Efficient Training of BERT by Progressively Stacking

Total Page:16

File Type:pdf, Size:1020Kb

Efficient Training of BERT by Progressively Stacking Efficient Training of BERT by Progressively Stacking Linyuan Gong 1 Di He 1 Zhuohan Li 1 Tao Qin 2 Liwei Wang 1 3 Tie-Yan Liu 2 Abstract especially in domains that require particular expertise. Unsupervised pre-training is commonly used in In natural language processing, using unsupervised pre- natural language processing: a deep neural net- trained models is one of the most effective ways to help work trained with proper unsupervised prediction train tasks in which labeled information is not rich enough. tasks are shown to be effective in many down- For example, word embedding learned from Wikipedia cor- stream tasks. Because it is easy to create a large pus (Mikolov et al., 2013; Pennington et al., 2014) can monolingual dataset by collecting data from the substantially improve the performance of sentence classifi- Web, we can train high-capacity models. There- cation and textual similarity systems (Socher et al., 2011; fore, training efficiency becomes a critical issue Tai et al., 2015; Kalchbrenner et al., 2014). Recently, pre- even when using high-performance hardware. In trained contextual representation approaches (Devlin et al., this paper, we explore an efficient training method 2018; Radford et al., 2018; Peters et al., 2018) have been for the state-of-the-art bidirectional Transformer developed and shown to be more effective than conven- (BERT) model. By visualizing the self-attention tional word embedding. Different from word embedding distributions of different layers at different po- that only extracts local semantic information of individual sitions in a well-trained BERT model, we find words, pre-trained contextual representations further learn that in most layers, the self-attention distribution sentence-level information by sentence-level encoders. will concentrate locally around its position and the BERT (Devlin et al., 2018) is the current state-of-the-art pre- start-of-sentence token. Motivated by this, we pro- trained contextual representations based on a huge multi- pose the stacking algorithm to transfer knowledge layer Transformer encoder architecture (BERT-Base has from a shallow model to a deep model; then we 110M parameters and BERT-Large has 330M parameters) apply stacking progressively to accelerate BERT and trained by masked language modeling and next-sentence training. Experiments showed that the models prediction tasks. Because these tasks require no human su- trained by our training strategy achieve similar pervision, the size of the available training data easily scales performance to models trained from scratch, but up to billions of tokens. Therefore, the training efficiency our algorithm is much faster. of such a model becomes the most critical issue, and the requirement of extremely high-performance hardware be- comes a barrier to its practical application. 1. Introduction In this paper, we aim to improve the training efficiency of the In recent years, deep neural networks have pushed the limits BERT model from in an algorithmic sense. Our motivation of many applications, including speech recognition (Hinton is from the observation of self-attention layers, which is the et al., 2012), image classification (He et al., 2016), and core component of the BERT model. We visualize a shallow machine translation (Vaswani et al., 2017). The keys to the BERT model and a deep BERT model and then study their success are the advanced neural network architectures and differences and relationships. By carefully investigating massive databases of labeled instances (Deng et al., 2009). the attention distributions in different layers at different However, human annotations may be very costly to collect, positions, we find some interesting phenomena: First, the The work was done while the first and third author were visit- attention distributions of the shallow model are quite similar ing Microsoft Research Asia. 1Key Laboratory of Machine Per- across different position and layers. At any position, the ception, MOE, School of EECS, Peking University 2Microsoft attention distribution is a mixture of two distributions. One Research 3Center for Data Science, Peking University, Beijing distribution is local attention that focuses on neighbors. The Institute of Big Data Research. Correspondence to: Tao Qin <tao- other distribution focuses on the start-of-sentence token. [email protected]>. Second, we find the attention distribution in the shallow Proceedings of the 36 th International Conference on Machine model is similar to that of a deep model. This suggests that Learning, Long Beach, California, PMLR 97, 2019. Copyright such knowledge can be shared from the shallow model to a 2019 by the author(s). Efficient Training of BERT by Progressive Stacking deep model: Once we have a shallow model, we can stack Output the shallow model into a deep model by sharing weight Probabilities between the top self-attention layers and the bottom self- Classifier attention layers, and then