DOCSLIB.ORG

Using the Output Embedding to Improve Language Models

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
Using the Output Embedding to Improve Language Models Using the Output Embedding to Improve Language Models Ofir Press and Lior Wolf School of Computer Science Tel-Aviv University, Israel ofir.press,wolf @cs.tau.ac.il { } Abstract results in an activation vector h2 that similarly to H U >c, is also in IR . In this case, U and V are of We study the topmost weight matrix of exactly the same size. neural network language models. We show that this matrix constitutes a valid word embedding. When training language We call U the input embedding, and V the out- models, we recommend tying the input put embedding. In both matrices, we expect rows embedding and this output embedding. that correspond to similar words to be similar: for We analyze the resulting update rules and the input embedding, we would like the network show that the tied embedding evolves in to react similarly to synonyms, while in the out- a more similar way to the output embed- put embedding, we would like the scores of words ding than to the input embedding in the that are interchangeable to be similar (Mnih and untied model. We also offer a new method Teh, 2012). of regularizing the output embedding. Our methods lead to a significant reduction in While U and V can both serve as word embed- perplexity, as we are able to show on a va- dings, in the literature, only the former serves this riety of neural network language models. role. In this paper, we compare the quality of the Finally, we show that weight tying can re- input embedding to that of the output embedding, duce the size of neural translation models and we show that the latter can be used to improve to less than half of their original size with- neural network language models. Our main results out harming their performance. are as follows: (i) We show that in the word2vec skip-gram model, the output embedding is only 1 Introduction slightly inferior to the input embedding. This is In a common family of neural network language shown using metrics that are commonly used in or- models, the current input word is represented as der to measure embedding quality. (ii) In recurrent the vector c IRC and is projected to a dense neural network based language models, the output ∈ representation using a word embedding matrix U. embedding outperforms the input embedding. (iii) Some computation is then performed on the word By tying the two embeddings together, i.e., enforc- embedding U >c, which results in a vector of ac- ing U = V , the joint embedding evolves in a more tivations h2. A second matrix V then projects h2 similar way to the output embedding than to the in- to a vector h3 containing one score per vocabulary put embedding of the untied model. (iv) Tying the word: h3 = V h2. The vector of scores is then con- input and output embeddings leads to an improve- verted to a vector of probability values p, which ment in the perplexity of various language mod- represents the models’ prediction of the next word, els. This is true both when using dropout or when using the softmax function. not using it. (v) When not using dropout, we pro- For example, in the LSTM-based language pose adding an additional projection P before V , models of (Sundermeyer et al., 2012; Zaremba and apply regularization to P . (vi) Weight tying et al., 2014), for vocabulary of size C, the one- in neural translation models can reduce their size hot encoding is used to represent the input c and (number of parameters) to less than half of their C H U IR × . An LSTM is then employed, which original size without harming their performance. ∈ 157 Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 2, Short Papers, pages 157–163, Valencia, Spain, April 3-7, 2017. c 2017 Association for Computational Linguistics 2 Related Work on the model’s performance was not tested. In- dependently and concurrently with our work (Inan Neural network language models (NNLMs) assign et al., 2016) presented an explanation for weight probabilities to word sequences. Their resurgence tying in NNLMs based on (Hinton et al., 2015). was initiated by (Bengio et al., 2003). Recur- rent neural networks were first used for language 3 Weight Tying modeling in (Mikolov et al., 2010) and (Pascanu et al., 2013). The first model that implemented In this work, we employ three different model cat- language modeling with LSTMs (Hochreiter and egories: NNLMs, the word2vec skip-gram model, Schmidhuber, 1997) was (Sundermeyer et al., and NMT models. Weight tying is applied sim- 2012). Following that, (Zaremba et al., 2014) in- ilarly in all models. For