DOCSLIB.ORG

Worse WER, but Better BLEU? Leveraging Word Embedding As Intermediate in Multitask End-To-End Speech Translation

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
Worse WER, but Better BLEU? Leveraging Word Embedding As Intermediate in Multitask End-To-End Speech Translation Worse WER, but Better BLEU? Leveraging Word Embedding as Intermediate in Multitask End-to-End Speech Translation Shun-Po Chuang1, Tzu-Wei Sung2, Alexander H. Liu1, and Hung-yi Lee1 1National Taiwan University, Taiwan 2University of California San Diego, USA ff04942141, b03902042, r07922013, [email protected] Abstract Although applying the text of source language as the intermediate information in multitask end-to- Speech translation (ST) aims to learn transfor- end ST empirically yielded improvement, we argue mations from speech in the source language to the text in the target language. Previous works whether this is the optimal solution. Even though show that multitask learning improves the ST the recognition part does not correctly transcribe performance, in which the recognition decoder the input speech into text, the final translation result generates the text of the source language, and would be correct if the output of the recognition the translation decoder obtains the final trans- part preserves sufficient semantic information for lations based on the output of the recognition translation. Therefore, we explore to leverage word decoder. Because whether the output of the embedding as the intermediate level instead of text. recognition decoder has the correct semantics is more critical than its accuracy, we propose In this paper, we apply pre-trained word embed- to improve the multitask ST model by utilizing ding as the intermediate level in the multitask ST word embedding as the intermediate. model. We propose to constrain the hidden states of the decoder of the recognition part to be close 1 Introduction to the pre-trained word embedding. Prior works on Speech translation (ST) increasingly receives atten- word embedding regression show improved results tion from the machine translation (MT) commu- on MT (Jauregi Unanue et al., 2019; Kumar and nity recently. To learn the transformation between Tsvetkov, 2018). Experimental results show that speech in the source language and the text in the the proposed approach obtains improvement to the target language, conventional models pipeline au- ST model. Further analysis also shows that con- tomatic speech recognition (ASR) and text-to-text strained hidden states are approximately isospectral MT model (Berard´ et al., 2016). However, such to word embedding space, indicating that the de- pipeline systems suffer from error propagation. coder achieves speech-to-semantic mappings. Previous works show that deep end-to-end mod- 2 Multitask End-to-End ST model els can outperform conventional pipeline systems with sufficient training data (Weiss et al., 2017; In- Our method is based on the multitask learning aguma et al., 2019; Sperber et al., 2019). Neverthe- for ST (Anastasopoulos and Chiang, 2018), in- less, well-annotated bilingual data is expensive and cluding speech recognition in the source language hard to collect (Bansal et al., 2018a,b; Duong et al., and translation in the target language, as shown in 2016). Multitask learning plays an essential role in Fig.1(a). The input audio feature sequence is first leveraging a large amount of monolingual data to encoded into the encoder hidden state sequence improve representation in ST. Multitask ST models h = h1; h2; : : : ; hT with length T by the pyramid have two jointly learned decoding parts, namely the encoder (Chan et al., 2015). To present speech recognition and translation part. The recognition recognition in the source language, the attention part firstly decodes the speech of source language mechanism and a decoder is employed to pro- into the text of source language, and then based on duce source decoder sequence s^ =s ^1; s^2;:::; s^M , the output of the recognition part, the translation where M is the number of decoding steps in the part generates the text in the target language. Vari- source language. For each decoding step m, the ant multitask models have been explored (Anas- probability P (^ym) of predicting the token y^m in tasopoulos and Chiang, 2018), which shows the the source language vocabulary can be computed improvement in low-resource scenario. based on the corresponding decoder state s^m. 5998 Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5998–6003 July 5 - 10, 2020. c 2020 Association for Computational Linguistics (a) Multitask End-to-End Model D set and e^v 2 R is the embedding vector with di- Recognition ( Intermedia Level ) Translation mension D for any word v 2 V , in the recognition [… , �%& , … ] [… , �),… ] Cross task. We choose the source language decoder state Entropy [… , �(� ),… ] [… , �(�%&),… ] ) (embedding) s^ to reinforce