fine-tune all the parameters. As we can train the model from a shallow one to a deep one, + training time can be largely reduced as training a shallow model usually requires less time. Feed Forward We conduct extensive experiments on our proposed method to see (1) whether it can improve the training efficiency and Layer Norm convergence rate at the pre-training step, and (2) whether the L x trained model can achieve similar performance compared + to the baseline models. According to our results, we find first during pre-training, our proposed method is about 25% Multi-Head faster than several baselines to achieve the same validation Attention accuracy. Second, our final model is competitive and even better than the baseline model on several downstream tasks. Layer Norm Positional + + 2. Related Work Encoding 2.1. Unsupervised Pre-training in Natural Language Token Segment Processing Embedding Embedding Pre-trained word vectors (Mikolov et al., 2013; Pennington Inputs et al., 2014) have been considered a standard component of most state-of-the-art NLP architectures, especially for those Figure 1. The model architecture of BERT. tasks where the amount of labeled data is not large enough (Socher et al., 2011; Tai et al., 2015; Kalchbrenner et al., 2014). However, these learned word vectors only capture transfer. Chen et al.(2015) tackles the problem about how the semantics of a single word independent of its surround- to train a deep neural network efficiently when we have a ing context. The rich syntactic and semantic structures of shallow neural network. In particular, function-preserving sentences are not effectively exploited. initialization is proposed which first initializes a deep neural Pre-trained contextual representations overcomes the short- network that represents the same function as the shallow comings of traditional word vectors by considering its sur- one, and then continue to train the deep network by standard rounding context. Peters et al.(2018) first train language optimization methods. However, when dealing with sophis- models using stacked LSTMs, and then use the hidden ticated structures such as Transformer, function-preserving states in the stacked LSTMs as the contextual represen- initialization is usually not effective. For example, the basic tation. Since LSTM processes word sequentially, the hidden component in the Transformer is a composition of a self- state of LSTM at one position contains the information of attention layer and a feed-forward layer. According to our the words in previous positions, and thus the representation empirical study, simply setting the feed-forward layer to be contains not only the word semantics but also the sentence near zero and randomly initializing the self-attention layer contexts. Radford et al.(2018) uses advanced self-attention is a function-preserving initialization, but it is ineffective as units instead of LSTM units in language models. Devlin most parameters in the self-attention layer stay untrained. In et al.(2018) further develops a masked language modeling our work, we propose a different and more efficient method task and achieves state-of-the-art performance on multiple to transfer knowledge from shallow models to deep models. natural language understanding tasks. As (masked) lan- guage modeling requires no human labeling effort, billions 3. Method of sentences on the web can be used to train a very deep network. Therefore, a major challenge in learning such a The BERT (Bidirectional Encoder Representation from model is training efficiency. Transformers) model is developed on a multi-layer bidi- rectional Transformer (Vaswani et al., 2017) encoder. The 2.2. Network Training by Knowledge Transfer architecture is shown in Figure1. The encoder consists of L encoder layers, each of which consists of a multi-head Our iterative training method is also closely related to ef- self-attention sub-layer and a feed forward sub-layer: both ficiently training deep neural networks using knowledge of them have residual connections (He et al., 2015). The Efficient Training of BERT by Progressive Stacking feed forward layer (FFN) is point-wise, i.e., it applies inde- pendently to each position of the input. The key component of the Transformer encoder is the multi- head self-attention layer. An attention function can be formulated as querying a dictionary with key-value pairs (Vaswani et al., 2017), e.g., QKT Attention(Q; K; V ) = softmax p · V; dk (1) n ;d n ;d n ;d where Q 2 R q k ;K 2 R e k ;V 2 R e v : dk is the dimension of each key and each query, nq is the number of queries, andpne is the number of key-value entries. T nq ;ne A = softmax(QK = dk) 2 R defines the attention distribution. The output of each query is a weighted average of the rows of V with A as the coefficient. The attention Figure 2. Visualization of attention distributions of BERT-Base. distribution A helps us understand the attention function: For a randomly chosen sample sentence, we visualize the attention A reflects the importance of the i-th key-value entry with distributions of 6 heads from different layers.