translation models, we troduced a dropout (Srivastava, 2013) augmented also present a three-way weight tying method. NNLM. (Gal, 2015; Gal and Ghahramani, 2016) NNLM models contain an input embedding ma- proposed a new dropout method, which is referred trix, two LSTM layers (h1 and h2), a third hidden to as Bayesian Dropout below, that improves on scores/logits layer h3, and a softmax layer. The the results of (Zaremba et al., 2014). loss used during training is the cross entropy loss The skip-gram word2vec model introduced without any regularization terms. in (Mikolov et al., 2013a; Mikolov et al., 2013b) Following (Zaremba et al., 2014), we employ learns representations of words. This model learns two models: large and small. The large model em- a representation for each word in its vocabulary, ploys dropout for regularization. The small model both in an input embedding matrix and in an out- is not regularized. Therefore, we propose the fol- lowing regularization scheme. A projection matrix put embedding matrix. When training is com- H H P IR × is inserted before the output embed- plete, the vectors that are returned are the input ∈ embeddings. The output embedding is typically ding, i.e., h3 = V P h2. The regularizing term λ P is then added to the small model’s loss ignored, although (Mitra et al., 2016; Mnih and k k2 Kavukcuoglu, 2013) use both the output and input function. In all of our experiments, λ = 0.15. embeddings of words in order to compute word Projection regularization allows us to use the similarity. Recently, (Goldberg and Levy, 2014) same embedding (as both the input/output embed- argued that the output embedding of the word2vec ding) with some adaptation that is under regular- skip-gram model needs to be different than the in- ization. It is, therefore, especially suited for WT. put embedding. While training a vanilla untied NNLM, at As we show, tying the input and the output em- timestep t, with current input word sequence beddings is indeed detrimental in word2vec. How- i1:t = [i1, ..., it] and current target output word ever, it improves performance in NNLMs. ot, the negative log likelihood loss is given by: = log p (o i ) p (o i ) = In neural machine translation (NMT) mod- t t t 1:t , where t t 1:t L exp (V −h(t)) | | o>t 2 els (Kalchbrenner and Blunsom, 2013; Cho et PC (t) , Uk (Vk) is the kth row of U (V ), x=1 exp(Vx>h2 ) al., 2014; Sutskever et al., 2014; Bahdanau et (t) al., 2014), the decoder, which generates the trans- which corresponds to word k, and h2 is the vector lation of the input sentence in the target lan- of activations of the topmost LSTM layer’s output guage, is a language model that is conditioned on at time t. For simplicity, we assume that at each timestep t, i = o . Optimization of the model is both the previous words of the output sentence t 6 t and on the source sentence. State of the art re- performed using stochastic gradient descent. sults in NMT have recently been achieved by sys- The update for row k of the input embedding is: tems that segment the source and target words ( (t) PC ∂h2 ∂ t ( x=1 pt(x i1:t) Vx> Vo>t ) ∂U k = it into subword units (Sennrich et al., 2016a). One L = | · − it ∂U k 0 k = it such method (Sennrich et al., 2016b) is based on 6 the byte pair encoding (BPE) compression algo- For the output embedding, row k’s update is: ( (t) rithm (Gage, 1994). BPE segments rare words into ∂ t (pt(ot i1:t) 1)h2 k = ot L = | −(t) ∂Vk pt(k i1:t) h k = ot their more commonly appearing subwords. | · 2 6 Weight tying was previously used in the log- Therefore, in the untied model, at every timestep, bilinear model of (Mnih and Hinton, 2009), but the the only row that is updated in the input embed- decision to use it was not explained, and its effect ding is the row Uit representing the current input 158 word. This means that vectors representing rare Language Subwords Subwords Subwords pairs only in source only in target in both words are updated only a small number of times. EN FR 2K 7K 85K The output embedding updates every row at each EN→DE 3K 11K 80K → timestep. In tied NNLMs, we set U = V = S. The Table 1: Shared BPE subwords between pairs of languages. update for each row in S is the sum of the updates obtained for the two roles of S as both an input and previous state at each timestep. ct is the context output embedding. vector at timestep t, ct = j r atjhj, where atj The update for row k = i is similar to the up- ∈ 6 t is the weight given to the jth annotation at time t: date of row k in the untied NNLM’s output embed- exp(etj ) P atj = P exp(e ) , and etj = at(hj), where a is ding (the only difference being that U and V are k r ik the alignment∈ model.