since it is later used in Source Language ������� ������� Speech Features Linear Linear the translation task. To be more specific, we argue [ �+̂ , �,̂ , … , �0̂ ] that the embedding generated by the source lan- Source Language Attention guage decoder should be more semantically correct Decoder Target Speech Concatenation Language in order to benefit the translation task. Given the Encoder Attention Decoder pre-trained source language word embedding E^, [ ℎ+ , ℎ, , … , ℎ-] Attention we proposed to constrain the source decoder state (b) Cosine Distance (CD) (c) Cosine Softmax (CS) [… , �%&, … ] Cosine Distance Cross s^ at step m to be close to its corresponding word Linear [… , �& , … ] Entropy m [… , �̂ , … ] [… , � , … ] [… , �9: (�%&), … ] 5%6 & [… , �̂ , … ] embedding e^y^m with the two approaches detailed & ������� Linear Cosine Similarity Source Source in the following sections. [… , �̂ , … ] Language & Language … �+̂ �̂, �̂|8| �B Decoder Decoder 3.1 Directly Learn Word Embedding Figure 1: (a) Multitask ST model. Dotted arrows indi- Since semantic-related words would be close in cate steps in the recognition part. Solid arrows indicate terms of cosine distance (Mikolov et al., 2018), a steps in the translation part. (b) Directly learn word simple idea is to minimize the cosine distance (CD) embedding via cosine distance. (c) Learn word embed- between the source language decoder hidden state ding via cosine softmax function. Both (b)(c) are the s^m and the corresponding word embedding e^y^m for recognition part in (a). every decode step m, X LCD = 1 − cos(fθ(^sm); e^y^m ) To perform speech translation in the target lan- m guage, both the source language decoder state se- (2) X fθ(^sm) · e^y^m quence s^ and the encoder state sequence h will = 1 − ; kfθ(^sm)kke^y^m k be attended and treated as the target language de- m coder’s input. The hidden state of target language where fθ(·) is a learnable linear projection to match decoder can then be used to derived the probability the dimensionality of word embedding and decoder state. With this design, the network architecture of P (yq) of predicting token yq in the target language vocabulary for every decoding step q. the target language decoder would not be limited Given the ground truth sequence in the source by the dimension of word embedding. Fig.1(b) il- lustrates this approach. By replacing L in Eq. (1) language y^ =y ^1; y^2;:::; y^M and the target lan- src with L , semantic learning from word embedding guage y = y1; y2; : : : ; yQ with length Q, multitask CD ST can be trained with maximizing log likelihood for source language recognition can be achieved. in both domains. Formally, the objective function 3.2 Learn Word Embedding via Probability of multitask ST can be written as: Ideally, using word embedding as the learning tar- α β get via minimizing CD can effectively train the LST = Lsrc + Ltgt M Q decoder to model the semantic relation existing in α X β X = − log P (^y ) + − log P (y ); the embedding space. However, such an approach M m Q q m q suffers from the hubness problem (Faruqui et al., (1) 2016) of word embedding in practice (as we later discuss in Sec. 4.5). where α and β are the trade-off factors to balance To address this problem, we introduce cosine between the two tasks. softmax (CS) function (Liu et al., 2017a,b) to learn speech-to-semantic embedding mappings. Given 3 Proposed Methods the decoder hidden state s^m and the word embed- ^ We propose two ways to help the multitask end- ding E, the probability of the target word y^m is to-end ST model capture the semantic relation defined as between word tokens by leveraging the source exp(cos(fθ(^sm); e^y^m )=τ) PCS(^ym) = P ; language word embedding as intermediate level. e^ 2E^ exp(cos(fθ(^sm); e^v)=τ) ^ v E = fe^1; e^2; :::e^jV jg, where V is the vocabulary (3) 5999 where cos(·) and fθ(·) are from Eq. (2), and τ is the (a) 160 hours (b) 40 hours temperature of softmax function. Note that since dev test dev test the temperature τ re-scales cosine similarity, the SE 34.50 34.51 17.41 15.44 hubness problem can be mitigated by selecting a ME 35.35 35.49 23.30 20.40 proper value for τ. Fig.1(c) illustrates the approach. CD 33.06 33.65 23.53 20.87 With the probability derived from cosine softmax in CS 35.84 36.32 23.54 21.72 Eq. (3), the objective function for source language decoder can be written as Table 1: BLEU scores trained on different size of data. X LCS = − log PCS(^ym): (4) m we could see that ME outperforms SE in all condi- tions. By replacing Lsrc in Eq. (1) with LCS, the decoder High-resource: Column (a) in Table1 showed hidden state sequence s^ is forced to contain seman- the results trained on 160 hours of data. CD and tic information provided by the word embedding. CS represent the proposed methods mentioned in Sec. 3.1 and 3.2 respectively. We got mixed results 4 Experiments on further applying pre-trained word embedding on ME.