Recommended publications
  • Malware Classification with BERT
    San Jose State University SJSU ScholarWorks Master's Projects Master's Theses and Graduate Research Spring 5-25-2021 Malware Classification with BERT Joel Lawrence Alvares Follow this and additional works at: https://scholarworks.sjsu.edu/etd_projects Part of the Artificial Intelligence and Robotics Commons, and the Information Security Commons Malware Classification with Word Embeddings Generated by BERT and Word2Vec Malware Classification with BERT Presented to Department of Computer Science San José State University In Partial Fulfillment of the Requirements for the Degree By Joel Alvares May 2021 Malware Classification with Word Embeddings Generated by BERT and Word2Vec The Designated Project Committee Approves the Project Titled Malware Classification with BERT by Joel Lawrence Alvares APPROVED FOR THE DEPARTMENT OF COMPUTER SCIENCE San Jose State University May 2021 Prof. Fabio Di Troia Department of Computer Science Prof. William Andreopoulos Department of Computer Science Prof. Katerina Potika Department of Computer Science 1 Malware Classification with Word Embeddings Generated by BERT and Word2Vec ABSTRACT Malware Classification is used to distinguish unique types of malware from each other. This project aims to carry out malware classification using word embeddings which are used in Natural Language Processing (NLP) to identify and evaluate the relationship between words of a sentence. Word embeddings generated by BERT and Word2Vec for malware samples to carry out multi-class classification. BERT is a transformer based pre- trained natural language processing (NLP) model which can be used for a wide range of tasks such as question answering, paraphrase generation and next sentence prediction. However, the attention mechanism of a pre-trained BERT model can also be used in malware classification by capturing information about relation between each opcode and every other opcode belonging to a malware family.
    [Show full text]
  • Bros:Apre-Trained Language Model for Understanding Textsin Document
    Under review as a conference paper at ICLR 2021 BROS: A PRE-TRAINED LANGUAGE MODEL FOR UNDERSTANDING TEXTS IN DOCUMENT Anonymous authors Paper under double-blind review ABSTRACT Understanding document from their visual snapshots is an emerging and chal- lenging problem that requires both advanced computer vision and NLP methods. Although the recent advance in OCR enables the accurate extraction of text seg- ments, it is still challenging to extract key information from documents due to the diversity of layouts. To compensate for the difficulties, this paper introduces a pre-trained language model, BERT Relying On Spatiality (BROS), that represents and understands the semantics of spatially distributed texts. Different from pre- vious pre-training methods on 1D text, BROS is pre-trained on large-scale semi- structured documents with a novel area-masking strategy while efficiently includ- ing the spatial layout information of input documents. Also, to generate structured outputs in various document understanding tasks, BROS utilizes a powerful graph- based decoder that can capture the relation between text segments. BROS achieves state-of-the-art results on four benchmark tasks: FUNSD, SROIE*, CORD, and SciTSR. Our experimental settings and implementation codes will be publicly available. 1 INTRODUCTION Document intelligence (DI)1, which understands industrial documents from their visual appearance, is a critical application of AI in business. One of the important challenges of DI is a key information extraction task (KIE) (Huang et al., 2019; Jaume et al., 2019; Park et al., 2019) that extracts struc- tured information from documents such as financial reports, invoices, business emails, insurance quotes, and many others.
    [Show full text]
  • Information Extraction Based on Named Entity for Tourism Corpus
    Information Extraction based on Named Entity for Tourism Corpus Chantana Chantrapornchai Aphisit Tunsakul Dept. of Computer Engineering Dept. of Computer Engineering Faculty of Engineering Faculty of Engineering Kasetsart University Kasetsart University Bangkok, Thailand Bangkok, Thailand [email protected] [email protected] Abstract— Tourism information is scattered around nowa- The ontology is extracted based on HTML web structure, days. To search for the information, it is usually time consuming and the corpus is based on WordNet. For these approaches, to browse through the results from search engine, select and the time consuming process is the annotation which is to view the details of each accommodation. In this paper, we present a methodology to extract particular information from annotate the type of name entity. In this paper, we target at full text returned from the search engine to facilitate the users. the tourism domain, and aim to extract particular information Then, the users can specifically look to the desired relevant helping for ontology data acquisition. information. The approach can be used for the same task in We present the framework for the given named entity ex- other domains. The main steps are 1) building training data traction. Starting from the web information scraping process, and 2) building recognition model. First, the tourism data is gathered and the vocabularies are built. The raw corpus is used the data are selected based on the HTML tag for corpus to train for creating vocabulary embedding. Also, it is used building. The data is used for model creation for automatic for creating annotated data.
