DOCSLIB.ORG

Arxiv:2004.03720V2 [Cs.CL] 5 Oct 2020

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
Arxiv:2004.03720V2 [Cs.CL] 5 Oct 2020 Byte Pair Encoding is Suboptimal for Language Model Pretraining Kaj Bostrom and Greg Durrett Department of Computer Science The University of Texas at Austin fkaj,[email protected] Abstract levels of granularity, from individual characters to full words. As a result, rare words are broken down The success of pretrained transformer lan- into a collection of subword units, bottoming out guage models (LMs) in natural language in characters in the worst case. processing has led to a wide range of pretraining setups. In particular, these models Critically, a pretrained language model’s sub- employ a variety of subword tokenization word vocabulary cannot be altered: any down- methods, most notably byte-pair encoding stream application of these models must tokenize (BPE) (Sennrich et al., 2016; Gage, 1994), the input or generate output using the original subword WordPiece method (Schuster and Nakajima, vocabulary, making the choice of tokenization a 2012), and unigram language modeling (Kudo, particularly significant decision. 2018), to segment text. However, to the A variety of subword tokenization methods have best of our knowledge, the literature does not seen use in pretrained language models. BERT contain a direct evaluation of the impact of tokenization on language model pretraining. uses the WordPiece method (Schuster and Naka- We analyze differences between BPE and un- jima, 2012), a language-modeling based variant of igram LM tokenization, finding that the latter BPE; T5 (Raffel et al., 2019) uses character-level method recovers subword units that align more BPE; GPT2 (Radford et al., 2019) and ROBERTA closely with morphology and avoids problems (Liu et al., 2019) use BPE over raw bytes instead stemming from BPE’s greedy construction of unicode characters; XLNET (Yang et al., 2019) procedure. We then compare the fine-tuned and ALBERT (Lan et al., 2019) use the Sentence- task performance of identical transformer masked language models pretrained with these Piece library (Kudo and Richardson, 2018) which tokenizations. Across downstream tasks and implements both BPE and unigram language model two languages (English and Japanese), we tokenization, but in both cases fail to clarify which find that the unigram LM tokenization method of these methods they chose. The effects of tok- matches or outperforms BPE. We hope that enization are not examined in a reported experi- developers of future pretrained LMs will ment in any of the above works except Liu et al. consider adopting the unigram LM method (2019), who note that WordPiece gave a small ad- over the more prevalent BPE. vantage over BPE in their preliminary investigation. In the machine translation literature, Kudo(2018) arXiv:2004.03720v2 [cs.CL] 5 Oct 2020 1 Introduction introduced the unigram language model tokeniza- Large transformers (Vaswani et al., 2017) pre- tion method in the context of machine translation trained with variants of a language modeling ob- and found it comparable in performance to BPE. jective, such as BERT (Devlin et al., 2019), have Domingo et al.(2018) performed further experi- proven their effectiveness at flexibly transferring to ments to investigate the effects of tokenization on a variety of domains and tasks. One design deci- neural machine translation, but used a shared BPE sion that makes them particularly adaptable is their vocabulary across all experiments. Galle´(2019) graceful handling of the open vocabulary problem examined algorithms in the BPE family, but did not through subword tokenization. Subword tokeniza- compare to unigram language modeling. tion, popularized in the neural machine translation In this work, we characterize the space of pro- literature (Sennrich et al., 2016; Vaswani et al., posed subword tokenization algorithms and ana- 2017; Wu et al., 2016), produces tokens at multiple lyze the differences between the two methods with publicly available implementations: BPE (merg- taining ordered merges and applies them to new ing tokens based on bigram frequency) and uni- text in the same order as they occurred during vo- gram language modeling (pruning tokens based on cabulary construction. unigram LM