Load more

Recommended publications
  • Arxiv:2002.06235V1
    Semantic Relatedness and Taxonomic Word Embeddings Magdalena Kacmajor John D. Kelleher Innovation Exchange ADAPT Research Centre IBM Ireland Technological University Dublin [email protected] [email protected] Filip Klubickaˇ Alfredo Maldonado ADAPT Research Centre Underwriters Laboratories Technological University Dublin [email protected] [email protected] Abstract This paper1 connects a series of papers dealing with taxonomic word embeddings. It begins by noting that there are different types of semantic relatedness and that different lexical representa- tions encode different forms of relatedness. A particularly important distinction within semantic relatedness is that of thematic versus taxonomic relatedness. Next, we present a number of ex- periments that analyse taxonomic embeddings that have been trained on a synthetic corpus that has been generated via a random walk over a taxonomy. These experiments demonstrate how the properties of the synthetic corpus, such as the percentage of rare words, are affected by the shape of the knowledge graph the corpus is generated from. Finally, we explore the interactions between the relative sizes of natural and synthetic corpora on the performance of embeddings when taxonomic and thematic embeddings are combined. 1 Introduction Deep learning has revolutionised natural language processing over the last decade. A key enabler of deep learning for natural language processing has been the development of word embeddings. One reason for this is that deep learning intrinsically involves the use of neural network models and these models only work with numeric inputs. Consequently, applying deep learning to natural language processing first requires developing a numeric representation for words. Word embeddings provide a way of creating numeric representations that have proven to have a number of advantages over traditional numeric rep- resentations of language.
    [Show full text]
  • Combining Word Embeddings with Semantic Resources
    AutoExtend: Combining Word Embeddings with Semantic Resources Sascha Rothe∗ LMU Munich Hinrich Schutze¨ ∗ LMU Munich We present AutoExtend, a system that combines word embeddings with semantic resources by learning embeddings for non-word objects like synsets and entities and learning word embeddings that incorporate the semantic information from the resource. The method is based on encoding and decoding the word embeddings and is flexible in that it can take any word embeddings as input and does not need an additional training corpus. The obtained embeddings live in the same vector space as the input word embeddings. A sparse tensor formalization guar- antees efficiency and parallelizability. We use WordNet, GermaNet, and Freebase as semantic resources. AutoExtend achieves state-of-the-art performance on Word-in-Context Similarity and Word Sense Disambiguation tasks. 1. Introduction Unsupervised methods for learning word embeddings are widely used in natural lan- guage processing (NLP). The only data these methods need as input are very large corpora. However, in addition to corpora, there are many other resources that are undoubtedly useful in NLP, including lexical resources like WordNet and Wiktionary and knowledge bases like Wikipedia and Freebase. We will simply refer to these as resources. In this article, we present AutoExtend, a method for enriching these valuable resources with embeddings for non-word objects they describe; for example, Auto- Extend enriches WordNet with embeddings for synsets. The word embeddings and the new non-word embeddings live in the same vector space. Many NLP applications benefit if non-word objects described by resources—such as synsets in WordNet—are also available as embeddings.
    [Show full text]
  • Word-Like Character N-Gram Embedding
    Word-like character n-gram embedding Geewook Kim and Kazuki Fukui and Hidetoshi Shimodaira Department of Systems Science, Graduate School of Informatics, Kyoto University Mathematical Statistics Team, RIKEN Center for Advanced Intelligence Project fgeewook, [email protected], [email protected] Abstract Table 1: Top-10 2-grams in Sina Weibo and 4-grams in We propose a new word embedding method Japanese Twitter (Experiment 1). Words are indicated called word-like character n-gram embed- by boldface and space characters are marked by . ding, which learns distributed representations FNE WNE (Proposed) Chinese Japanese Chinese Japanese of words by embedding word-like character n- 1 ][ wwww 自己 フォロー grams. Our method is an extension of recently 2 。␣ ！！！！ 。␣ ありがと proposed segmentation-free word embedding, 3 ！␣ ありがと ][ wwww 4 .. りがとう 一个 ！！！！ which directly embeds frequent character n- 5 ]␣ ございま 微博 めっちゃ grams from a raw corpus. However, its n-gram 6 。。 うござい 什么 んだけど vocabulary tends to contain too many non- 7 ，我 とうござ 可以 うござい 8 ！！ ざいます 没有 line word n-grams. We solved this problem by in- 9 ␣我 がとうご 吗？ 2018 troducing an idea of expected word frequency. 10 了， ください 哈哈 じゃない Compared to the previously proposed meth- ods, our method can embed more words, along tion tools are used to determine word boundaries with the words that are not included in a given in the raw corpus. However, these segmenters re- basic word dictionary. Since our method does quire rich dictionaries for accurate segmentation, not rely on word segmentation with rich word which are expensive to prepare and not always dictionaries, it is especially effective when the available.
