model state-of-the-art BLEU score of 41.8 after\ntraining for 3.5 days on eight GPUs, a small fraction of the training costs of the\nbest models from the literature. We show that the Transformer generalizes well to\nother tasks by applying it successfully to English constituency parsing both with\nlarge and limited training data.\n∗Equal contribution. Listing order is random. Jakob proposed replacing RNNs with self-attention and started\nthe effort to evaluate this idea. Ashish, with Illia, designed and implemented the first Transformer models and\nhas been crucially involved in every aspect of this work. Noam proposed scaled dot-product attention, multi-head\nattention and the parameter-free position representation and became the other person involved in nearly every\ndetail. Niki designed, implemented, tuned and evaluated countless model variants in our original codebase and\ntensor2tensor. Llion also experimented with novel model variants, was responsible for our initial codebase, and\nefficient inference and visualizations. Lukasz and Aidan spent countless long days designing various parts of and\nimplementing tensor2tensor, replacing our earlier codebase, greatly improving results and massively accelerating\nour research.\n†Work performed while at Google Brain.\n‡Work performed while at Google Research.\n31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA.\narXiv:1706.03762v7  [cs.CL]  2 Aug 2023\n1\nIntroduction\nRecurrent neural networks, long short-term memory [13] and gated recurrent [7] neural networks\nin particular, have been firmly established as state of the art approaches in sequence modeling and\ntransduction problems such as language modeling and machine translation [35, 2, 5]. Numerous\nefforts have since continued to push the boundaries of recurrent language models and encoder-decoder\narchitectures [38, 24, 15].\nRecurrent models typically factor computation along the symbol positions of the input and output\nsequences. Aligning the positions to steps in computation time, they generate a sequence of hidden\nstates ht, as a function of the previous hidden state ht−1 and the input for position t. This inherently\nsequential nature precludes parallelization within training examples, which becomes critical at longer\nsequence lengths, as memory constraints limit batching across examples. Recent work has achieved\nsignificant improvements in computational efficiency through factorization tricks [21] and conditional\ncomputation [32], while also improving model performance in case of the latter. The fundamental\nconstraint of sequential computation, however, remains.\nAttention mechanisms have become an integral part of compelling sequence modeling and transduc-\ntion models in various tasks, allowing modeling of dependencies without regard to their distance in\nthe input or output sequences [2, 19]. In all but a few cases [27], however, such attention mechanisms\nare used in conjunction with a recurrent network.\nIn this work we propose the Transformer, a model architecture eschewing recurrence and instead\nrelying entirely on an attention mechanism to draw global dependencies between input and output.\nThe Transformer allows for significantly more parallelization and can reach a new state of the art in\ntranslation quality after being trained for as little as twelve hours on eight P100 GPUs.\n2\nBackground\nThe goal of reducing sequential computation also forms the foundation of the Extended Neural GPU\n[16], ByteNet [18] and ConvS2S [9], all of which use convolutional neural networks as basic building\nblock, computing hidden representations in parallel for all input and output positions. In these models,\nthe number of operations required to relate signals from two arbitrary input or output positions grows\nin the distance between positions, linearly for ConvS2S and logarithmically for ByteNet. This makes\nit more difficult to learn dependencies between distant positions [12]. In the Transformer this is\nreduced to a constant number of operations, albeit at the cost of reduced effective resolution due\nto averaging attention-weighted positions, an effect we counteract with Multi-Head Attention as\ndescribed in section 3.2.\nSelf-attention, sometimes called intra-attention is an attention mechanism relating different positions\nof a single sequence in order to compute a representation of the sequence. Self-attention has been\nused successfully in a variety of tasks including reading comprehension, abstractive summarization,\ntextual entailment and learning task-independent sentence representations [4, 27, 28, 22].\nEnd-to-end memory networks are based on a recurrent attention mechanism instead of sequence-\naligned recurrence and have been shown to perform well on simple-language question answering and\nlanguage modeling tasks [34].\nTo the best of our knowledge, however, the Transformer is the first transduction model relying\nentirely on self-attention to compute representations of its input and output without using sequence-\naligned RNNs or convolution. In the following sections, we will describe the Transformer, motivate\nself-attention and discuss its advantages over models such as [17, 18] and [9].\n3\nModel Architecture\nMost competitive neural sequence transduction models have an encoder-decoder structure [5, 2, 35].\nHere, the encoder maps an input sequence of symbol representations (x1, ..., xn) to a sequence\nof continuous representations z = (z1, ..., zn). Given z, the decoder then generates an output\nsequence (y1, ..., ym) of symbols one element at a time. At each step the model is auto-regressive\n[10], consuming the previously generated symbols as additional input when generating the next.\n2\nFigure 1: The Transformer - model architecture.\nThe Transformer follows this overall architecture using stacked self-attention and point-wise, fully\nconnected layers for both the encoder and decoder, shown in the left and right halves of Figure 1,\nrespectively.\n3.1\nEncoder and Decoder Stacks\nEncoder:\nThe encoder is composed of a stack of N = 6 identical layers. Each layer has two\nsub-layers. The first is a multi-head self-attention mechanism, and the second is a simple, position-\nwise fully connected feed-forward network. We employ a residual connection [11] around each of\nthe two sub-layers, followed by layer normalization [1]. That is, the output of each sub-layer is\nLayerNorm(x + Sublayer(x)), where Sublayer(x) is the function implemented by the sub-layer\nitself. To facilitate these residual connections, all sub-layers in the model, as well as the embedding\nlayers, produce outputs of dimension dmodel = 512.\nDecoder:\nThe decoder is also composed of a stack of N = 6 identical layers. In addition to the two\nsub-layers in each encoder layer, the decoder inserts a third sub-layer, which performs multi-head\nattention over the output of the encoder stack. Similar to the encoder, we employ residual connections\naround each of the sub-layers, followed by layer normalization. We also modify the self-attention\nsub-layer in the decoder stack to prevent positions from attending to subsequent positions. This\nmasking, combined with fact that the output embeddings are offset by one position, ensures that the\npredictions for position i can depend only on the known outputs at positions less than i.\n3.2\nAttention\nAn attention function can be described as mapping a query and a set of key-value pairs to an output,\nwhere the query, keys, values, and output are all vectors. The output is computed as a weighted sum\n3\nScaled Dot-Product Attention\nMulti-Head Attention\nFigure 2: (left) Scaled Dot-Product Attention. (right) Multi-Head Attention consists of several\nattention layers running in parallel.\nof the values, where the weight assigned to each value is computed by a compatibility function of the\nquery with the corresponding key.\n3.2.1\nScaled Dot-Product Attention\nWe call our particular attention "Scaled Dot-Product Attention" (Figure 2). The input consists of\nqueries and keys of dimension dk, and values of dimension dv. We compute the dot products of the\nquery with all keys, divide each by √dk, and apply a softmax function to obtain the weights on the\nvalues.\nIn practice, we compute the attention function on a set of queries simultaneously, packed together\ninto a matrix Q. The keys and values are also packed together into matrices K and V . We compute\nthe matrix of outputs as:\nAttention(Q, K, V ) = softmax(QKT\n√dk\n)V\n(1)\nThe two most commonly used attention functions are additive attention [2], and dot-product (multi-\nplicative) attention. Dot-product attention is identical to our algorithm, except for the scaling factor\nof\n1\n√dk . Additive attention computes the compatibility function using a feed-forward network with\na single hidden layer. While the two are similar in theoretical complexity, dot-product attention is\nmuch faster and more space-efficient in practice, since it can be implemented using highly optimized\nmatrix multiplication code.\nWhile for small values of dk the two mechanisms perform similarly, additive attention outperforms\ndot product attention without scaling for larger values of dk [3]. We suspect that for large values of\ndk, the dot products grow large in magnitude, pushing the softmax function into regions where it has\nextremely small gradients 4. To counteract this effect, we scale the dot products by\n1\n√dk .\n3.2.2\nMulti-Head Attention\nInstead of performing a single attention function with dmodel-dimensional keys, values and queries,\nwe found it beneficial to linearly project the queries, keys and values h times with different, learned\nlinear projections to dk, dk and dv dimensions, respectively. On each of these projected versions of\nqueries, keys and values we then perform the attention function in parallel, yielding dv-dimensional\n4To illustrate why the dot products get large, assume that the components of q and k are independent random\nvariables with mean 0 and variance 1. Then their dot product, q · k = Pdk\ni=1 qiki, has mean 0 and variance dk.\n4\noutput values. These are concatenated and once again projected, resulting in the final values, as\ndepicted in Figure 2.\nMulti-head attention allows the model to jointly attend to information from different representation\nsubspaces at different positions. With a single attention head, averaging inhibits this.\nMultiHead(Q, K, V ) = Concat(head1, ..., headh)W O\nwhere headi = Attention(QW Q\ni , KW K\ni , V W V\ni )\nWhere the projections are parameter matrices W Q\ni\n∈ Rdmodel×dk, W K\ni\n∈ Rdmodel×dk, W V\ni\n∈ Rdmodel×dv\nand W O ∈ Rhdv×dmodel.\nIn this work we employ h = 8 parallel attention layers, or heads. For each of these we use\ndk = dv = dmodel/h = 64. Due to the reduced dimension of each head, the total computational cost\nis similar to that of single-head attention with full dimensionality.\n3.2.3\nApplications of Attention in our Model\nThe Transformer uses multi-head attention in three different ways:\n• In "encoder-decoder attention" layers, the queries come from the previous decoder layer,\nand the memory keys and values come from the output of the encoder. This allows every\nposition in the decoder to attend over all positions in the input sequence. This mimics the\ntypical encoder-decoder attention mechanisms in sequence-to-sequence models such as\n[38, 2, 9].\n• The encoder contains self-attention layers. In a self-attention layer all of the keys, values\nand queries come from the same place, in this case, the output of the previous layer in the\nencoder. Each position in the encoder can attend to all positions in the previous layer of the\nencoder.