Load more

Recommended publications
  • Knowledge-Powered Deep Learning for Word Embedding
    Knowledge-Powered Deep Learning for Word Embedding Jiang Bian, Bin Gao, and Tie-Yan Liu Microsoft Research {jibian,bingao,tyliu}@microsoft.com Abstract. The basis of applying deep learning to solve natural language process- ing tasks is to obtain high-quality distributed representations of words, i.e., word embeddings, from large amounts of text data. However, text itself usually con- tains incomplete and ambiguous information, which makes necessity to leverage extra knowledge to understand it. Fortunately, text itself already contains well- defined morphological and syntactic knowledge; moreover, the large amount of texts on the Web enable the extraction of plenty of semantic knowledge. There- fore, it makes sense to design novel deep learning algorithms and systems in order to leverage the above knowledge to compute more effective word embed- dings. In this paper, we conduct an empirical study on the capacity of leveraging morphological, syntactic, and semantic knowledge to achieve high-quality word embeddings. Our study explores these types of knowledge to define new basis for word representation, provide additional input information, and serve as auxiliary supervision in deep learning, respectively. Experiments on an analogical reason- ing task, a word similarity task, and a word completion task have all demonstrated that knowledge-powered deep learning can enhance the effectiveness of word em- bedding. 1 Introduction With rapid development of deep learning techniques in recent years, it has drawn in- creasing attention to train complex and deep models on large amounts of data, in order to solve a wide range of text mining and natural language processing (NLP) tasks [4, 1, 8, 13, 19, 20].
    [Show full text]
  • Recurrent Neural Network Based Language Model
    INTERSPEECH 2010 Recurrent neural network based language model Toma´sˇ Mikolov1;2, Martin Karafiat´ 1, Luka´sˇ Burget1, Jan “Honza” Cernockˇ y´1, Sanjeev Khudanpur2 1Speech@FIT, Brno University of Technology, Czech Republic 2 Department of Electrical and Computer Engineering, Johns Hopkins University, USA fimikolov,karafiat,burget,[email protected], [email protected] Abstract INPUT(t) OUTPUT(t) A new recurrent neural network based language model (RNN CONTEXT(t) LM) with applications to speech recognition is presented. Re- sults indicate that it is possible to obtain around 50% reduction of perplexity by using mixture of several RNN LMs, compared to a state of the art backoff language model. Speech recognition experiments show around 18% reduction of word error rate on the Wall Street Journal task when comparing models trained on the same amount of data, and around 5% on the much harder NIST RT05 task, even when the backoff model is trained on much more data than the RNN LM. We provide ample empiri- cal evidence to suggest that connectionist language models are superior to standard n-gram techniques, except their high com- putational (training) complexity. Index Terms: language modeling, recurrent neural networks, speech recognition CONTEXT(t-1) 1. Introduction Figure 1: Simple recurrent neural network. Sequential data prediction is considered by many as a key prob- lem in machine learning and artificial intelligence (see for ex- ample [1]). The goal of statistical language modeling is to plex and often work well only for systems based on very limited predict the next word in textual data given context; thus we amounts of training data.