perplexity). While the vocabularies The WordPiece algorithm (Schuster and Naka- resulting from these schemes are heavily overlap- jima, 2012), used to construct BERT’s vocabulary, ping, we compare each method to reference mor- closely resembles BPE. However, instead of merg- phological segmentations and find that the unigram ing the most frequent token bigram, each poten- LM method produces tokens better aligned with tial merge is scored based on the likelihood of an morphology. To understand whether this more nat- n-gram language model trained on a version of ural tokenization leads to improved performance, the corpus incorporating that merge. Schuster and we pretrain separate language models using the Nakajima(2012) note that the process of estimat- ROBERTA objective (Liu et al., 2019) with each ing language model parameters for every potential tokenization for both English and Japanese, two merge is prohibitive, so they employ aggressive typologically distant languages. On downstream heuristics to reduce the number of potential merges tasks, we find a performance gap across tasks and considered. As their implementation is not public,1 languages, with the unigram LM method provid- we are unable to make a comparison to this method. ing an improvement over BPE of up to 10% in The unigram LM method (Kudo, 2018), in con- our Japanese QA experiments, indicating the ben- trast to the bottom-up construction process of BPE efits of adopting this technique in the context of and WordPiece, begins with a superset of the final language model pretraining. vocabulary, pruning it to the desired size: 2 Algorithms Algorithm 2 Unigram LM (Kudo, 2018) Subword tokenization algorithms consist of two 1: Input: set of strings D, target vocab size k components: a vocabulary construction procedure, 2: procedure UNIGRAMLM(D; k) which takes a corpus of text and returns a vocabu- 3: V all substrings occurring more than lary with the desired size, and a tokenization proce- 4: once in D (not crossing words) dure, which takes the built vocabulary and applies it 5: while jV j > k do . Prune tokens to new text, returning a sequence of tokens. In the- 6: Fit unigram LM θ to D ory, these two steps can be independent, although 7: for t 2 V do . Estimate token ‘loss’ for the algorithms we examine the tokenization 8: Lt pθ(D) − pθ0 (D) 0 procedure is tightly coupled to the vocabulary con- 9: where θ is the LM without token t struction procedure. 10: end for A BPE vocabulary is constructed as follows: 11: Remove min(jV j − k; bαjV jc) of the 12: tokens t with highest Lt from V , Algorithm 1 Byte-pair encoding (Sennrich et al., 13: where α 2 [0; 1] is a hyperparameter 2016; Gage, 1994) 14: end while 15: Fit final unigram LM θ to D 1: Input: set of strings D, target vocab size k 16: return V; θ 2: procedure BPE(D; k) 17: end procedure 3: V all unique characters in D 4: (about 4,000 in English Wikipedia) 5: while jV j < k do . Merge tokens Unigram LM tokenization takes the vocabulary 6: tL; tR Most frequent bigram in D V and unigram LM parameters θ and performs Viterbi inference to decode the segmentation with 7: tNEW tL + tR . Make new token θ 8: V V + [tNEW] maximum likelihood under . This method is 9: Replace each occurrence of tL; tR in similar to Morfessor’s unsupervised segmentation 10: D with tNEW (Creutz and Lagus, 2005) without its informed prior 11: end while over token length. 12: return V 1Although its name and association with Google might sug- 13: end procedure gest otherwise, the SentencePiece library (Kudo and Richard- son, 2018) does not, in fact, implement the WordPiece algo- rithm; it provides implementations of BPE and unigram LM BPE tokenization takes the vocabulary V con- based tokenization. Original: furiously Original: tricycles Original: nanotechnology BPE: fur iously BPE: t ric y cles BPE: n an ote chn ology Uni. LM: fur ious ly Uni. LM: tri cycle s Uni. LM: nano technology Original: Completely preposterous suggestions BPE: Comple t ely prep ost erous suggest ions Unigram LM: Complete ly pre post er ous suggestion s Original: corrupted Original: 1848 and 1852, BPE: cor rupted BPE: 184 8 and 185 2, Unigram LM: corrupt ed Unigram LM: 1848 and 1852 , Original n性/様々+分Cがなされている。 BPE n 性/ 様々 +分C がなされている 。 Unigram LM n 性 / 様々 + 分C がなされている 。 