\n• Similarly, self-attention layers in the decoder allow each position in the decoder to attend to\nall positions in the decoder up to and including that position. We need to prevent leftward\ninformation flow in the decoder to preserve the auto-regressive property. We implement this\ninside of scaled dot-product attention by masking out (setting to −∞) all values in the input\nof the softmax which correspond to illegal connections. See Figure 2.\n3.3\nPosition-wise Feed-Forward Networks\nIn addition to attention sub-layers, each of the layers in our encoder and decoder contains a fully\nconnected feed-forward network, which is applied to each position separately and identically. This\nconsists of two linear transformations with a ReLU activation in between.\nFFN(x) = max(0, xW1 + b1)W2 + b2\n(2)\nWhile the linear transformations are the same across different positions, they use different parameters\nfrom layer to layer. Another way of describing this is as two convolutions with kernel size 1.\nThe dimensionality of input and output is dmodel = 512, and the inner-layer has dimensionality\ndff = 2048.\n3.4\nEmbeddings and Softmax\nSimilarly to other sequence transduction models, we use learned embeddings to convert the input\ntokens and output tokens to vectors of dimension dmodel. We also use the usual learned linear transfor-\nmation and softmax function to convert the decoder output to predicted next-token probabilities. In\nour model, we share the same weight matrix between the two embedding layers and the pre-softmax\nlinear transformation, similar to [30]. In the embedding layers, we multiply those weights by √dmodel.\n5\nTable 1: Maximum path lengths, per-layer complexity and minimum number of sequential operations\nfor different layer types. n is the sequence length, d is the representation dimension, k is the kernel\nsize of convolutions and r the size of the neighborhood in restricted self-attention.\nLayer Type\nComplexity per Layer\nSequential\nMaximum Path Length\nOperations\nSelf-Attention\nO(n2 · d)\nO(1)\nO(1)\nRecurrent\nO(n · d2)\nO(n)\nO(n)\nConvolutional\nO(k · n · d2)\nO(1)\nO(logk(n))\nSelf-Attention (restricted)\nO(r · n · d)\nO(1)\nO(n/r)\n3.5\nPositional Encoding\nSince our model contains no recurrence and no convolution, in order for the model to make use of the\norder of the sequence, we must inject some information about the relative or absolute position of the\ntokens in the sequence. To this end, we add "positional encodings" to the input embeddings at the\nbottoms of the encoder and decoder stacks. The positional encodings have the same dimension dmodel\nas the embeddings, so that the two can be summed. There are many choices of positional encodings,\nlearned and fixed [9].\nIn this work, we use sine and cosine functions of different frequencies:\nPE(pos,2i) = sin(pos/100002i/dmodel)\nPE(pos,2i+1) = cos(pos/100002i/dmodel)\nwhere pos is the position and i is the dimension. That is, each dimension of the positional encoding\ncorresponds to a sinusoid. The wavelengths form a geometric progression from 2π to 10000 · 2π. We\nchose this function because we hypothesized it would allow the model to easily learn to attend by\nrelative positions, since for any fixed offset k, PEpos+k can be represented as a linear function of\nPEpos.\nWe also experimented with using learned positional embeddings [9] instead, and found that the two\nversions produced nearly identical results (see Table 3 row (E)). We chose the sinusoidal version\nbecause it may allow the model to extrapolate to sequence lengths longer than the ones encountered\nduring training.\n4\nWhy Self-Attention\nIn this section we compare various aspects of self-attention layers to the recurrent and convolu-\ntional layers commonly used for mapping one variable-length sequence of symbol representations\n(x1, ..., xn) to another sequence of equal length (z1, ..., zn), with xi, zi ∈ Rd, such as a hidden\nlayer in a typical sequence transduction encoder or decoder. Motivating our use of self-attention we\nconsider three desiderata.\nOne is the total computational complexity per layer. Another is the amount of computation that can\nbe parallelized, as measured by the minimum number of sequential operations required.\nThe third is the path length between long-range dependencies in the network. Learning long-range\ndependencies is a key challenge in many sequence transduction tasks. One key factor affecting the\nability to learn such dependencies is the length of the paths forward and backward signals have to\ntraverse in the network. The shorter these paths between any combination of positions in the input\nand output sequences, the easier it is to learn long-range dependencies [12]. Hence we also compare\nthe maximum path length between any two input and output positions in networks composed of the\ndifferent layer types.\nAs noted in Table 1, a self-attention layer connects all positions with a constant number of sequentially\nexecuted operations, whereas a recurrent layer requires O(n) sequential operations. In terms of\ncomputational complexity, self-attention layers are faster than recurrent layers when the sequence\n6\nlength n is smaller than the representation dimensionality d, which is most often the case with\nsentence representations used by state-of-the-art models in machine translations, such as word-piece\n[38] and byte-pair [31] representations. To improve computational performance for tasks involving\nvery long sequences, self-attention could be restricted to considering only a neighborhood of size r in\nthe input sequence centered around the respective output position. This would increase the maximum\npath length to O(n/r). We plan to investigate this approach further in future work.\nA single convolutional layer with kernel width k < n does not connect all pairs of input and output\npositions. Doing so requires a stack of O(n/k) convolutional layers in the case of contiguous kernels,\nor O(logk(n)) in the case of dilated convolutions [18], increasing the length of the longest paths\nbetween any two positions in the network. Convolutional layers are generally more expensive than\nrecurrent layers, by a factor of k. Separable convolutions [6], however, decrease the complexity\nconsiderably, to O(k · n · d + n · d2). Even with k = n, however, the complexity of a separable\nconvolution is equal to the combination of a self-attention layer and a point-wise feed-forward layer,\nthe approach we take in our model.\nAs side benefit, self-attention could yield more interpretable models. We inspect attention distributions\nfrom our models and present and discuss examples in the appendix. Not only do individual attention\nheads clearly learn to perform different tasks, many appear to exhibit behavior related to the syntactic\nand semantic structure of the sentences.\n5\nTraining\nThis section describes the training regime for our models.\n5.1\nTraining Data and Batching\nWe trained on the standard WMT 2014 English-German dataset consisting of about 4.5 million\nsentence pairs. Sentences were encoded using byte-pair encoding [3], which has a shared source-\ntarget vocabulary of about 37000 tokens. For English-French, we used the significantly larger WMT\n2014 English-French dataset consisting of 36M sentences and split tokens into a 32000 word-piece\nvocabulary [38]. Sentence pairs were batched together by approximate sequence length. Each training\nbatch contained a set of sentence pairs containing approximately 25000 source tokens and 25000\ntarget tokens.\n5.2\nHardware and Schedule\nWe trained our models on one machine with 8 NVIDIA P100 GPUs. For our base models using\nthe hyperparameters described throughout the paper, each training step took about 0.4 seconds. We\ntrained the base models for a total of 100,000 steps or 12 hours. For our big models,(described on the\nbottom line of table 3), step time was 1.0 seconds. The big models were trained for 300,000 steps\n(3.5 days).\n5.3\nOptimizer\nWe used the Adam optimizer [20] with β1 = 0.9, β2 = 0.98 and ϵ = 10−9. We varied the learning\nrate over the course of training, according to the formula:\nlrate = d−0.5\nmodel · min(step_num−0.5, step_num · warmup_steps−1.5)\n(3)\nThis corresponds to increasing the learning rate linearly for the first warmup_steps training steps,\nand decreasing it thereafter proportionally to the inverse square root of the step number. We used\nwarmup_steps = 4000.\n5.4\nRegularization\nWe employ three types of regularization during training:\n7\nTable 2: The Transformer achieves better BLEU scores than previous state-of-the-art models on the\nEnglish-to-German and English-to-French newstest2014 tests at a fraction of the training cost.\nModel\nBLEU\nTraining Cost (FLOPs)\nEN-DE\nEN-FR\nEN-DE\nEN-FR\nByteNet [18]\n23.75\nDeep-Att + PosUnk [39]\n39.2\n1.0 · 1020\nGNMT + RL [38]\n24.6\n39.92\n2.3 · 1019\n1.4 · 1020\nConvS2S [9]\n25.16\n40.46\n9.6 · 1018\n1.5 · 1020\nMoE [32]\n26.03\n40.56\n2.0 · 1019\n1.2 · 1020\nDeep-Att + PosUnk Ensemble [39]\n40.4\n8.0 · 1020\nGNMT + RL Ensemble [38]\n26.30\n41.16\n1.8 · 1020\n1.1 · 1021\nConvS2S Ensemble [9]\n26.36\n41.29\n7.7 · 1019\n1.2 · 1021\nTransformer (base model)\n27.3\n38.1\n3.3 · 1018\nTransformer (big)\n28.4\n41.8\n2.3 · 1019\nResidual Dropout\nWe apply dropout [33] to the output of each sub-layer, before it is added to the\nsub-layer input and normalized. In addition, we apply dropout to the sums of the embeddings and the\npositional encodings in both the encoder and decoder stacks. For the base model, we use a rate of\nPdrop = 0.1.\nLabel Smoothing\nDuring training, we employed label smoothing of value ϵls = 0.1 [36]. This\nhurts perplexity, as the model learns to be more unsure, but improves accuracy and BLEU score.\n6\nResults\n6.1\nMachine Translation\nOn the WMT 2014 English-to-German translation task, the big transformer model (Transformer (big)\nin Table 2) outperforms the best previously reported models (including ensembles) by more than 2.0\nBLEU, establishing a new state-of-the-art BLEU score of 28.4. The configuration of this model is\nlisted in the bottom line of Table 3. Training took 3.5 days on 8 P100 GPUs. Even our base model\nsurpasses all previously published models and ensembles, at a fraction of the training cost of any of\nthe competitive models.\nOn the WMT 2014 English-to-French translation task, our big model achieves a BLEU score of 41.0,\noutperforming all of the previously published single models, at less than 1/4 the training cost of the\nprevious state-of-the-art model. The Transformer (big) model trained for English-to-French used\ndropout rate Pdrop = 0.1, instead of 0.3.\nFor the base models, we used a single model obtained by averaging the last 5 checkpoints, which\nwere written at 10-minute intervals. For the big models, we averaged the last 20 checkpoints. We\nused beam search with a beam size of 4 and length penalty α = 0.6 [38]. These hyperparameters\nwere chosen after experimentation on the development set. We set the maximum output length during\ninference to input length + 50, but terminate early when possible [38].\nTable 2 summarizes our results and compares our translation quality and training costs to other model\narchitectures from the literature. We estimate the number of floating point operations used to train a\nmodel by multiplying the training time, the number of GPUs used, and an estimate of the sustained\nsingle-precision floating-point capacity of each GPU 5.