    [Show full text]
  • Learned in Speech Recognition: Contextual Acoustic Word Embeddings
    LEARNED IN SPEECH RECOGNITION: CONTEXTUAL ACOUSTIC WORD EMBEDDINGS Shruti Palaskar∗, Vikas Raunak∗ and Florian Metze Carnegie Mellon University, Pittsburgh, PA, U.S.A. fspalaska j vraunak j fmetze [email protected] ABSTRACT model [10, 11, 12] trained for direct Acoustic-to-Word (A2W) speech recognition [13]. Using this model, we jointly learn to End-to-end acoustic-to-word speech recognition models have re- automatically segment and classify input speech into individual cently gained popularity because they are easy to train, scale well to words, hence getting rid of the problem of chunking or requiring large amounts of training data, and do not require a lexicon. In addi- pre-defined word boundaries. As our A2W model is trained at the tion, word models may also be easier to integrate with downstream utterance level, we show that we can not only learn acoustic word tasks such as spoken language understanding, because inference embeddings, but also learn them in the proper context of their con- (search) is much simplified compared to phoneme, character or any taining sentence. We also evaluate our contextual acoustic word other sort of sub-word units. In this paper, we describe methods embeddings on a spoken language understanding task, demonstrat- to construct contextual acoustic word embeddings directly from a ing that they can be useful in non-transcription downstream tasks. supervised sequence-to-sequence acoustic-to-word speech recog- Our main contributions in this paper are the following: nition model using the learned attention distribution. On a suite 1. We demonstrate the usability of attention not only for aligning of 16 standard sentence evaluation tasks, our embeddings show words to acoustic frames without any forced alignment but also for competitive performance against a word2vec model trained on the constructing Contextual Acoustic Word Embeddings (CAWE).
    [Show full text]
  • 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]
  • Local Homology of Word Embeddings
    Local Homology of Word Embeddings Tadas Temcinasˇ 1 Abstract Intuitively, stratification is a decomposition of a topologi- Topological data analysis (TDA) has been widely cal space into manifold-like pieces. When thinking about used to make progress on a number of problems. stratification learning and word embeddings, it seems intu- However, it seems that TDA application in natural itive that vectors of words corresponding to the same broad language processing (NLP) is at its infancy. In topic would constitute a structure, which we might hope this paper we try to bridge the gap by arguing why to be a manifold. Hence, for example, by looking at the TDA tools are a natural choice when it comes to intersections between those manifolds or singularities on analysing word embedding data. We describe the manifolds (both of which can be recovered using local a parallelisable unsupervised learning algorithm homology-based algorithms (Nanda, 2017)) one might hope based on local homology of datapoints and show to find vectors of homonyms like ‘bank’ (which can mean some experimental results on word embedding either a river bank, or a financial institution). This, in turn, data. We see that local homology of datapoints in has potential to help solve the word sense disambiguation word embedding data contains some information (WSD) problem in NLP, which is pinning down a particular that can potentially be used to solve the word meaning of a word used in a sentence when the word has sense disambiguation problem. multiple meanings. In this work we present a clustering algorithm based on lo- cal homology, which is more relaxed1 that the stratification- 1.