Gloss magnetism (top.) various ways in classification is done . Translation Magnetism is classified in various ways. Figure 1: Example tokenizations. The character ‘ ’ is a word boundary marker. BPE merges common tokens, such as English inflectional suffixes and Japanese particles, into their neighbors even when the resulting unit is not semantically meaningful. 7 10 4000 BPE BPE 6 Unigram LM 10 Unigram LM 3000 5 10 2000 4 10 3 1000 Token frequency Number of tokens 10 2 0 10 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 1 20000 Token length Token frequency rank (a) Token length distributions within each vocabulary (b) Token frequency profiles over the corpus Figure 2: English subword vocabulary and corpus profiles. The unigram LM method produces longer tokens on average (a) and uses its vocabulary space more effectively (b), with more tokens of moderate frequency. In the course of our experiments we did not ob- More frequent in serve a major difference in speed between the two BPE Unigram LM algorithms. Both require similar amounts of time to H L M T B s . , ed d construct a vocabulary, and both have a negligible P C K D R ing e ly t a impact on overall model inference latency. Table 1: Tokens with the highest difference in fre- quency between tokenizations.
Load more

Recommended publications
  • Evaluation of Machine Learning Algorithms for Sms Spam Filtering
    EVALUATION OF MACHINE LEARNING ALGORITHMS FOR SMS SPAM FILTERING David Bäckman Bachelor Thesis, 15 credits Bachelor Of Science Programme in Computing Science 2019 Abstract The purpose of this thesis is to evaluate dierent machine learning algorithms and methods for text representation in order to determine what is best suited to use to distinguish between spam SMS and legitimate SMS. A data set that contains 5573 real SMS has been used to train the algorithms K-Nearest Neighbor, Support Vector Machine, Naive Bayes and Logistic Regression. The dierent methods that have been used to represent text are Bag of Words, Bigram and Word2Vec. In particular, it has been investigated if semantic text representations can improve the performance of classication. A total of 12 combinations have been evaluated with help of the metrics accuracy and F1-score. The results shows that Logistic Regression together with Bag of Words reach the highest accuracy and F1-score. Bigram as text representation seems to work worse then the others methods. Word2Vec can increase the performnce for K- Nearst Neigbor but not for the other algorithms. Acknowledgements I would like to thank my supervisor Kai-Florian Richter for all good advice and guidance through the project. I would also like to thank all my classmates for help and support during the education, you have made it possible for me to reach this day. Contents 1 Introduction 1 1.1 Background1 1.2 Purpose and Research Questions1 2 Related Work 3 3 Theoretical Background 5 3.1 The Choice of Algorithms5 3.2 Classication
    [Show full text]
  • NLP - Assignment 2
    NLP - Assignment 2 Week 2 December 27th, 2016 1. A 5-gram model is a order Markov Model: (a) Six (b) Five (c) Four (d) Constant Ans : c) Four 2. For the following corpus C1 of 3 sentences, what is the total count of unique bi- grams for which the likelihood will be estimated? Assume we do not perform any pre-processing, and we are using the corpus as given. (i) ice cream tastes better than any other food (ii) ice cream is generally served after the meal (iii) many of us have happy childhood memories linked to ice cream (a) 22 (b) 27 (c) 30 (d) 34 Ans : b) 27 3. Arrange the words \curry, oil and tea" in descending order, based on the frequency of their occurrence in the Google Books n-grams. The Google Books n-gram viewer is available at https://books.google.com/ngrams: (a) tea, oil, curry (c) curry, tea, oil (b) curry, oil, tea (d) oil, tea, curry Ans: d) oil, tea, curry 4. Given a corpus C2, The Maximum Likelihood Estimation (MLE) for the bigram \ice cream" is 0.4 and the count of occurrence of the word \ice" is 310. The likelihood of \ice cream" after applying add-one smoothing is 0:025, for the same corpus C2. What is the vocabulary size of C2: 1 (a) 4390 (b) 4690 (c) 5270 (d) 5550 Ans: b)4690 The Questions from 5 to 10 require you to analyse the data given in the corpus C3, using a programming language of your choice.