\n6.2\nModel Variations\nTo evaluate the importance of different components of the Transformer, we varied our base model\nin different ways, measuring the change in performance on English-to-German translation on the\n5We used values of 2.8, 3.7, 6.0 and 9.5 TFLOPS for K80, K40, M40 and P100, respectively.\n8\nTable 3: Variations on the Transformer architecture. Unlisted values are identical to those of the base\nmodel. All metrics are on the English-to-German translation development set, newstest2013. Listed\nperplexities are per-wordpiece, according to our byte-pair encoding, and should not be compared to\nper-word perplexities.\nN\ndmodel\ndff\nh\ndk\ndv\nPdrop\nϵls\ntrain\nPPL\nBLEU\nparams\nsteps\n(dev)\n(dev)\n×106\nbase\n6\n512\n2048\n8\n64\n64\n0.1\n0.1\n100K\n4.92\n25.8\n65\n(A)\n1\n512\n512\n5.29\n24.9\n4\n128\n128\n5.00\n25.5\n16\n32\n32\n4.91\n25.8\n32\n16\n16\n5.01\n25.4\n(B)\n16\n5.16\n25.1\n58\n32\n5.01\n25.4\n60\n(C)\n2\n6.11\n23.7\n36\n4\n5.19\n25.3\n50\n8\n4.88\n25.5\n80\n256\n32\n32\n5.75\n24.5\n28\n1024\n128\n128\n4.66\n26.0\n168\n1024\n5.12\n25.4\n53\n4096\n4.75\n26.2\n90\n(D)\n0.0\n5.77\n24.6\n0.2\n4.95\n25.5\n0.0\n4.67\n25.3\n0.2\n5.47\n25.7\n(E)\npositional embedding instead of sinusoids\n4.92\n25.7\nbig\n6\n1024\n4096\n16\n0.3\n300K\n4.33\n26.4\n213\ndevelopment set, newstest2013. We used beam search as described in the previous section, but no\ncheckpoint averaging. We present these results in Table 3.\nIn Table 3 rows (A), we vary the number of attention heads and the attention key and value dimensions,\nkeeping the amount of computation constant, as described in Section 3.2.2. While single-head\nattention is 0.9 BLEU worse than the best setting, quality also drops off with too many heads.\nIn Table 3 rows (B), we observe that reducing the attention key size dk hurts model quality. This\nsuggests that determining compatibility is not easy and that a more sophisticated compatibility\nfunction than dot product may be beneficial. We further observe in rows (C) and (D) that, as expected,\nbigger models are better, and dropout is very helpful in avoiding over-fitting. In row (E) we replace our\nsinusoidal positional encoding with learned positional embeddings [9], and observe nearly identical\nresults to the base model.\n6.3\nEnglish Constituency Parsing\nTo evaluate if the Transformer can generalize to other tasks we performed experiments on English\nconstituency parsing. This task presents specific challenges: the output is subject to strong structural\nconstraints and is significantly longer than the input. Furthermore, RNN sequence-to-sequence\nmodels have not been able to attain state-of-the-art results in small-data regimes [37].\nWe trained a 4-layer transformer with dmodel = 1024 on the Wall Street Journal (WSJ) portion of the\nPenn Treebank [25], about 40K training sentences. We also trained it in a semi-supervised setting,\nusing the larger high-confidence and BerkleyParser corpora from with approximately 17M sentences\n[37]. We used a vocabulary of 16K tokens for the WSJ only setting and a vocabulary of 32K tokens\nfor the semi-supervised setting.\nWe performed only a small number of experiments to select the dropout, both attention and residual\n(section 5.4), learning rates and beam size on the Section 22 development set, all other parameters\nremained unchanged from the English-to-German base translation model. During inference, we\n9\nTable 4: The Transformer generalizes well to English constituency parsing (Results are on Section 23\nof WSJ)\nParser\nTraining\nWSJ 23 F1\nVinyals & Kaiser el al. (2014) [37]\nWSJ only, discriminative\n88.3\nPetrov et al. (2006) [29]\nWSJ only, discriminative\n90.4\nZhu et al. (2013) [40]\nWSJ only, discriminative\n90.4\nDyer et al. (2016) [8]\nWSJ only, discriminative\n91.7\nTransformer (4 layers)\nWSJ only, discriminative\n91.3\nZhu et al. (2013) [40]\nsemi-supervised\n91.3\nHuang & Harper (2009) [14]\nsemi-supervised\n91.3\nMcClosky et al. (2006) [26]\nsemi-supervised\n92.1\nVinyals & Kaiser el al. (2014) [37]\nsemi-supervised\n92.1\nTransformer (4 layers)\nsemi-supervised\n92.7\nLuong et al. (2015) [23]\nmulti-task\n93.0\nDyer et al. (2016) [8]\ngenerative\n93.3\nincreased the maximum output length to input length + 300. We used a beam size of 21 and α = 0.3\nfor both WSJ only and the semi-supervised setting.\nOur results in Table 4 show that despite the lack of task-specific tuning our model performs sur-\nprisingly well, yielding better results than all previously reported models with the exception of the\nRecurrent Neural Network Grammar [8].\nIn contrast to RNN sequence-to-sequence models [37], the Transformer outperforms the Berkeley-\nParser [29] even when training only on the WSJ training set of 40K sentences.\n7\nConclusion\nIn this work, we presented the Transformer, the first sequence transduction model based entirely on\nattention, replacing the recurrent layers most commonly used in encoder-decoder architectures with\nmulti-headed self-attention.\nFor translation tasks, the Transformer can be trained significantly faster than architectures based\non recurrent or convolutional layers. On both WMT 2014 English-to-German and WMT 2014\nEnglish-to-French translation tasks, we achieve a new state of the art. In the former task our best\nmodel outperforms even all previously reported ensembles.\nWe are excited about the future of attention-based models and plan to apply them to other tasks. We\nplan to extend the Transformer to problems involving input and output modalities other than text and\nto investigate local, restricted attention mechanisms to efficiently handle large inputs and outputs\nsuch as images, audio and video. Making generation less sequential is another research goals of ours.\nThe code we used to train and evaluate our models is available at https://github.com/\ntensorflow/tensor2tensor.\nAcknowledgements\nWe are grateful to Nal Kalchbrenner and Stephan Gouws for their fruitful\ncomments, corrections and inspiration.\nReferences\n[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint\narXiv:1607.06450, 2016.\n[2] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly\nlearning to align and translate. CoRR, abs/1409.0473, 2014.\n[3] Denny Britz, Anna Goldie, Minh-Thang Luong, and Quoc V. Le. Massive exploration of neural\nmachine translation architectures. CoRR, abs/1703.03906, 2017.\n[4] Jianpeng Cheng, Li Dong, and Mirella Lapata. Long short-term memory-networks for machine\nreading. arXiv preprint arXiv:1601.06733, 2016.\n10\n[5] Kyunghyun Cho, Bart van Merrienboer, Caglar Gulcehre, Fethi Bougares, Holger Schwenk,\nand Yoshua Bengio. Learning phrase representations using rnn encoder-decoder for statistical\nmachine translation. CoRR, abs/1406.1078, 2014.\n[6] Francois Chollet. Xception: Deep learning with depthwise separable convolutions. arXiv\npreprint arXiv:1610.02357, 2016.\n[7] Junyoung Chung, Çaglar Gülçehre, Kyunghyun Cho, and Yoshua Bengio. Empirical evaluation\nof gated recurrent neural networks on sequence modeling. CoRR, abs/1412.3555, 2014.\n[8] Chris Dyer, Adhiguna Kuncoro, Miguel Ballesteros, and Noah A. Smith. Recurrent neural\nnetwork grammars. In Proc. of NAACL, 2016.\n[9] Jonas Gehring, Michael Auli, David Grangier, Denis Yarats, and Yann N. Dauphin. Convolu-\ntional sequence to sequence learning. arXiv preprint arXiv:1705.03122v2, 2017.\n[10] Alex Graves.\nGenerating sequences with recurrent neural networks.\narXiv preprint\narXiv:1308.0850, 2013.\n[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for im-\nage recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern\nRecognition, pages 770–778, 2016.\n[12] Sepp Hochreiter, Yoshua Bengio, Paolo Frasconi, and Jürgen Schmidhuber. Gradient flow in\nrecurrent nets: the difficulty of learning long-term dependencies, 2001.\n[13] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation,\n9(8):1735–1780, 1997.\n[14] Zhongqiang Huang and Mary Harper. Self-training PCFG grammars with latent annotations\nacross languages. In Proceedings of the 2009 Conference on Empirical Methods in Natural\nLanguage Processing, pages 832–841. ACL, August 2009.\n[15] Rafal Jozefowicz, Oriol Vinyals, Mike Schuster, Noam Shazeer, and Yonghui Wu. Exploring\nthe limits of language modeling. arXiv preprint arXiv:1602.02410, 2016.\n[16] Łukasz Kaiser and Samy Bengio. Can active memory replace attention? In Advances in Neural\nInformation Processing Systems, (NIPS), 2016.\n[17] Łukasz Kaiser and Ilya Sutskever. Neural GPUs learn algorithms. In International Conference\non Learning Representations (ICLR), 2016.\n[18] Nal Kalchbrenner, Lasse Espeholt, Karen Simonyan, Aaron van den Oord, Alex Graves, and Ko-\nray Kavukcuoglu. Neural machine translation in linear time. arXiv preprint arXiv:1610.10099v2,\n2017.\n[19] Yoon Kim, Carl Denton, Luong Hoang, and Alexander M. Rush. Structured attention networks.\nIn International Conference on Learning Representations, 2017.\n[20] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015.\n[21] Oleksii Kuchaiev and Boris Ginsburg. Factorization tricks for LSTM networks. arXiv preprint\narXiv:1703.10722, 2017.\n[22] Zhouhan Lin, Minwei Feng, Cicero Nogueira dos Santos, Mo Yu, Bing Xiang, Bowen\nZhou, and Yoshua Bengio. A structured self-attentive sentence embedding. arXiv preprint\narXiv:1703.03130, 2017.\n[23] Minh-Thang Luong, Quoc V. Le, Ilya Sutskever, Oriol Vinyals, and Lukasz Kaiser. Multi-task\nsequence to sequence learning. arXiv preprint arXiv:1511.06114, 2015.\n[24] Minh-Thang Luong, Hieu Pham, and Christopher D Manning. Effective approaches to attention-\nbased neural machine translation. arXiv preprint arXiv:1508.04025, 2015.\n11\n[25] Mitchell P Marcus, Mary Ann Marcinkiewicz, and Beatrice Santorini. Building a large annotated\ncorpus of english: The penn treebank. Computational linguistics, 19(2):313–330, 1993.\n[26] David McClosky, Eugene Charniak, and Mark Johnson. Effective self-training for parsing. In\nProceedings of the Human Language Technology Conference of the NAACL, Main Conference,\npages 152–159. ACL, June 2006.\n[27] Ankur Parikh, Oscar Täckström, Dipanjan Das, and Jakob Uszkoreit. A decomposable attention\nmodel. In Empirical Methods in Natural Language Processing, 2016.\n[28] Romain Paulus, Caiming Xiong, and Richard Socher. A deep reinforced model for abstractive\nsummarization. arXiv preprint arXiv:1705.04304, 2017.\n[29] Slav Petrov, Leon Barrett, Romain Thibaux, and Dan Klein. Learning accurate, compact,\nand interpretable tree annotation. In Proceedings of the 21st International Conference on\nComputational Linguistics and 44th Annual Meeting of the ACL, pages 433–440. ACL, July\n2006.\n[30] Ofir Press and Lior Wolf. Using the output embedding to improve language models. arXiv\npreprint arXiv:1608.05859, 2016.\n[31] Rico Sennrich, Barry Haddow, and Alexandra Birch. Neural machine translation of rare words\nwith subword units. arXiv preprint arXiv:1508.07909, 2015.\n[32] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton,\nand Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts\nlayer. arXiv preprint arXiv:1701.06538, 2017.\n[33] Nitish Srivastava, Geoffrey E Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdi-\nnov. Dropout: a simple way to prevent neural networks from overfitting. Journal of Machine\nLearning Research, 15(1):1929–1958, 2014.\n[34] Sainbayar Sukhbaatar, Arthur Szlam, Jason Weston, and Rob Fergus. End-to-end memory\nnetworks. In C. Cortes, N. D. Lawrence, D. D. Lee, M. Sugiyama, and R. Garnett, editors,\nAdvances in Neural Information Processing Systems 28, pages 2440–2448. Curran Associates,\nInc., 2015.\n[35] Ilya Sutskever, Oriol Vinyals, and Quoc VV Le. Sequence to sequence learning with neural\nnetworks. In Advances in Neural Information Processing Systems, pages 3104–3112, 2014.