    [Show full text]
  • Unified Language Model Pre-Training for Natural
    Unified Language Model Pre-training for Natural Language Understanding and Generation Li Dong∗ Nan Yang∗ Wenhui Wang∗ Furu Wei∗† Xiaodong Liu Yu Wang Jianfeng Gao Ming Zhou Hsiao-Wuen Hon Microsoft Research {lidong1,nanya,wenwan,fuwei}@microsoft.com {xiaodl,yuwan,jfgao,mingzhou,hon}@microsoft.com Abstract This paper presents a new UNIfied pre-trained Language Model (UNILM) that can be fine-tuned for both natural language understanding and generation tasks. The model is pre-trained using three types of language modeling tasks: unidirec- tional, bidirectional, and sequence-to-sequence prediction. The unified modeling is achieved by employing a shared Transformer network and utilizing specific self-attention masks to control what context the prediction conditions on. UNILM compares favorably with BERT on the GLUE benchmark, and the SQuAD 2.0 and CoQA question answering tasks. Moreover, UNILM achieves new state-of- the-art results on five natural language generation datasets, including improving the CNN/DailyMail abstractive summarization ROUGE-L to 40.51 (2.04 absolute improvement), the Gigaword abstractive summarization ROUGE-L to 35.75 (0.86 absolute improvement), the CoQA generative question answering F1 score to 82.5 (37.1 absolute improvement), the SQuAD question generation BLEU-4 to 22.12 (3.75 absolute improvement), and the DSTC7 document-grounded dialog response generation NIST-4 to 2.67 (human performance is 2.65). The code and pre-trained models are available at https://github.com/microsoft/unilm. 1 Introduction Language model (LM) pre-training has substantially advanced the state of the art across a variety of natural language processing tasks [8, 29, 19, 31, 9, 1].
    [Show full text]
  • Hello, It's GPT-2
    Hello, It’s GPT-2 - How Can I Help You? Towards the Use of Pretrained Language Models for Task-Oriented Dialogue Systems Paweł Budzianowski1;2;3 and Ivan Vulic´2;3 1Engineering Department, Cambridge University, UK 2Language Technology Lab, Cambridge University, UK 3PolyAI Limited, London, UK [email protected], [email protected] Abstract (Young et al., 2013). On the other hand, open- domain conversational bots (Li et al., 2017; Serban Data scarcity is a long-standing and crucial et al., 2017) can leverage large amounts of freely challenge that hinders quick development of available unannotated data (Ritter et al., 2010; task-oriented dialogue systems across multiple domains: task-oriented dialogue models are Henderson et al., 2019a). Large corpora allow expected to learn grammar, syntax, dialogue for training end-to-end neural models, which typ- reasoning, decision making, and language gen- ically rely on sequence-to-sequence architectures eration from absurdly small amounts of task- (Sutskever et al., 2014). Although highly data- specific data. In this paper, we demonstrate driven, such systems are prone to producing unre- that recent progress in language modeling pre- liable and meaningless responses, which impedes training and transfer learning shows promise their deployment in the actual conversational ap- to overcome this problem. We propose a task- oriented dialogue model that operates solely plications (Li et al., 2017). on text input: it effectively bypasses ex- Due to the unresolved issues with the end-to- plicit policy and language generation modules. end architectures, the focus has been extended to Building on top of the TransferTransfo frame- retrieval-based models.
    [Show full text]
  • N-Gram Language Modeling Tutorial Dustin Hillard and Sarah Petersen Lecture Notes Courtesy of Prof
    N-gram Language Modeling Tutorial Dustin Hillard and Sarah Petersen Lecture notes courtesy of Prof. Mari Ostendorf Outline: • Statistical Language Model (LM) Basics • n-gram models • Class LMs • Cache LMs • Mixtures • Empirical observations (Goodman CSL 2001) • Factored LMs Part I: Statistical Language Model (LM) Basics • What is a statistical LM and why are they interesting? • Evaluating LMs • History equivalence classes What is a statistical language model? A stochastic process model for word sequences. A mechanism for computing the probability of: p(w1, . , wT ) 1 Why are LMs interesting? • Important component of a speech recognition system – Helps discriminate between similar sounding words – Helps reduce search costs • In statistical machine translation, a language model characterizes the target language, captures fluency • For selecting alternatives in summarization, generation • Text classification (style, reading level, language, topic, . ) • Language models can be used for more than just words – letter sequences (language identification) – speech act sequence modeling – case and punctuation restoration 2 Evaluating LMs Evaluating LMs in the context of an application can be expensive, so LMs are usually evaluated on the own in terms of perplexity:: ˜ 1 PP = 2Hr where H˜ = − log p(w , . , w ) r T 2 1 T where {w1, . , wT } is held out test data that provides the empirical distribution q(·) in the cross-entropy formula H˜ = − X q(x) log p(x) x and p(·) is the LM estimated on a training set. Interpretations: • Entropy rate: lower entropy means that it is easier to predict the next symbol and hence easier to rule out alternatives when combined with other models small H˜r → small PP • Average branching factor: When a distribution is uniform for a vocabulary of size V , then entropy is log2 V , and perplexity is V .