    [Show full text]
  • Word Embedding, Sense Embedding and Their Application to Word Sense Induction
    Word Embeddings, Sense Embeddings and their Application to Word Sense Induction Linfeng Song The University of Rochester Computer Science Department Rochester, NY 14627 Area Paper April 2016 Abstract This paper investigates the cutting-edge techniques for word embedding, sense embedding, and our evaluation results on large-scale datasets. Word embedding refers to a kind of methods that learn a distributed dense vector for each word in a vocabulary. Traditional word embedding methods first obtain the co-occurrence matrix then perform dimension reduction with PCA. Recent methods use neural language models that directly learn word vectors by predicting the context words of the target word. Moving one step forward, sense embedding learns a distributed vector for each sense of a word. They either define a sense as a cluster of contexts where the target word appears or define a sense based on a sense inventory. To evaluate the performance of the state-of-the-art sense embedding methods, I first compare them on the dominant word similarity datasets, then compare them on my experimental settings. In addition, I show that sense embedding is applicable to the task of word sense induction (WSI). Actually we are the first to show that sense embedding methods are competitive on WSI by building sense-embedding-based systems that demonstrate highly competitive performances on the SemEval 2010 WSI shared task. Finally, I propose several possible future research directions on word embedding and sense embedding. The University of Rochester Computer Science Department supported this work. Contents 1 Introduction 3 2 Word Embedding 5 2.1 Skip-gramModel................................
    [Show full text]
  • Learning Word Meta-Embeddings by Autoencoding
    Learning Word Meta-Embeddings by Autoencoding Cong Bao Danushka Bollegala Department of Computer Science Department of Computer Science University of Liverpool University of Liverpool [email protected] [email protected] Abstract Distributed word embeddings have shown superior performances in numerous Natural Language Processing (NLP) tasks. However, their performances vary significantly across different tasks, implying that the word embeddings learnt by those methods capture complementary aspects of lexical semantics. Therefore, we believe that it is important to combine the existing word em- beddings to produce more accurate and complete meta-embeddings of words. We model the meta-embedding learning problem as an autoencoding problem, where we would like to learn a meta-embedding space that can accurately reconstruct all source embeddings simultaneously. Thereby, the meta-embedding space is enforced to capture complementary information in differ- ent source embeddings via a coherent common embedding space. We propose three flavours of autoencoded meta-embeddings motivated by different requirements that must be satisfied by a meta-embedding. Our experimental results on a series of benchmark evaluations show that the proposed autoencoded meta-embeddings outperform the existing state-of-the-art meta- embeddings in multiple tasks. 1 Introduction Representing the meanings of words is a fundamental task in Natural Language Processing (NLP). A popular approach to represent the meaning of a word is to embed it in some fixed-dimensional
    [Show full text]
  • A Comparison of Word Embeddings and N-Gram Models for Dbpedia Type and Invalid Entity Detection †
    information Article A Comparison of Word Embeddings and N-gram Models for DBpedia Type and Invalid Entity Detection † Hanqing Zhou *, Amal Zouaq and Diana Inkpen School of Electrical Engineering and Computer Science, University of Ottawa, Ottawa ON K1N 6N5, Canada; [email protected] (A.Z.); [email protected] (D.I.) * Correspondence: [email protected]; Tel.: +1-613-562-5800 † This paper is an extended version of our conference paper: Hanqing Zhou, Amal Zouaq, and Diana Inkpen. DBpedia Entity Type Detection using Entity Embeddings and N-Gram Models. In Proceedings of the International Conference on Knowledge Engineering and Semantic Web (KESW 2017), Szczecin, Poland, 8–10 November 2017, pp. 309–322. Received: 6 November 2018; Accepted: 20 December 2018; Published: 25 December 2018 Abstract: This article presents and evaluates a method for the detection of DBpedia types and entities that can be used for knowledge base completion and maintenance. This method compares entity embeddings with traditional N-gram models coupled with clustering and classification. We tackle two challenges: (a) the detection of entity types, which can be used to detect invalid DBpedia types and assign DBpedia types for type-less entities; and (b) the detection of invalid entities in the resource description of a DBpedia entity. Our results show that entity embeddings outperform n-gram models for type and entity detection and can contribute to the improvement of DBpedia’s quality, maintenance, and evolution. Keywords: semantic web; DBpedia; entity embedding; n-grams; type identification; entity identification; data mining; machine learning 1. Introduction The Semantic Web is defined by Berners-Lee et al.