\n[36] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna.\nRethinking the inception architecture for computer vision. CoRR, abs/1512.00567, 2015.\n[37] Vinyals & Kaiser, Koo, Petrov, Sutskever, and Hinton. Grammar as a foreign language. In\nAdvances in Neural Information Processing Systems, 2015.\n[38] Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V Le, Mohammad Norouzi, Wolfgang\nMacherey, Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, et al. Google’s neural machine\ntranslation system: Bridging the gap between human and machine translation. arXiv preprint\narXiv:1609.08144, 2016.\n[39] Jie Zhou, Ying Cao, Xuguang Wang, Peng Li, and Wei Xu. Deep recurrent models with\nfast-forward connections for neural machine translation. CoRR, abs/1606.04199, 2016.\n[40] Muhua Zhu, Yue Zhang, Wenliang Chen, Min Zhang, and Jingbo Zhu. Fast and accurate\nshift-reduce constituent parsing. In Proceedings of the 51st Annual Meeting of the ACL (Volume\n1: Long Papers), pages 434–443. ACL, August 2013.\n12\nAttention Visualizations\nInput-Input Layer5\nIt\nis\nin\nthis\nspirit\nthat\na\nmajority\nof\nAmerican\ngovernments\nhave\npassed\nnew\nlaws\nsince\n2009\nmaking\nthe\nregistration\nor\nvoting\nprocess\nmore\ndifficult\n.\n<EOS>\n<pad>\n<pad>\n<pad>\n<pad>\n<pad>\n<pad>\nIt\nis\nin\nthis\nspirit\nthat\na\nmajority\nof\nAmerican\ngovernments\nhave\npassed\nnew\nlaws\nsince\n2009\nmaking\nthe\nregistration\nor\nvoting\nprocess\nmore\ndifficult\n.\n<EOS>\n<pad>\n<pad>\n<pad>\n<pad>\n<pad>\n<pad>\nFigure 3: An example of the attention mechanism following long-distance dependencies in the\nencoder self-attention in layer 5 of 6. Many of the attention heads attend to a distant dependency of\nthe verb ‘making’, completing the phrase ‘making...more difficult’. Attentions here shown only for\nthe word ‘making’. Different colors represent different heads. Best viewed in color.\n13\nInput-Input Layer5\nThe\nLaw\nwill\nnever\nbe\nperfect\n,\nbut\nits\napplication\nshould\nbe\njust\n-\nthis\nis\nwhat\nwe\nare\nmissing\n,\nin\nmy\nopinion\n.\n<EOS>\n<pad>\nThe\nLaw\nwill\nnever\nbe\nperfect\n,\nbut\nits\napplication\nshould\nbe\njust\n-\nthis\nis\nwhat\nwe\nare\nmissing\n,\nin\nmy\nopinion\n.\n<EOS>\n<pad>\nInput-Input Layer5\nThe\nLaw\nwill\nnever\nbe\nperfect\n,\nbut\nits\napplication\nshould\nbe\njust\n-\nthis\nis\nwhat\nwe\nare\nmissing\n,\nin\nmy\nopinion\n.\n<EOS>\n<pad>\nThe\nLaw\nwill\nnever\nbe\nperfect\n,\nbut\nits\napplication\nshould\nbe\njust\n-\nthis\nis\nwhat\nwe\nare\nmissing\n,\nin\nmy\nopinion\n.\n<EOS>\n<pad>\nFigure 4: Two attention heads, also in layer 5 of 6, apparently involved in anaphora resolution. Top:\nFull attentions for head 5. Bottom: Isolated attentions from just the word ‘its’ for attention heads 5\nand 6. Note that the attentions are very sharp for this word.\n14\nInput-Input Layer5\nThe\nLaw\nwill\nnever\nbe\nperfect\n,\nbut\nits\napplication\nshould\nbe\njust\n-\nthis\nis\nwhat\nwe\nare\nmissing\n,\nin\nmy\nopinion\n.\n<EOS>\n<pad>\nThe\nLaw\nwill\nnever\nbe\nperfect\n,\nbut\nits\napplication\nshould\nbe\njust\n-\nthis\nis\nwhat\nwe\nare\nmissing\n,\nin\nmy\nopinion\n.\n<EOS>\n<pad>\nInput-Input Layer5\nThe\nLaw\nwill\nnever\nbe\nperfect\n,\nbut\nits\napplication\nshould\nbe\njust\n-\nthis\nis\nwhat\nwe\nare\nmissing\n,\nin\nmy\nopinion\n.\n<EOS>\n<pad>\nThe\nLaw\nwill\nnever\nbe\nperfect\n,\nbut\nits\napplication\nshould\nbe\njust\n-\nthis\nis\nwhat\nwe\nare\nmissing\n,\nin\nmy\nopinion\n.\n<EOS>\n<pad>\nFigure 5: Many of the attention heads exhibit behaviour that seems related to the structure of the\nsentence. We give two such examples above, from two different heads from the encoder self-attention\nat layer 5 of 6. The heads clearly learned to perform different tasks.\n15\n', metadata={'Published': '2023-08-02', 'Title': 'Attention Is All You Need', 'Authors': 'Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, Illia Polosukhin', 'Summary': 'The dominant sequence transduction models are based on complex recurrent or\nconvolutional neural networks in an encoder-decoder configuration. The best\nperforming models also connect the encoder and decoder through an attention\nmechanism. We propose a new simple network architecture, the Transformer, based\nsolely on attention mechanisms, dispensing with recurrence and convolutions\nentirely. Experiments on two machine translation tasks show these models to be\nsuperior in quality while being more parallelizable and requiring significantly\nless time to train. Our model achieves 28.4 BLEU on the WMT 2014\nEnglish-to-German translation task, improving over the existing best results,\nincluding ensembles by over 2 BLEU. On the WMT 2014 English-to-French\ntranslation task, our model establishes a new single-model state-of-the-art\nBLEU score of 41.8 after training for 3.5 days on eight GPUs, a small fraction\nof the training costs of the best models from the literature. We show that the\nTransformer generalizes well to other tasks by applying it successfully to\nEnglish constituency parsing both with large and limited training data.'})][Document(page_content='Provided proper attribution is provided, Google hereby grants permission to\nreproduce the tables and figures in this paper solely for use in journalistic or\nscholarly works.\nAttention Is All You Need\nAshish Vaswani∗\nGoogle Brain\navaswani@google.com\nNoam Shazeer∗\nGoogle Brain\nnoam@google.com\nNiki Parmar∗\nGoogle Research\nnikip@google.com\nJakob Uszkoreit∗\nGoogle Research\nusz@google.com\nLlion Jones∗\nGoogle Research\nllion@google.com\nAidan N. Gomez∗ †\nUniversity of Toronto\naidan@cs.toronto.edu\nŁukasz Kaiser∗\nGoogle Brain\nlukaszkaiser@google.com\nIllia Polosukhin∗ ‡\nillia.polosukhin@gmail.com\nAbstract\nThe dominant sequence transduction models are based on complex recurrent or\nconvolutional neural networks that include an encoder and a decoder. The best\nperforming models also connect the encoder and decoder through an attention\nmechanism. We propose a new simple network architecture, the Transformer,\nbased solely on attention mechanisms, dispensing with recurrence and convolutions\nentirely. Experiments on two machine translation tasks show these models to\nbe superior in quality while being more parallelizable and requiring significantly\nless time to train. Our model achieves 28.4 BLEU on the WMT 2014 English-\nto-German translation task, improving over the existing best results, including\nensembles, by over 2 BLEU. On the WMT 2014 English-to-French translation task,\nour model establishes a new single-model state-of-the-art BLEU score of 41.8 after\ntraining for 3.5 days on eight GPUs, a small fraction of the training costs of the\nbest models from the literature. We show that the Transformer generalizes well to\nother tasks by applying it successfully to English constituency parsing both with\nlarge and limited training data.\n∗Equal contribution. Listing order is random. Jakob proposed replacing RNNs with self-attention and started\nthe effort to evaluate this idea. Ashish, with Illia, designed and implemented the first Transformer models and\nhas been crucially involved in every aspect of this work. Noam proposed scaled dot-product attention, multi-head\nattention and the parameter-free position representation and became the other person involved in nearly every\ndetail. Niki designed, implemented, tuned and evaluated countless model variants in our original codebase and\ntensor2tensor. Llion also experimented with novel model variants, was responsible for our initial codebase, and\nefficient inference and visualizations. Lukasz and Aidan spent countless long days designing various parts of and\nimplementing tensor2tensor, replacing our earlier codebase, greatly improving results and massively accelerating\nour research.\n†Work performed while at Google Brain.\n‡Work performed while at Google Research.\n31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA.\narXiv:1706.03762v7  [cs.CL]  2 Aug 2023\n1\nIntroduction\nRecurrent neural networks, long short-term memory [13] and gated recurrent [7] neural networks\nin particular, have been firmly established as state of the art approaches in sequence modeling and\ntransduction problems such as language modeling and machine translation [35, 2, 5]. Numerous\nefforts have since continued to push the boundaries of recurrent language models and encoder-decoder\narchitectures [38, 24, 15].\nRecurrent models typically factor computation along the symbol positions of the input and output\nsequences. Aligning the positions to steps in computation time, they generate a sequence of hidden\nstates ht, as a function of the previous hidden state ht−1 and the input for position t. This inherently\nsequential nature precludes parallelization within training examples, which becomes critical at longer\nsequence lengths, as memory constraints limit batching across examples. Recent work has achieved\nsignificant improvements in computational efficiency through factorization tricks [21] and conditional\ncomputation [32], while also improving model performance in case of the latter. The fundamental\nconstraint of sequential computation, however, remains.\nAttention mechanisms have become an integral part of compelling sequence modeling and transduc-\ntion models in various tasks, allowing modeling of dependencies without regard to their distance in\nthe input or output sequences [2, 19]. In all but a few cases [27], however, such attention mechanisms\nare used in conjunction with a recurrent network.\nIn this work we propose the Transformer, a model architecture eschewing recurrence and instead\nrelying entirely on an attention mechanism to draw global dependencies between input and output.\nThe Transformer allows for significantly more parallelization and can reach a new state of the art in\ntranslation quality after being trained for as little as twelve hours on eight P100 GPUs.\n2\nBackground\nThe goal of reducing sequential computation also forms the foundation of the Extended Neural GPU\n[16], ByteNet [18] and ConvS2S [9], all of which use convolutional neural networks as basic building\nblock, computing hidden representations in parallel for all input and output positions. In these models,\nthe number of operations required to relate signals from two arbitrary input or output positions grows\nin the distance between positions, linearly for ConvS2S and logarithmically for ByteNet. This makes\nit more difficult to learn dependencies between distant positions [12]. In the Transformer this is\nreduced to a constant number of operations, albeit at the cost of reduced effective resolution due\nto averaging attention-weighted positions, an effect we counteract with Multi-Head Attention as\ndescribed in section 3.2.\nSelf-attention, sometimes called intra-attention is an attention mechanism relating different positions\nof a single sequence in order to compute a representation of the sequence. Self-attention has been\nused successfully in a variety of tasks including reading comprehension, abstractive summarization,\ntextual entailment and learning task-independent sentence representations [4, 27, 28, 22].\nEnd-to-end memory networks are based on a recurrent attention mechanism instead of sequence-\naligned recurrence and have been shown to perform well on simple-language question answering and\nlanguage modeling tasks [34].\nTo the best of our knowledge, however, the Transformer is the first transduction model relying\nentirely on self-attention to compute representations of its input and output without using sequence-\naligned RNNs or convolution. In the following sections, we will describe the Transformer, motivate\nself-attention and discuss its advantages over models such as [17, 18] and [9].