    [Show full text]
  • A General Language Model for Information Retrieval Fei Song W
    A General Language Model for Information Retrieval Fei Song W. Bruce Croft Dept. of Computing and Info. Science Dept. of Computer Science University of Guelph University of Massachusetts Guelph, Ontario, Canada N1G 2W1 Amherst, Massachusetts 01003 [email protected] [email protected] ABSTRACT However, estimating the probabilities of word sequences can be expensive, since sentences can be arbitrarily long and the size of Statistical language modeling has been successfully used for a corpus needs to be very large. In practice, the statistical speech recognition, part-of-speech tagging, and syntactic parsing. language model is often approximated by N-gram models. Recently, it has also been applied to information retrieval. Unigram: P(w ) = P(w )P(w ) ¡ P(w ) According to this new paradigm, each document is viewed as a 1,n 1 2 n language sample, and a query as a generation process. The Bigram: P(w ) = P(w )P(w | w ) ¡ P(w | w ) retrieved documents are ranked based on the probabilities of 1,n 1 2 1 n n−1 producing a query from the corresponding language models of Trigram: P(w ) = P(w )P(w | w )P(w | w ) ¡ P(w | w ) these documents. In this paper, we will present a new language 1,n 1 2 1 3 1,2 n n−2,n−1 model for information retrieval, which is based on a range of data The unigram model makes a strong assumption that each word smoothing techniques, including the Good-Turing estimate, occurs independently, and consequently, the probability of a curve-fitting functions, and model combinations.
    [Show full text]
  • A Unified Multilingual Handwriting Recognition System Using
    1 A Unified Multilingual Handwriting Recognition System using multigrams sub-lexical units Wassim Swaileha,∗, Yann Soullarda, Thierry Paqueta aLITIS Laboratory - EA 4108 Normandie University - University of Rouen Rouen, France ABSTRACT We address the design of a unified multilingual system for handwriting recognition. Most of multi- lingual systems rests on specialized models that are trained on a single language and one of them is selected at test time. While some recognition systems are based on a unified optical model, dealing with a unified language model remains a major issue, as traditional language models are generally trained on corpora composed of large word lexicons per language. Here, we bring a solution by con- sidering language models based on sub-lexical units, called multigrams. Dealing with multigrams strongly reduces the lexicon size and thus decreases the language model complexity. This makes pos- sible the design of an end-to-end unified multilingual recognition system where both a single optical model and a single language model are trained on all the languages. We discuss the impact of the language unification on each model and show that our system reaches state-of-the-art methods perfor- mance with a strong reduction of the complexity. c 2018 Elsevier Ltd. All rights reserved. 1. Introduction Offline handwriting recognition is a challenging task due to the high variability of data, as writing styles, the quality of the input image and the lack of contextual information. The recognition is commonly done in two steps. First, an optical character recognition (OCR) model is optimized for recognizing sequences of characters from input images of handwritten texts (Plotz¨ and Fink (2009); Singh (2013)).