    [Show full text]
  • 3 Dictionaries and Tolerant Retrieval
    Online edition (c)2009 Cambridge UP DRAFT! © April 1, 2009 Cambridge University Press. Feedback welcome. 49 Dictionaries and tolerant 3 retrieval In Chapters 1 and 2 we developed the ideas underlying inverted indexes for handling Boolean and proximity queries. Here, we develop techniques that are robust to typographical errors in the query, as well as alternative spellings. In Section 3.1 we develop data structures that help the search for terms in the vocabulary in an inverted index. In Section 3.2 we study WILDCARD QUERY the idea of a wildcard query: a query such as *a*e*i*o*u*, which seeks doc- uments containing any term that includes all the five vowels in sequence. The * symbol indicates any (possibly empty) string of characters. Users pose such queries to a search engine when they are uncertain about how to spell a query term, or seek documents containing variants of a query term; for in- stance, the query automat* would seek documents containing any of the terms automatic, automation and automated. We then turn to other forms of imprecisely posed queries, focusing on spelling errors in Section 3.3. Users make spelling errors either by accident, or because the term they are searching for (e.g., Herman) has no unambiguous spelling in the collection. We detail a number of techniques for correcting spelling errors in queries, one term at a time as well as for an entire string of query terms. Finally, in Section 3.4 we study a method for seeking vo- cabulary terms that are phonetically close to the query term(s).
    [Show full text]
  • N-Gram-Based Machine Translation
    N-gram-based Machine Translation ∗ JoseB.Mari´ no˜ ∗ Rafael E. Banchs ∗ Josep M. Crego ∗ Adria` de Gispert ∗ Patrik Lambert ∗ Jose´ A. R. Fonollosa ∗ Marta R. Costa-jussa` Universitat Politecnica` de Catalunya This article describes in detail an n-gram approach to statistical machine translation. This ap- proach consists of a log-linear combination of a translation model based on n-grams of bilingual units, which are referred to as tuples, along with four specific feature functions. Translation performance, which happens to be in the state of the art, is demonstrated with Spanish-to-English and English-to-Spanish translations of the European Parliament Plenary Sessions (EPPS). 1. Introduction The beginnings of statistical machine translation (SMT) can be traced back to the early fifties, closely related to the ideas from which information theory arose (Shannon and Weaver 1949) and inspired by works on cryptography (Shannon 1949, 1951) during World War II. According to this view, machine translation was conceived as the problem of finding a sentence by decoding a given “encrypted” version of it (Weaver 1955). Although the idea seemed very feasible, enthusiasm faded shortly afterward because of the computational limitations of the time (Hutchins 1986). Finally, during the nineties, two factors made it possible for SMT to become an actual and practical technology: first, significant increment in both the computational power and storage capacity of computers, and second, the availability of large volumes of bilingual data. The first SMT systems were developed in the early nineties (Brown et al. 1990, 1993). These systems were based on the so-called noisy channel approach, which models the probability of a target language sentence T given a source language sentence S as the product of a translation-model probability p(S|T), which accounts for adequacy of trans- lation contents, times a target language probability p(T), which accounts for fluency of target constructions.