\n3\nModel Architecture\nMost competitive neural sequence transduction models have an encoder-decoder structure [5, 2, 35].\nHere, the encoder maps an input sequence of symbol representations (x1, ..., xn) to a sequence\nof continuous representations z = (z1, ..., zn). Given z, the decoder then generates an output\nsequence (y1, ..., ym) of symbols one element at a time. At each step the model is auto-regressive\n[10], consuming the previously generated symbols as additional input when generating the next.\n2\nFigure 1: The Transformer - model architecture.\nThe Transformer follows this overall architecture using stacked self-attention and point-wise, fully\nconnected layers for both the encoder and decoder, shown in the left and right halves of Figure 1,\nrespectively.\n3.1\nEncoder and Decoder Stacks\nEncoder:\nThe encoder is composed of a stack of N = 6 identical layers. Each layer has two\nsub-layers. The first is a multi-head self-attention mechanism, and the second is a simple, position-\nwise fully connected feed-forward network. We employ a residual connection [11] around each of\nthe two sub-layers, followed by layer normalization [1]. That is, the output of each sub-layer is\nLayerNorm(x + Sublayer(x)), where Sublayer(x) is the function implemented by the sub-layer\nitself. To facilitate these residual connections, all sub-layers in the model, as well as the embedding\nlayers, produce outputs of dimension dmodel = 512.\nDecoder:\nThe decoder is also composed of a stack of N = 6 identical layers. In addition to the two\nsub-layers in each encoder layer, the decoder inserts a third sub-layer, which performs multi-head\nattention over the output of the encoder stack. Similar to the encoder, we employ residual connections\naround each of the sub-layers, followed by layer normalization. We also modify the self-attention\nsub-layer in the decoder stack to prevent positions from attending to subsequent positions. This\nmasking, combined with fact that the output embeddings are offset by one position, ensures that the\npredictions for position i can depend only on the known outputs at positions less than i.\n3.2\nAttention\nAn attention function can be described as mapping a query and a set of key-value pairs to an output,\nwhere the query, keys, values, and output are all vectors. The output is computed as a weighted sum\n3\nScaled Dot-Product Attention\nMulti-Head Attention\nFigure 2: (left) Scaled Dot-Product Attention. (right) Multi-Head Attention consists of several\nattention layers running in parallel.\nof the values, where the weight assigned to each value is computed by a compatibility function of the\nquery with the corresponding key.\n3.2.1\nScaled Dot-Product Attention\nWe call our particular attention "Scaled Dot-Product Attention" (Figure 2). The input consists of\nqueries and keys of dimension dk, and values of dimension dv. We compute the dot products of the\nquery with all keys, divide each by √dk, and apply a softmax function to obtain the weights on the\nvalues.\nIn practice, we compute the attention function on a set of queries simultaneously, packed together\ninto a matrix Q. The keys and values are also packed together into matrices K and V . We compute\nthe matrix of outputs as:\nAttention(Q, K, V ) = softmax(QKT\n√dk\n)V\n(1)\nThe two most commonly used attention functions are additive attention [2], and dot-product (multi-\nplicative) attention. Dot-product attention is identical to our algorithm, except for the scaling factor\nof\n1\n√dk . Additive attention computes the compatibility function using a feed-forward network with\na single hidden layer. While the two are similar in theoretical complexity, dot-product attention is\nmuch faster and more space-efficient in practice, since it can be implemented using highly optimized\nmatrix multiplication code.\nWhile for small values of dk the two mechanisms perform similarly, additive attention outperforms\ndot product attention without scaling for larger values of dk [3]. We suspect that for large values of\ndk, the dot products grow large in magnitude, pushing the softmax function into regions where it has\nextremely small gradients 4. To counteract this effect, we scale the dot products by\n1\n√dk .\n3.2.2\nMulti-Head Attention\nInstead of performing a single attention function with dmodel-dimensional keys, values and queries,\nwe found it beneficial to linearly project the queries, keys and values h times with different, learned\nlinear projections to dk, dk and dv dimensions, respectively. On each of these projected versions of\nqueries, keys and values we then perform the attention function in parallel, yielding dv-dimensional\n4To illustrate why the dot products get large, assume that the components of q and k are independent random\nvariables with mean 0 and variance 1. Then their dot product, q · k = Pdk\ni=1 qiki, has mean 0 and variance dk.\n4\noutput values. These are concatenated and once again projected, resulting in the final values, as\ndepicted in Figure 2.\nMulti-head attention allows the model to jointly attend to information from different representation\nsubspaces at different positions. With a single attention head, averaging inhibits this.\nMultiHead(Q, K, V ) = Concat(head1, ..., headh)W O\nwhere headi = Attention(QW Q\ni , KW K\ni , V W V\ni )\nWhere the projections are parameter matrices W Q\ni\n∈ Rdmodel×dk, W K\ni\n∈ Rdmodel×dk, W V\ni\n∈ Rdmodel×dv\nand W O ∈ Rhdv×dmodel.\nIn this work we employ h = 8 parallel attention layers, or heads. For each of these we use\ndk = dv = dmodel/h = 64. Due to the reduced dimension of each head, the total computational cost\nis similar to that of single-head attention with full dimensionality.\n3.2.3\nApplications of Attention in our Model\nThe Transformer uses multi-head attention in three different ways:\n• In "encoder-decoder attention" layers, the queries come from the previous decoder layer,\nand the memory keys and values come from the output of the encoder. This allows every\nposition in the decoder to attend over all positions in the input sequence. This mimics the\ntypical encoder-decoder attention mechanisms in sequence-to-sequence models such as\n[38, 2, 9].\n• The encoder contains self-attention layers. In a self-attention layer all of the keys, values\nand queries come from the same place, in this case, the output of the previous layer in the\nencoder. Each position in the encoder can attend to all positions in the previous layer of the\nencoder.\n• Similarly, self-attention layers in the decoder allow each position in the decoder to attend to\nall positions in the decoder up to and including that position. We need to prevent leftward\ninformation flow in the decoder to preserve the auto-regressive property. We implement this\ninside of scaled dot-product attention by masking out (setting to −∞) all values in the input\nof the softmax which correspond to illegal connections. See Figure 2.\n3.3\nPosition-wise Feed-Forward Networks\nIn addition to attention sub-layers, each of the layers in our encoder and decoder contains a fully\nconnected feed-forward network, which is applied to each position separately and identically. This\nconsists of two linear transformations with a ReLU activation in between.\nFFN(x) = max(0, xW1 + b1)W2 + b2\n(2)\nWhile the linear transformations are the same across different positions, they use different parameters\nfrom layer to layer. Another way of describing this is as two convolutions with kernel size 1.\nThe dimensionality of input and output is dmodel = 512, and the inner-layer has dimensionality\ndff = 2048.\n3.4\nEmbeddings and Softmax\nSimilarly to other sequence transduction models, we use learned embeddings to convert the input\ntokens and output tokens to vectors of dimension dmodel. We also use the usual learned linear transfor-\nmation and softmax function to convert the decoder output to predicted next-token probabilities. In\nour model, we share the same weight matrix between the two embedding layers and the pre-softmax\nlinear transformation, similar to [30]. In the embedding layers, we multiply those weights by √dmodel.\n5\nTable 1: Maximum path lengths, per-layer complexity and minimum number of sequential operations\nfor different layer types. n is the sequence length, d is the representation dimension, k is the kernel\nsize of convolutions and r the size of the neighborhood in restricted self-attention.\nLayer Type\nComplexity per Layer\nSequential\nMaximum Path Length\nOperations\nSelf-Attention\nO(n2 · d)\nO(1)\nO(1)\nRecurrent\nO(n · d2)\nO(n)\nO(n)\nConvolutional\nO(k · n · d2)\nO(1)\nO(logk(n))\nSelf-Attention (restricted)\nO(r · n · d)\nO(1)\nO(n/r)\n3.5\nPositional Encoding\nSince our model contains no recurrence and no convolution, in order for the model to make use of the\norder of the sequence, we must inject some information about the relative or absolute position of the\ntokens in the sequence. To this end, we add "positional encodings" to the input embeddings at the\nbottoms of the encoder and decoder stacks. The positional encodings have the same dimension dmodel\nas the embeddings, so that the two can be summed. There are many choices of positional encodings,\nlearned and fixed [9].\nIn this work, we use sine and cosine functions of different frequencies:\nPE(pos,2i) = sin(pos/100002i/dmodel)\nPE(pos,2i+1) = cos(pos/100002i/dmodel)\nwhere pos is the position and i is the dimension. That is, each dimension of the positional encoding\ncorresponds to a sinusoid. The wavelengths form a geometric progression from 2π to 10000 · 2π. We\nchose this function because we hypothesized it would allow the model to easily learn to attend by\nrelative positions, since for any fixed offset k, PEpos+k can be represented as a linear function of\nPEpos.\nWe also experimented with using learned positional embeddings [9] instead, and found that the two\nversions produced nearly identical results (see Table 3 row (E)). We chose the sinusoidal version\nbecause it may allow the model to extrapolate to sequence lengths longer than the ones encountered\nduring training.\n4\nWhy Self-Attention\nIn this section we compare various aspects of self-attention layers to the recurrent and convolu-\ntional layers commonly used for mapping one variable-length sequence of symbol representations\n(x1, ..., xn) to another sequence of equal length (z1, ..., zn), with xi, zi ∈ Rd, such as a hidden\nlayer in a typical sequence transduction encoder or decoder. Motivating our use of self-attention we\nconsider three desiderata.\nOne is the total computational complexity per layer. Another is the amount of computation that can\nbe parallelized, as measured by the minimum number of sequential operations required.\nThe third is the path length between long-range dependencies in the network. Learning long-range\ndependencies is a key challenge in many sequence transduction tasks. One key factor affecting the\nability to learn such dependencies is the length of the paths forward and backward signals have to\ntraverse in the network. The shorter these paths between any combination of positions in the input\nand output sequences, the easier it is to learn long-range dependencies [12]. Hence we also compare\nthe maximum path length between any two input and output positions in networks composed of the\ndifferent layer types.