    [Show full text]
  • Word Embeddings: History & Behind the Scene
    Word Embeddings: History & Behind the Scene Alfan F. Wicaksono Information Retrieval Lab. Faculty of Computer Science Universitas Indonesia References • Bengio, Y., Ducharme, R., Vincent, P., & Janvin, C. (2003). A Neural Probabilistic Language Model. The Journal of Machine Learning Research, 3, 1137–1155. • Mikolov, T., Corrado, G., Chen, K., & Dean, J. (2013). Efficient Estimation of Word Representations in Vector Space. Proceedings of the International Conference on Learning Representations (ICLR 2013), 1–12 • Mikolov, T., Chen, K., Corrado, G., & Dean, J. (2013). Distributed Representations of Words and Phrases and their Compositionality. NIPS, 1–9. • Morin, F., & Bengio, Y. (2005). Hierarchical Probabilistic Neural Network Language Model. Aistats, 5. References Good weblogs for high-level understanding: • http://sebastianruder.com/word-embeddings-1/ • Sebastian Ruder. On word embeddings - Part 2: Approximating the Softmax. http://sebastianruder.com/word- embeddings-softmax • https://www.tensorflow.org/tutorials/word2vec • https://www.gavagai.se/blog/2015/09/30/a-brief-history-of- word-embeddings/ Some slides were also borrowed from From Dan Jurafsky’s course slide: Word Meaning and Similarity. Stanford University. Terminology • The term “Word Embedding” came from deep learning community • For computational linguistic community, they prefer “Distributional Semantic Model” • Other terms: – Distributed Representation – Semantic Vector Space – Word Space https://www.gavagai.se/blog/2015/09/30/a-brief-history-of-word-embeddings/ Before we learn Word Embeddings... Semantic Similarity • Word Similarity – Near-synonyms – “boat” and “ship”, “car” and “bicycle” • Word Relatedness – Can be related any way – Similar: “boat” and “ship” – Topical Similarity: • “boat” and “water” • “car” and “gasoline” Why Word Similarity? • Document Classification • Document Clustering • Language Modeling • Information Retrieval • ..
    [Show full text]
  • Language Modelling for Speech Recognition
    Language Modelling for Speech Recognition • Introduction • n-gram language models • Probability estimation • Evaluation • Beyond n-grams 6.345 Automatic Speech Recognition Language Modelling 1 Language Modelling for Speech Recognition • Speech recognizers seek the word sequence Wˆ which is most likely to be produced from acoustic evidence A P(Wˆ |A)=m ax P(W |A) ∝ max P(A|W )P(W ) W W • Speech recognition involves acoustic processing, acoustic modelling, language modelling, and search • Language models (LMs) assign a probability estimate P(W ) to word sequences W = {w1,...,wn} subject to � P(W )=1 W • Language models help guide and constrain the search among alternative word hypotheses during recognition 6.345 Automatic Speech Recognition Language Modelling 2 Language Model Requirements � Constraint Coverage � NLP Understanding � 6.345 Automatic Speech Recognition Language Modelling 3 Finite-State Networks (FSN) show me all the flights give restaurants display • Language space defined by a word network or graph • Describable by a regular phrase structure grammar A =⇒ aB | a • Finite coverage can present difficulties for ASR • Graph arcs or rules can be augmented with probabilities 6.345 Automatic Speech Recognition Language Modelling 4 Context-Free Grammars (CFGs) VP NP V D N display the flights • Language space defined by context-free rewrite rules e.g., A =⇒ BC | a • More powerful representation than FSNs • Stochastic CFG rules have associated probabilities which can be learned automatically from a corpus • Finite coverage can present difficulties
    [Show full text]
  • A Pretrained Language Model for Scientific Text
    SCIBERT: A Pretrained Language Model for Scientific Text Iz Beltagy Kyle Lo Arman Cohan Allen Institute for Artificial Intelligence, Seattle, WA, USA fbeltagy,kylel,[email protected] Abstract neural architectures. Leveraging the success of un- supervised pretraining has become especially im- Obtaining large-scale annotated data for NLP portant especially when task-specific annotations tasks in the scientific domain is challenging are difficult to obtain, like in scientific NLP. Yet and expensive. We release SCIBERT, a pre- while both BERT and ELMo have released pre- trained language model based on BERT (De- vlin et al., 2019) to address the lack of high- trained models, they are still trained on general do- quality, large-scale labeled scientific data. main corpora such as news articles and Wikipedia. SCIBERT leverages unsupervised pretraining In this work, we make the following contribu- on a large multi-domain corpus of scien- tions: tific publications to improve performance on (i) We release SCIBERT, a new resource demon- downstream scientific NLP tasks. We eval- strated to improve performance on a range of NLP uate on a suite of tasks including sequence tasks in the scientific domain. SCIBERT is a pre- tagging, sentence classification and depen- dency parsing, with datasets from a variety trained language model based on BERT but trained of scientific domains. We demonstrate sta- on a large corpus of scientific text. tistically significant improvements over BERT (ii) We perform extensive experimentation to and achieve new state-of-the-art results on sev- investigate the performance of finetuning ver- eral of these tasks. The code and pretrained sus task-specific architectures atop frozen embed- models are available at https://github.