    [Show full text]
  • Title and Link Description Improving Distributional Similarity With
    title and link description Improving Distributional Similarity • word similarity with Lessons Learned from Word ... • analogy • qualitativ: compare neighborhoods across Dependency-Based Word Embeddings embeddings • word similarity • word similarity Multi-Granularity Chinese Word • analogy Embedding • qualitative: local neighborhoods A Mixture Model for Learning Multi- • word similarity Sense Word Embeddings • analogy Learning Crosslingual Word • multilingual Embeddings without Bilingual Corpora • word similarities (monolingual) Word Embeddings with Limited • word similarity Memory • phrase similarity A Latent Variable Model Approach to • analogy PMI-based Word Embeddings Prepositional Phrase Attachment over • extrinsic Word Embedding Products A Simple Word Embedding Model for • lexical substitution (nearest Lexical Substitution neighbors after vector averaging) Segmentation-Free Word Embedding • extrinsic for Unsegmented Languages • word similarity Evaluation methods for unsupervised • analogy word embeddings • concept categorization • selectional preference Dict2vec : Learning Word Embeddings • word similarity using Lexical Dictionaries • extrinsic Task-Oriented Learning of Word • extrinsic Embeddings for Semantic Relation ... Refining Word Embeddings for • extrinsic Sentiment Analysis Language classification from bilingual • word similarity word embedding graphs Learning principled bilingual mappings • multilingual of word embeddings while ... A Word Embedding Approach to • extrinsic Identifying Verb-Noun Idiomatic ... Deep Multilingual
    [Show full text]
  • Arxiv:2006.14799V2 [Cs.CL] 18 May 2021 the Same User Input
    Evaluation of Text Generation: A Survey Evaluation of Text Generation: A Survey Asli Celikyilmaz [email protected] Facebook AI Research Elizabeth Clark [email protected] University of Washington Jianfeng Gao [email protected] Microsoft Research Abstract The paper surveys evaluation methods of natural language generation (NLG) systems that have been developed in the last few years. We group NLG evaluation methods into three categories: (1) human-centric evaluation metrics, (2) automatic metrics that require no training, and (3) machine-learned metrics. For each category, we discuss the progress that has been made and the challenges still being faced, with a focus on the evaluation of recently proposed NLG tasks and neural NLG models. We then present two examples for task-specific NLG evaluations for automatic text summarization and long text generation, and conclude the paper by proposing future research directions.1 1. Introduction Natural language generation (NLG), a sub-field of natural language processing (NLP), deals with building software systems that can produce coherent and readable text (Reiter & Dale, 2000a) NLG is commonly considered a general term which encompasses a wide range of tasks that take a form of input (e.g., a structured input like a dataset or a table, a natural language prompt or even an image) and output a sequence of text that is coherent and understandable by humans. Hence, the field of NLG can be applied to a broad range of NLP tasks, such as generating responses to user questions in a chatbot, translating a sentence or a document from one language into another, offering suggestions to help write a story, or generating summaries of time-intensive data analysis.
    [Show full text]
  • Word Senses and Wordnet Lady Bracknell
    Speech and Language Processing. Daniel Jurafsky & James H. Martin. Copyright © 2021. All rights reserved. Draft of September 21, 2021. CHAPTER 18 Word Senses and WordNet Lady Bracknell. Are your parents living? Jack. I have lost both my parents. Lady Bracknell. To lose one parent, Mr. Worthing, may be regarded as a misfortune; to lose both looks like carelessness. Oscar Wilde, The Importance of Being Earnest ambiguous Words are ambiguous: the same word can be used to mean different things. In Chapter 6 we saw that the word “mouse” has (at least) two meanings: (1) a small rodent, or (2) a hand-operated device to control a cursor. The word “bank” can mean: (1) a financial institution or (2) a sloping mound. In the quote above from his play The Importance of Being Earnest, Oscar Wilde plays with two meanings of “lose” (to misplace an object, and to suffer the death of a close person). We say that the words ‘mouse’ or ‘bank’ are polysemous (from Greek ‘having word sense many senses’, poly- ‘many’ + sema, ‘sign, mark’).1 A sense (or word sense) is a discrete representation of one aspect of the meaning of a word. In this chapter WordNet we discuss word senses in more detail and introduce WordNet, a large online the- saurus —a database that represents word senses—with versions in many languages. WordNet also represents relations between senses. For example, there is an IS-A relation between dog and mammal (a dog is a kind of mammal) and a part-whole relation between engine and car (an engine is a part of a car).