\nAs noted in Table 1, a self-attention layer connects all positions with a constant number of sequentially\nexecuted operations, whereas a recurrent layer requires O(n) sequential operations. In terms of\ncomputational complexity, self-attention layers are faster than recurrent layers when the sequence\n6\nlength n is smaller than the representation dimensionality d, which is most often the case with\nsentence representations used by state-of-the-art models in machine translations, such as word-piece\n[38] and byte-pair [31] representations. To improve computational performance for tasks involving\nvery long sequences, self-attention could be restricted to considering only a neighborhood of size r in\nthe input sequence centered around the respective output position. This would increase the maximum\npath length to O(n/r). We plan to investigate this approach further in future work.\nA single convolutional layer with kernel width k < n does not connect all pairs of input and output\npositions. Doing so requires a stack of O(n/k) convolutional layers in the case of contiguous kernels,\nor O(logk(n)) in the case of dilated convolutions [18], increasing the length of the longest paths\nbetween any two positions in the network. Convolutional layers are generally more expensive than\nrecurrent layers, by a factor of k. Separable convolutions [6], however, decrease the complexity\nconsiderably, to O(k · n · d + n · d2). Even with k = n, however, the complexity of a separable\nconvolution is equal to the combination of a self-attention layer and a point-wise feed-forward layer,\nthe approach we take in our model.\nAs side benefit, self-attention could yield more interpretable models. We inspect attention distributions\nfrom our models and present and discuss examples in the appendix. Not only do individual attention\nheads clearly learn to perform different tasks, many appear to exhibit behavior related to the syntactic\nand semantic structure of the sentences.\n5\nTraining\nThis section describes the training regime for our models.\n5.1\nTraining Data and Batching\nWe trained on the standard WMT 2014 English-German dataset consisting of about 4.5 million\nsentence pairs. Sentences were encoded using byte-pair encoding [3], which has a shared source-\ntarget vocabulary of about 37000 tokens. For English-French, we used the significantly larger WMT\n2014 English-French dataset consisting of 36M sentences and split tokens into a 32000 word-piece\nvocabulary [38]. Sentence pairs were batched together by approximate sequence length. Each training\nbatch contained a set of sentence pairs containing approximately 25000 source tokens and 25000\ntarget tokens.\n5.2\nHardware and Schedule\nWe trained our models on one machine with 8 NVIDIA P100 GPUs. For our base models using\nthe hyperparameters described throughout the paper, each training step took about 0.4 seconds. We\ntrained the base models for a total of 100,000 steps or 12 hours. For our big models,(described on the\nbottom line of table 3), step time was 1.0 seconds. The big models were trained for 300,000 steps\n(3.5 days).\n5.3\nOptimizer\nWe used the Adam optimizer [20] with β1 = 0.9, β2 = 0.98 and ϵ = 10−9. We varied the learning\nrate over the course of training, according to the formula:\nlrate = d−0.5\nmodel · min(step_num−0.5, step_num · warmup_steps−1.5)\n(3)\nThis corresponds to increasing the learning rate linearly for the first warmup_steps training steps,\nand decreasing it thereafter proportionally to the inverse square root of the step number. We used\nwarmup_steps = 4000.\n5.4\nRegularization\nWe employ three types of regularization during training:\n7\nTable 2: The Transformer achieves better BLEU scores than previous state-of-the-art models on the\nEnglish-to-German and English-to-French newstest2014 tests at a fraction of the training cost.\nModel\nBLEU\nTraining Cost (FLOPs)\nEN-DE\nEN-FR\nEN-DE\nEN-FR\nByteNet [18]\n23.75\nDeep-Att + PosUnk [39]\n39.2\n1.0 · 1020\nGNMT + RL [38]\n24.6\n39.92\n2.3 · 1019\n1.4 · 1020\nConvS2S [9]\n25.16\n40.46\n9.6 · 1018\n1.5 · 1020\nMoE [32]\n26.03\n40.56\n2.0 · 1019\n1.2 · 1020\nDeep-Att + PosUnk Ensemble [39]\n40.4\n8.0 · 1020\nGNMT + RL Ensemble [38]\n26.30\n41.16\n1.8 · 1020\n1.1 · 1021\nConvS2S Ensemble [9]\n26.36\n41.29\n7.7 · 1019\n1.2 · 1021\nTransformer (base model)\n27.3\n38.1\n3.3 · 1018\nTransformer (big)\n28.4\n41.8\n2.3 · 1019\nResidual Dropout\nWe apply dropout [33] to the output of each sub-layer, before it is added to the\nsub-layer input and normalized. In addition, we apply dropout to the sums of the embeddings and the\npositional encodings in both the encoder and decoder stacks. For the base model, we use a rate of\nPdrop = 0.1.\nLabel Smoothing\nDuring training, we employed label smoothing of value ϵls = 0.1 [36]. This\nhurts perplexity, as the model learns to be more unsure, but improves accuracy and BLEU score.\n6\nResults\n6.1\nMachine Translation\nOn the WMT 2014 English-to-German translation task, the big transformer model (Transformer (big)\nin Table 2) outperforms the best previously reported models (including ensembles) by more than 2.0\nBLEU, establishing a new state-of-the-art BLEU score of 28.4. The configuration of this model is\nlisted in the bottom line of Table 3. Training took 3.5 days on 8 P100 GPUs. Even our base model\nsurpasses all previously published models and ensembles, at a fraction of the training cost of any of\nthe competitive models.\nOn the WMT 2014 English-to-French translation task, our big model achieves a BLEU score of 41.0,\noutperforming all of the previously published single models, at less than 1/4 the training cost of the\nprevious state-of-the-art model. The Transformer (big) model trained for English-to-French used\ndropout rate Pdrop = 0.1, instead of 0.3.\nFor the base models, we used a single model obtained by averaging the last 5 checkpoints, which\nwere written at 10-minute intervals. For the big models, we averaged the last 20 checkpoints. We\nused beam search with a beam size of 4 and length penalty α = 0.6 [38]. These hyperparameters\nwere chosen after experimentation on the development set. We set the maximum output length during\ninference to input length + 50, but terminate early when possible [38].\nTable 2 summarizes our results and compares our translation quality and training costs to other model\narchitectures from the literature. We estimate the number of floating point operations used to train a\nmodel by multiplying the training time, the number of GPUs used, and an estimate of the sustained\nsingle-precision floating-point capacity of each GPU 5.\n6.2\nModel Variations\nTo evaluate the importance of different components of the Transformer, we varied our base model\nin different ways, measuring the change in performance on English-to-German translation on the\n5We used values of 2.8, 3.7, 6.0 and 9.5 TFLOPS for K80, K40, M40 and P100, respectively.\n8\nTable 3: Variations on the Transformer architecture. Unlisted values are identical to those of the base\nmodel. All metrics are on the English-to-German translation development set, newstest2013. Listed\nperplexities are per-wordpiece, according to our byte-pair encoding, and should not be compared to\nper-word perplexities.\nN\ndmodel\ndff\nh\ndk\ndv\nPdrop\nϵls\ntrain\nPPL\nBLEU\nparams\nsteps\n(dev)\n(dev)\n×106\nbase\n6\n512\n2048\n8\n64\n64\n0.1\n0.1\n100K\n4.92\n25.8\n65\n(A)\n1\n512\n512\n5.29\n24.9\n4\n128\n128\n5.00\n25.5\n16\n32\n32\n4.91\n25.8\n32\n16\n16\n5.01\n25.4\n(B)\n16\n5.16\n25.1\n58\n32\n5.01\n25.4\n60\n(C)\n2\n6.11\n23.7\n36\n4\n5.19\n25.3\n50\n8\n4.88\n25.5\n80\n256\n32\n32\n5.75\n24.5\n28\n1024\n128\n128\n4.66\n26.0\n168\n1024\n5.12\n25.4\n53\n4096\n4.75\n26.2\n90\n(D)\n0.0\n5.77\n24.6\n0.2\n4.95\n25.5\n0.0\n4.67\n25.3\n0.2\n5.47\n25.7\n(E)\npositional embedding instead of sinusoids\n4.92\n25.7\nbig\n6\n1024\n4096\n16\n0.3\n300K\n4.33\n26.4\n213\ndevelopment set, newstest2013. We used beam search as described in the previous section, but no\ncheckpoint averaging. We present these results in Table 3.\nIn Table 3 rows (A), we vary the number of attention heads and the attention key and value dimensions,\nkeeping the amount of computation constant, as described in Section 3.2.2. While single-head\nattention is 0.9 BLEU worse than the best setting, quality also drops off with too many heads.\nIn Table 3 rows (B), we observe that reducing the attention key size dk hurts model quality. This\nsuggests that determining compatibility is not easy and that a more sophisticated compatibility\nfunction than dot product may be beneficial. We further observe in rows (C) and (D) that, as expected,\nbigger models are better, and dropout is very helpful in avoiding over-fitting. In row (E) we replace our\nsinusoidal positional encoding with learned positional embeddings [9], and observe nearly identical\nresults to the base model.\n6.3\nEnglish Constituency Parsing\nTo evaluate if the Transformer can generalize to other tasks we performed experiments on English\nconstituency parsing. This task presents specific challenges: the output is subject to strong structural\nconstraints and is significantly longer than the input. Furthermore, RNN sequence-to-sequence\nmodels have not been able to attain state-of-the-art results in small-data regimes [37].\nWe trained a 4-layer transformer with dmodel = 1024 on the Wall Street Journal (WSJ) portion of the\nPenn Treebank [25], about 40K training sentences. We also trained it in a semi-supervised setting,\nusing the larger high-confidence and BerkleyParser corpora from with approximately 17M sentences\n[37]. We used a vocabulary of 16K tokens for the WSJ only setting and a vocabulary of 32K tokens\nfor the semi-supervised setting.\nWe performed only a small number of experiments to select the dropout, both attention and residual\n(section 5.4), learning rates and beam size on the Section 22 development set, all other parameters\nremained unchanged from the English-to-German base translation model. During inference, we\n9\nTable 4: The Transformer generalizes well to English constituency parsing (Results are on Section 23\nof WSJ)\nParser\nTraining\nWSJ 23 F1\nVinyals & Kaiser el al. (2014) [37]\nWSJ only, discriminative\n88.3\nPetrov et al. (2006) [29]\nWSJ only, discriminative\n90.4\nZhu et al. (2013) [40]\nWSJ only, discriminative\n90.4\nDyer et al. (2016) [8]\nWSJ only, discriminative\n91.7\nTransformer (4 layers)\nWSJ only, discriminative\n91.3\nZhu et al. (2013) [40]\nsemi-supervised\n91.3\nHuang & Harper (2009) [14]\nsemi-supervised\n91.3\nMcClosky et al. (2006) [26]\nsemi-supervised\n92.1\nVinyals & Kaiser el al. (2014) [37]\nsemi-supervised\n92.1\nTransformer (4 layers)\nsemi-supervised\n92.7\nLuong et al. (2015) [23]\nmulti-task\n93.0\nDyer et al. (2016) [8]\ngenerative\n93.3\nincreased the maximum output length to input length + 300. We used a beam size of 21 and α = 0.3\nfor both WSJ only and the semi-supervised setting.\nOur results in Table 4 show that despite the lack of task-specific tuning our model performs sur-\nprisingly well, yielding better results than all previously reported models with the exception of the\nRecurrent Neural Network Grammar [8].\nIn contrast to RNN sequence-to-sequence models [37], the Transformer outperforms the Berkeley-\nParser [29] even when training only on the WSJ training set of 40K sentences.\n7\nConclusion\nIn this work, we presented the Transformer, the first sequence transduction model based entirely on\nattention, replacing the recurrent layers most commonly used in encoder-decoder architectures with\nmulti-headed self-attention.\nFor translation tasks, the Transformer can be trained significantly faster than architectures based\non recurrent or convolutional layers. On both WMT 2014 English-to-German and WMT 2014\nEnglish-to-French translation tasks, we achieve a new state of the art. In the former task our best\nmodel outperforms even all previously reported ensembles.