    [Show full text]
  • Improving Text Recognition Using Optical and Language Model Writer Adaptation
    Improving text recognition using optical and language model writer adaptation Yann Soullard, Wassim Swaileh, Pierrick Tranouez, Thierry Paquet Clement´ Chatelain LITIS lab LITIS lab Normandie University, Universite de Rouen Normandie INSA Rouen Normandie 76800 Saint Etienne du Rouvray, France 76800 Saint Etienne du Rouvray, France Email: ffi[email protected] Email: [email protected] Abstract—State-of-the-art methods for handwriting text shown to be very efficient on various recognition tasks recognition are based on deep learning approaches and lan- such as handwriting recognition, signature identification or guage modeling that require large data sets during training. medical imaging [4], [5]. In addition, the language itself, but In practice, there are some applications where the system processes mono-writer documents, and would thus benefit from also lexicons as well as word frequencies may not match being trained on examples from that writer. However, this is the statistical distributions learned during language model not common to have numerous examples coming from just one training. Statistical language model adaptation algorithms writer. In this paper, we propose an approach to adapt both the [6] have also been designed in this respect. optical model and the language model to a particular writer, This paper presents an approach to adapt a generic text from a generic system trained on large data sets with a variety of examples. We show the benefits of the optical and language recognition system to a particular writer. In this respect, both model writer adaptation. Our approach reaches competitive the optical model and the language model of the generic sys- results on the READ 2018 data set, which is dedicated to model tem are considered during the adaptation phase.
    [Show full text]
  • Exploring the Limits of Language Modeling
    Exploring the Limits of Language Modeling Rafal Jozefowicz [email protected] Oriol Vinyals [email protected] Mike Schuster [email protected] Noam Shazeer [email protected] Yonghui Wu [email protected] Google Brain Abstract language models improves the underlying metrics of the In this work we explore recent advances in Re- downstream task (such as word error rate for speech recog- current Neural Networks for large scale Lan- nition, or BLEU score for translation), which makes the guage Modeling, a task central to language un- task of training better LMs valuable by itself. derstanding. We extend current models to deal Further, when trained on vast amounts of data, language with two key challenges present in this task: cor- models compactly extract knowledge encoded in the train- pora and vocabulary sizes, and complex, long ing data. For example, when trained on movie subti- term structure of language. We perform an ex- tles (Serban et al., 2015; Vinyals & Le, 2015), these lan- haustive study on techniques such as character guage models are able to generate basic answers to ques- Convolutional Neural Networks or Long-Short tions about object colors, facts about people, etc. Lastly, Term Memory, on the One Billion Word Bench- recently proposed sequence-to-sequence models employ mark. Our best single model significantly im- conditional language models (Mikolov & Zweig, 2012) as proves state-of-the-art perplexity from 51.3 down their key component to solve diverse tasks like machine to 30.0 (whilst reducing the number of param- translation (Sutskever et al., 2014; Cho et al., 2014; Kalch- eters by a factor of 20), while an ensemble of brenner et al., 2014) or video generation (Srivastava et al., models sets a new record by improving perplex- 2015a).
    [Show full text]