    [Show full text]
  • Arxiv:2007.00183V2 [Eess.AS] 24 Nov 2020
    WHOLE-WORD SEGMENTAL SPEECH RECOGNITION WITH ACOUSTIC WORD EMBEDDINGS Bowen Shi, Shane Settle, Karen Livescu TTI-Chicago, USA fbshi,settle.shane,[email protected] ABSTRACT Segmental models are sequence prediction models in which scores of hypotheses are based on entire variable-length seg- ments of frames. We consider segmental models for whole- word (“acoustic-to-word”) speech recognition, with the feature vectors defined using vector embeddings of segments. Such models are computationally challenging as the number of paths is proportional to the vocabulary size, which can be orders of magnitude larger than when using subword units like phones. We describe an efficient approach for end-to-end whole-word segmental models, with forward-backward and Viterbi de- coding performed on a GPU and a simple segment scoring function that reduces space complexity. In addition, we inves- tigate the use of pre-training via jointly trained acoustic word embeddings (AWEs) and acoustically grounded word embed- dings (AGWEs) of written word labels. We find that word error rate can be reduced by a large margin by pre-training the acoustic segment representation with AWEs, and additional Fig. 1. Whole-word segmental model for speech recognition. (smaller) gains can be obtained by pre-training the word pre- Note: boundary frames are not shared. diction layer with AGWEs. Our final models improve over segmental models, where the sequence probability is com- prior A2W models. puted based on segment scores instead of frame probabilities. Index Terms— speech recognition, segmental model, Segmental models have a long history in speech recognition acoustic-to-word, acoustic word embeddings, pre-training research, but they have been used primarily for phonetic recog- nition or as phone-level acoustic models [11–18].
    [Show full text]
  • Arxiv:2104.05274V2 [Cs.CL] 27 Aug 2021 NLP field Due to Its Excellent Performance in Various Tasks
    Learning to Remove: Towards Isotropic Pre-trained BERT Embedding Yuxin Liang1, Rui Cao1, Jie Zheng?1, Jie Ren2, and Ling Gao1 1 Northwest University, Xi'an, China fliangyuxin,[email protected], fjzheng,[email protected] 2 Shannxi Normal University, Xi'an, China [email protected] Abstract. Research in word representation shows that isotropic em- beddings can significantly improve performance on downstream tasks. However, we measure and analyze the geometry of pre-trained BERT embedding and find that it is far from isotropic. We find that the word vectors are not centered around the origin, and the average cosine similar- ity between two random words is much higher than zero, which indicates that the word vectors are distributed in a narrow cone and deteriorate the representation capacity of word embedding. We propose a simple, and yet effective method to fix this problem: remove several dominant directions of BERT embedding with a set of learnable weights. We train the weights on word similarity tasks and show that processed embed- ding is more isotropic. Our method is evaluated on three standardized tasks: word similarity, word analogy, and semantic textual similarity. In all tasks, the word embedding processed by our method consistently out- performs the original embedding (with average improvement of 13% on word analogy and 16% on semantic textual similarity) and two baseline methods. Our method is also proven to be more robust to changes of hyperparameter. Keywords: Natural language processing · Pre-trained embedding · Word representation · Anisotropic. 1 Introduction With the rise of Transformers [12], its derivative model BERT [1] stormed the arXiv:2104.05274v2 [cs.CL] 27 Aug 2021 NLP field due to its excellent performance in various tasks.
    [Show full text]
  • A Survey of Embeddings in Clinical Natural Language Processing Kalyan KS, S
    This paper got published in Journal of Biomedical Informatics. Updated version is available @ https://doi.org/10.1016/j.jbi.2019.103323 SECNLP: A Survey of Embeddings in Clinical Natural Language Processing Kalyan KS, S. Sangeetha Text Analytics and Natural Language Processing Lab Department of Computer Applications National Institute of Technology Tiruchirappalli, INDIA. [email protected], [email protected] ABSTRACT Traditional representations like Bag of words are high dimensional, sparse and ignore the order as well as syntactic and semantic information. Distributed vector representations or embeddings map variable length text to dense fixed length vectors as well as capture prior knowledge which can transferred to downstream tasks. Even though embedding has become de facto standard for representations in deep learning based NLP tasks in both general and clinical domains, there is no survey paper which presents a detailed review of embeddings in Clinical Natural Language Processing. In this survey paper, we discuss various medical corpora and their characteristics, medical codes and present a brief overview as well as comparison of popular embeddings models. We classify clinical embeddings into nine types and discuss each embedding type in detail. We discuss various evaluation methods followed by possible solutions to various challenges in clinical embeddings. Finally, we conclude with some of the future directions which will advance the research in clinical embeddings. Keywords: embeddings, distributed representations, medical, natural language processing, survey 1. INTRODUCTION Distributed vector representation or embedding is one of the recent as well as prominent addition to modern natural language processing. Embedding has gained lot of attention and has become a part of NLP researcher’s toolkit.
    [Show full text]