\nWe are excited about the future of attention-based models and plan to apply them to other tasks. We\nplan to extend the Transformer to problems involving input and output modalities other than text and\nto investigate local, restricted attention mechanisms to efficiently handle large inputs and outputs\nsuch as images, audio and video. Making generation less sequential is another research goals of ours.\nThe code we used to train and evaluate our models is available at https://github.com/\ntensorflow/tensor2tensor.\nAcknowledgements\nWe are grateful to Nal Kalchbrenner and Stephan Gouws for their fruitful\ncomments, corrections and inspiration.\nReferences\n[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint\narXiv:1607.06450, 2016.\n[2] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly\nlearning to align and translate. CoRR, abs/1409.0473, 2014.\n[3] Denny Britz, Anna Goldie, Minh-Thang Luong, and Quoc V. Le. Massive exploration of neural\nmachine translation architectures. CoRR, abs/1703.03906, 2017.\n[4] Jianpeng Cheng, Li Dong, and Mirella Lapata. Long short-term memory-networks for machine\nreading. arXiv preprint arXiv:1601.06733, 2016.\n10\n[5] Kyunghyun Cho, Bart van Merrienboer, Caglar Gulcehre, Fethi Bougares, Holger Schwenk,\nand Yoshua Bengio. Learning phrase representations using rnn encoder-decoder for statistical\nmachine translation. CoRR, abs/1406.1078, 2014.\n[6] Francois Chollet. Xception: Deep learning with depthwise separable convolutions. arXiv\npreprint arXiv:1610.02357, 2016.\n[7] Junyoung Chung, Çaglar Gülçehre, Kyunghyun Cho, and Yoshua Bengio. Empirical evaluation\nof gated recurrent neural networks on sequence modeling. CoRR, abs/1412.3555, 2014.\n[8] Chris Dyer, Adhiguna Kuncoro, Miguel Ballesteros, and Noah A. Smith. Recurrent neural\nnetwork grammars. In Proc. of NAACL, 2016.\n[9] Jonas Gehring, Michael Auli, David Grangier, Denis Yarats, and Yann N. Dauphin. Convolu-\ntional sequence to sequence learning. arXiv preprint arXiv:1705.03122v2, 2017.\n[10] Alex Graves.\nGenerating sequences with recurrent neural networks.\narXiv preprint\narXiv:1308.0850, 2013.\n[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for im-\nage recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern\nRecognition, pages 770–778, 2016.\n[12] Sepp Hochreiter, Yoshua Bengio, Paolo Frasconi, and Jürgen Schmidhuber. Gradient flow in\nrecurrent nets: the difficulty of learning long-term dependencies, 2001.\n[13] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation,\n9(8):1735–1780, 1997.\n[14] Zhongqiang Huang and Mary Harper. Self-training PCFG grammars with latent annotations\nacross languages. In Proceedings of the 2009 Conference on Empirical Methods in Natural\nLanguage Processing, pages 832–841. ACL, August 2009.\n[15] Rafal Jozefowicz, Oriol Vinyals, Mike Schuster, Noam Shazeer, and Yonghui Wu. Exploring\nthe limits of language modeling. arXiv preprint arXiv:1602.02410, 2016.\n[16] Łukasz Kaiser and Samy Bengio. Can active memory replace attention? In Advances in Neural\nInformation Processing Systems, (NIPS), 2016.\n[17] Łukasz Kaiser and Ilya Sutskever. Neural GPUs learn algorithms. In International Conference\non Learning Representations (ICLR), 2016.\n[18] Nal Kalchbrenner, Lasse Espeholt, Karen Simonyan, Aaron van den Oord, Alex Graves, and Ko-\nray Kavukcuoglu. Neural machine translation in linear time. arXiv preprint arXiv:1610.10099v2,\n2017.\n[19] Yoon Kim, Carl Denton, Luong Hoang, and Alexander M. Rush. Structured attention networks.\nIn International Conference on Learning Representations, 2017.\n[20] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015.\n[21] Oleksii Kuchaiev and Boris Ginsburg. Factorization tricks for LSTM networks. arXiv preprint\narXiv:1703.10722, 2017.\n[22] Zhouhan Lin, Minwei Feng, Cicero Nogueira dos Santos, Mo Yu, Bing Xiang, Bowen\nZhou, and Yoshua Bengio. A structured self-attentive sentence embedding. arXiv preprint\narXiv:1703.03130, 2017.\n[23] Minh-Thang Luong, Quoc V. Le, Ilya Sutskever, Oriol Vinyals, and Lukasz Kaiser. Multi-task\nsequence to sequence learning. arXiv preprint arXiv:1511.06114, 2015.\n[24] Minh-Thang Luong, Hieu Pham, and Christopher D Manning. Effective approaches to attention-\nbased neural machine translation. arXiv preprint arXiv:1508.04025, 2015.\n11\n[25] Mitchell P Marcus, Mary Ann Marcinkiewicz, and Beatrice Santorini. Building a large annotated\ncorpus of english: The penn treebank. Computational linguistics, 19(2):313–330, 1993.\n[26] David McClosky, Eugene Charniak, and Mark Johnson. Effective self-training for parsing. In\nProceedings of the Human Language Technology Conference of the NAACL, Main Conference,\npages 152–159. ACL, June 2006.\n[27] Ankur Parikh, Oscar Täckström, Dipanjan Das, and Jakob Uszkoreit. A decomposable attention\nmodel. In Empirical Methods in Natural Language Processing, 2016.\n[28] Romain Paulus, Caiming Xiong, and Richard Socher. A deep reinforced model for abstractive\nsummarization. arXiv preprint arXiv:1705.04304, 2017.\n[29] Slav Petrov, Leon Barrett, Romain Thibaux, and Dan Klein. Learning accurate, compact,\nand interpretable tree annotation. In Proceedings of the 21st International Conference on\nComputational Linguistics and 44th Annual Meeting of the ACL, pages 433–440. ACL, July\n2006.\n[30] Ofir Press and Lior Wolf. Using the output embedding to improve language models. arXiv\npreprint arXiv:1608.05859, 2016.\n[31] Rico Sennrich, Barry Haddow, and Alexandra Birch. Neural machine translation of rare words\nwith subword units. arXiv preprint arXiv:1508.07909, 2015.\n[32] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton,\nand Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts\nlayer. arXiv preprint arXiv:1701.06538, 2017.\n[33] Nitish Srivastava, Geoffrey E Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdi-\nnov. Dropout: a simple way to prevent neural networks from overfitting. Journal of Machine\nLearning Research, 15(1):1929–1958, 2014.\n[34] Sainbayar Sukhbaatar, Arthur Szlam, Jason Weston, and Rob Fergus. End-to-end memory\nnetworks. In C. Cortes, N. D. Lawrence, D. D. Lee, M. Sugiyama, and R. Garnett, editors,\nAdvances in Neural Information Processing Systems 28, pages 2440–2448. Curran Associates,\nInc., 2015.\n[35] Ilya Sutskever, Oriol Vinyals, and Quoc VV Le. Sequence to sequence learning with neural\nnetworks. In Advances in Neural Information Processing Systems, pages 3104–3112, 2014.\n[36] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna.\nRethinking the inception architecture for computer vision. CoRR, abs/1512.00567, 2015.\n[37] Vinyals & Kaiser, Koo, Petrov, Sutskever, and Hinton. Grammar as a foreign language. In\nAdvances in Neural Information Processing Systems, 2015.\n[38] Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V Le, Mohammad Norouzi, Wolfgang\nMacherey, Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, et al. Google’s neural machine\ntranslation system: Bridging the gap between human and machine translation. arXiv preprint\narXiv:1609.08144, 2016.\n[39] Jie Zhou, Ying Cao, Xuguang Wang, Peng Li, and Wei Xu. Deep recurrent models with\nfast-forward connections for neural machine translation. CoRR, abs/1606.04199, 2016.\n[40] Muhua Zhu, Yue Zhang, Wenliang Chen, Min Zhang, and Jingbo Zhu. Fast and accurate\nshift-reduce constituent parsing. In Proceedings of the 51st Annual Meeting of the ACL (Volume\n1: Long Papers), pages 434–443. ACL, August 2013.\n12\nAttention Visualizations\nInput-Input Layer5\nIt\nis\nin\nthis\nspirit\nthat\na\nmajority\nof\nAmerican\ngovernments\nhave\npassed\nnew\nlaws\nsince\n2009\nmaking\nthe\nregistration\nor\nvoting\nprocess\nmore\ndifficult\n.\n<EOS>\n<pad>\n<pad>\n<pad>\n<pad>\n<pad>\n<pad>\nIt\nis\nin\nthis\nspirit\nthat\na\nmajority\nof\nAmerican\ngovernments\nhave\npassed\nnew\nlaws\nsince\n2009\nmaking\nthe\nregistration\nor\nvoting\nprocess\nmore\ndifficult\n.\n<EOS>\n<pad>\n<pad>\n<pad>\n<pad>\n<pad>\n<pad>\nFigure 3: An example of the attention mechanism following long-distance dependencies in the\nencoder self-attention in layer 5 of 6. Many of the attention heads attend to a distant dependency of\nthe verb ‘making’, completing the phrase ‘making...more difficult’. Attentions here shown only for\nthe word ‘making’. Different colors represent different heads. Best viewed in color.\n13\nInput-Input Layer5\nThe\nLaw\nwill\nnever\nbe\nperfect\n,\nbut\nits\napplication\nshould\nbe\njust\n-\nthis\nis\nwhat\nwe\nare\nmissing\n,\nin\nmy\nopinion\n.\n<EOS>\n<pad>\nThe\nLaw\nwill\nnever\nbe\nperfect\n,\nbut\nits\napplication\nshould\nbe\njust\n-\nthis\nis\nwhat\nwe\nare\nmissing\n,\nin\nmy\nopinion\n.\n<EOS>\n<pad>\nInput-Input Layer5\nThe\nLaw\nwill\nnever\nbe\nperfect\n,\nbut\nits\napplication\nshould\nbe\njust\n-\nthis\nis\nwhat\nwe\nare\nmissing\n,\nin\nmy\nopinion\n.\n<EOS>\n<pad>\nThe\nLaw\nwill\nnever\nbe\nperfect\n,\nbut\nits\napplication\nshould\nbe\njust\n-\nthis\nis\nwhat\nwe\nare\nmissing\n,\nin\nmy\nopinion\n.\n<EOS>\n<pad>\nFigure 4: Two attention heads, also in layer 5 of 6, apparently involved in anaphora resolution. Top:\nFull attentions for head 5. Bottom: Isolated attentions from just the word ‘its’ for attention heads 5\nand 6. Note that the attentions are very sharp for this word.\n14\nInput-Input Layer5\nThe\nLaw\nwill\nnever\nbe\nperfect\n,\nbut\nits\napplication\nshould\nbe\njust\n-\nthis\nis\nwhat\nwe\nare\nmissing\n,\nin\nmy\nopinion\n.\n<EOS>\n<pad>\nThe\nLaw\nwill\nnever\nbe\nperfect\n,\nbut\nits\napplication\nshould\nbe\njust\n-\nthis\nis\nwhat\nwe\nare\nmissing\n,\nin\nmy\nopinion\n.\n<EOS>\n<pad>\nInput-Input Layer5\nThe\nLaw\nwill\nnever\nbe\nperfect\n,\nbut\nits\napplication\nshould\nbe\njust\n-\nthis\nis\nwhat\nwe\nare\nmissing\n,\nin\nmy\nopinion\n.\n<EOS>\n<pad>\nThe\nLaw\nwill\nnever\nbe\nperfect\n,\nbut\nits\napplication\nshould\nbe\njust\n-\nthis\nis\nwhat\nwe\nare\nmissing\n,\nin\nmy\nopinion\n.\n<EOS>\n<pad>\nFigure 5: Many of the attention heads exhibit behaviour that seems related to the structure of the\nsentence. We give two such examples above, from two different heads from the encoder self-attention\nat layer 5 of 6. The heads clearly learned to perform different tasks.\n15\n', metadata={'Published': '2023-08-02', 'Title': 'Attention Is All You Need', 'Authors': 'Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, Illia Polosukhin', 'Summary': 'The dominant sequence transduction models are based on complex recurrent or\nconvolutional neural networks in an encoder-decoder configuration. The best\nperforming models also connect the encoder and decoder through an attention\nmechanism. We propose a new simple network architecture, the Transformer, based\nsolely on attention mechanisms, dispensing with recurrence and convolutions\nentirely. Experiments on two machine translation tasks show these models to be\nsuperior in quality while being more parallelizable and requiring significantly\nless time to train. Our model achieves 28.4 BLEU on the WMT 2014\nEnglish-to-German translation task, improving over the existing best results,\nincluding ensembles by over 2 BLEU. On the WMT 2014 English-to-French\ntranslation task, our model establishes a new single-model state-of-the-art\nBLEU score of 41.8 after training for 3.5 days on eight GPUs, a small fraction\nof the training costs of the best models from the literature. We show that the\nTransformer generalizes well to other tasks by applying it successfully to\nEnglish constituency parsing both with large and limited training data.'})]