    [Show full text]
  • A Unigram Orientation Model for Statistical Machine Translation
    A Unigram Orientation Model for Statistical Machine Translation Christoph Tillmann IBM T.J. Watson Research Center Yorktown Heights, NY 10598 [email protected] Abstract b5 airspace In this paper, we present a unigram segmen- b4 tation model for statistical machine transla- Lebanese b tion where the segmentation units are blocks: 3 pairs of phrases without internal structure. The violate b2 segmentation model uses a novel orientation warplanes component to handle swapping of neighbor b1 blocks. During training, we collect block un- Israeli igram counts with orientation: we count how A A A t A A A often a block occurs to the left or to the right of l l l n l l l some predecessor block. The orientation model T H A t m j l is shown to improve translation performance A r s h j w b } over two models: 1) no block re-ordering is b r k A y n r y A l A used, and 2) the block swapping is controlled A P } n only by a language model. We show exper- t y y imental results on a standard Arabic-English l y translation task. P 1 Introduction Figure 1: An Arabic-English block translation example taken from the devtest set. The Arabic words are roman- In recent years, phrase-based systems for statistical ma- ized. chine translation (Och et al., 1999; Koehn et al., 2003; Venugopal et al., 2003) have delivered state-of-the-art performance on standard translation tasks. In this pa- section 3, block swapping where only a trigram language per, we present a phrase-based unigram system similar model is used to compute probabilities between neighbor to the one in (Tillmann and Xia, 2003), which is ex- blocks fails to improve translation performance.
    [Show full text]
  • Heafield, K. and A. Lavie. "Voting on N-Grams for Machine Translation
    Voting on N-grams for Machine Translation System Combination Kenneth Heafield and Alon Lavie Language Technologies Institute Carnegie Mellon University 5000 Forbes Avenue Pittsburgh, PA 15213 fheafield,[email protected] Abstract by a single weight and further current features only guide decisions at the word level. To remedy this System combination exploits differences be- situation, we propose new features that account for tween machine translation systems to form a multiword behavior, each with a separate set of sys- combined translation from several system out- tem weights. puts. Core to this process are features that reward n-gram matches between a candidate 2 Features combination and each system output. Systems differ in performance at the n-gram level de- Most combination schemes generate many hypothe- spite similar overall scores. We therefore ad- sis combinations, score them using a battery of fea- vocate a new feature formulation: for each tures, and search for a hypothesis with highest score system and each small n, a feature counts to output. Formally, the system generates hypoth- n-gram matches between the system and can- esis h, evaluates feature function f, and multiplies didate. We show post-evaluation improvement of 6.67 BLEU over the best system on NIST linear weight vector λ by feature vector f(h) to ob- T MT09 Arabic-English test data. Compared to tain score λ f(h). The score is used to rank fi- a baseline system combination scheme from nal hypotheses and to prune partial hypotheses dur- WMT 2009, we show improvement in the ing beam search. The beams contain hypotheses of range of 1 BLEU point.