Let’s check the metadata.

data[0].metadata

{'Published': '2023-08-02',
 'Title': 'Attention Is All You Need',
 'Authors': 'Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, Illia Polosukhin',
 'Summary': 'The dominant sequence transduction models are based on complex recurrent or\nconvolutional neural networks in an encoder-decoder configuration. The best\nperforming models also connect the encoder and decoder through an attention\nmechanism. We propose a new simple network architecture, the Transformer, based\nsolely on attention mechanisms, dispensing with recurrence and convolutions\nentirely. Experiments on two machine translation tasks show these models to be\nsuperior in quality while being more parallelizable and requiring significantly\nless time to train. Our model achieves 28.4 BLEU on the WMT 2014\nEnglish-to-German translation task, improving over the existing best results,\nincluding ensembles by over 2 BLEU. On the WMT 2014 English-to-French\ntranslation task, our model establishes a new single-model state-of-the-art\nBLEU score of 41.8 after training for 3.5 days on eight GPUs, a small fraction\nof the training costs of the best models from the literature. We show that the\nTransformer generalizes well to other tasks by applying it successfully to\nEnglish constituency parsing both with large and limited training data.'}

So, we have covered some document loaders in LangChain. There are more loaders which you can read about in LangChain documentation.

Right, now let’s use this knowledge that we have learned and create a bot that can answer questions based on Wikipedia articles. Sounds interesting? Let’s get our hands dirty!

First, let’s set up the environment and import all necessary libraries.

import os
from langchain_openai import ChatOpenAI #model
from langchain.globals import set_llm_cache #response caching
from langchain.cache import InMemoryCache
from langchain.prompts import HumanMessagePromptTemplate, ChatPromptTemplate #prompt templating
from langchain.document_loaders import WikipediaLoader #document loader

import os
with open('../../openai_api_key.txt') as f:
    api_key = f.read()
os.environ['OPENAI_API_KEY'] = api_key

chat = ChatOpenAI()
set_llm_cache(InMemoryCache())

Let’s create a function that takes in a topic and a question and returns a LLM response.

def qa_bot(topic, question):
    chat = ChatOpenAI()
    loader = WikipediaLoader(query = topic, load_max_docs=1)
    data = loader.load()
    
    document = data[0]
    title = document.metadata['title']
    summary = document.metadata['summary']
    
    question = question

    human_template = "Read about the wikipedia article on {title} having content {content} and answer the {question}"
    
    human_message_prompt= HumanMessagePromptTemplate.from_template(human_template)
    
    chat_prompt = ChatPromptTemplate.from_messages([human_message_prompt])
    
    prompt = chat_prompt.format_prompt(title = title, content = summary, question = question)
    
    response = chat(messages = prompt.to_messages())
    
    return response.content

Here, we are first passing the wikipedia article to the loader and extracting metadata: title and summary. Instead of summary we can also use page_content. I have used summary because of the memory constraints my system. These will be passed onto the LLM. Next we create a prompt template where we pass some instructions to the model as well the input. I have covered prompt templates in my first series of Langchain articles. You can go through it in case you some doubts. Let’s test this.

qa_bot('Artificial Intelligence', 'When was AI born?')

'AI was founded as an academic discipline in 1956.'

So, this is how we just built a question answering bot that can answer questions based on the Wikipedia article.

That brings this article to an end. Hope this read was refreshing and worth it. I’ll see you in the next one.

Thanks for your time. I have added all the code in my Github repo. Keep learning!

References:

Document loaders | 🦜️🔗 Langchain

Efficient Few-Shot Prompting in LangChain: Response Caching, Prompt Templating and Prompt…