    [Show full text]
  • Part I Word Vectors I: Introduction, Svd and Word2vec 2 Natural Language in Order to Perform Some Task
    CS224n: Natural Language Processing with Deep 1 Learning 1 Course Instructors: Christopher Lecture Notes: Part I Manning, Richard Socher 2 Word Vectors I: Introduction, SVD and Word2Vec 2 Authors: Francois Chaubard, Michael Fang, Guillaume Genthial, Rohit Winter 2019 Mundra, Richard Socher Keyphrases: Natural Language Processing. Word Vectors. Singu- lar Value Decomposition. Skip-gram. Continuous Bag of Words (CBOW). Negative Sampling. Hierarchical Softmax. Word2Vec. This set of notes begins by introducing the concept of Natural Language Processing (NLP) and the problems NLP faces today. We then move forward to discuss the concept of representing words as numeric vectors. Lastly, we discuss popular approaches to designing word vectors. 1 Introduction to Natural Language Processing We begin with a general discussion of what is NLP. 1.1 What is so special about NLP? What’s so special about human (natural) language? Human language is a system specifically constructed to convey meaning, and is not produced by a physical manifestation of any kind. In that way, it is very different from vision or any other machine learning task. Natural language is a dis- Most words are just symbols for an extra-linguistic entity : the crete/symbolic/categorical system word is a signifier that maps to a signified (idea or thing). For instance, the word "rocket" refers to the concept of a rocket, and by extension can designate an instance of a rocket. There are some exceptions, when we use words and letters for expressive sig- naling, like in "Whooompaa". On top of this, the symbols of language can be encoded in several modalities : voice, gesture, writing, etc that are transmitted via continuous signals to the brain, which itself appears to encode things in a continuous manner.
    [Show full text]
  • Phrase Based Language Model for Statistical Machine Translation
    Phrase Based Language Model for Statistical Machine Translation Technical Report by Chen, Geliang 1 Abstract Reordering is a challenge to machine translation (MT) systems. In MT, the widely used approach is to apply word based language model (LM) which considers the constituent units of a sentence as words. In speech recognition (SR), some phrase based LM have been proposed. However, those LMs are not necessarily suitable or optimal for reordering. We propose two phrase based LMs which considers the constituent units of a sentence as phrases. Experiments show that our phrase based LMs outperform the word based LM with the respect of perplexity and n-best list re-ranking. Key words: machine translation, language model, phrase based This version of report is identical to the dissertation approved from Yuanpei Institute of Peking University to receive the Bachelor Degree of Probability and Statistics by Chen, Geliang. Advisor: Assistant Professor Dr.rer.nat. Jia Xu Evaluation Date: 20. June. 2013 2 Contents Abstract .......................................................................................................................... 2 Contents …………………………………………………………………………………………………………………..3 1 Introduction ................................................................................................................. 4 2 Review of the Word Based LM ................................................................................... 5 2.1 Sentence probability ............................................................................................
    [Show full text]
  • Phrase-Based Attentions
    Under review as a conference paper at ICLR 2019 PHRASE-BASED ATTENTIONS Phi Xuan Nguyen, Shafiq Joty Nanyang Technological University, Singapore [email protected], [email protected] ABSTRACT Most state-of-the-art neural machine translation systems, despite being different in architectural skeletons (e.g., recurrence, convolutional), share an indispensable feature: the Attention. However, most existing attention methods are token-based and ignore the importance of phrasal alignments, the key ingredient for the suc- cess of phrase-based statistical machine translation. In this paper, we propose novel phrase-based attention methods to model n-grams of tokens as attention entities. We incorporate our phrase-based attentions into the recently proposed Transformer network, and demonstrate that our approach yields improvements of 1.3 BLEU for English-to-German and 0.5 BLEU for German-to-English transla- tion tasks on WMT newstest2014 using WMT’16 training data. 1 INTRODUCTION Neural Machine Translation (NMT) has established breakthroughs in many different translation tasks, and has quickly become the standard approach to machine translation. NMT offers a sim- ple encoder-decoder architecture that is trained end-to-end. Most NMT models (except a few like (Kaiser & Bengio, 2016) and (Huang et al., 2018)) possess attention mechanisms to perform align- ments of the target tokens to the source tokens. The attention module plays a role analogous to the word alignment model in Statistical Machine Translation or SMT (Koehn, 2010). In fact, the trans- former network introduced recently by Vaswani et al. (2017) achieves state-of-the-art performance in both speed and BLEU scores (Papineni et al., 2002) by using only attention modules.
    [Show full text]