    [Show full text]
  • NLP with BERT: Sentiment Analysis Using SAS® Deep Learning and Dlpy Doug Cairns and Xiangxiang Meng, SAS Institute Inc
    Paper SAS4429-2020 NLP with BERT: Sentiment Analysis Using SAS® Deep Learning and DLPy Doug Cairns and Xiangxiang Meng, SAS Institute Inc. ABSTRACT A revolution is taking place in natural language processing (NLP) as a result of two ideas. The first idea is that pretraining a deep neural network as a language model is a good starting point for a range of NLP tasks. These networks can be augmented (layers can be added or dropped) and then fine-tuned with transfer learning for specific NLP tasks. The second idea involves a paradigm shift away from traditional recurrent neural networks (RNNs) and toward deep neural networks based on Transformer building blocks. One architecture that embodies these ideas is Bidirectional Encoder Representations from Transformers (BERT). BERT and its variants have been at or near the top of the leaderboard for many traditional NLP tasks, such as the general language understanding evaluation (GLUE) benchmarks. This paper provides an overview of BERT and shows how you can create your own BERT model by using SAS® Deep Learning and the SAS DLPy Python package. It illustrates the effectiveness of BERT by performing sentiment analysis on unstructured product reviews submitted to Amazon. INTRODUCTION Providing a computer-based analog for the conceptual and syntactic processing that occurs in the human brain for spoken or written communication has proven extremely challenging. As a simple example, consider the abstract for this (or any) technical paper. If well written, it should be a concise summary of what you will learn from reading the paper. As a reader, you expect to see some or all of the following: • Technical context and/or problem • Key contribution(s) • Salient result(s) If you were tasked to create a computer-based tool for summarizing papers, how would you translate your expectations as a reader into an implementable algorithm? This is the type of problem that the field of natural language processing (NLP) addresses.
    [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]
  • Question Answering by Bert
    Running head: QUESTION AND ANSWERING USING BERT 1 Question and Answering Using BERT Suman Karanjit, Computer SCience Major Minnesota State University Moorhead Link to GitHub https://github.com/skaranjit/BertQA QUESTION AND ANSWERING USING BERT 2 Table of Contents ABSTRACT .................................................................................................................................................... 3 INTRODUCTION .......................................................................................................................................... 4 SQUAD ............................................................................................................................................................ 5 BERT EXPLAINED ...................................................................................................................................... 5 WHAT IS BERT? .......................................................................................................................................... 5 ARCHITECTURE ............................................................................................................................................ 5 INPUT PROCESSING ...................................................................................................................................... 6 GETTING ANSWER ........................................................................................................................................ 8 SETTING UP THE ENVIRONMENT. ....................................................................................................
    [Show full text]
  • Distilling BERT for Natural Language Understanding
    TinyBERT: Distilling BERT for Natural Language Understanding Xiaoqi Jiao1∗,y Yichun Yin2∗z, Lifeng Shang2z, Xin Jiang2 Xiao Chen2, Linlin Li3, Fang Wang1z and Qun Liu2 1Key Laboratory of Information Storage System, Huazhong University of Science and Technology, Wuhan National Laboratory for Optoelectronics 2Huawei Noah’s Ark Lab 3Huawei Technologies Co., Ltd. fjiaoxiaoqi,[email protected] fyinyichun,shang.lifeng,[email protected] fchen.xiao2,lynn.lilinlin,[email protected] Abstract natural language processing (NLP). Pre-trained lan- guage models (PLMs), such as BERT (Devlin et al., Language model pre-training, such as BERT, has significantly improved the performances 2019), XLNet (Yang et al., 2019), RoBERTa (Liu of many natural language processing tasks. et al., 2019), ALBERT (Lan et al., 2020), T5 (Raf- However, pre-trained language models are usu- fel et al., 2019) and ELECTRA (Clark et al., 2020), ally computationally expensive, so it is diffi- have achieved great success in many NLP tasks cult to efficiently execute them on resource- (e.g., the GLUE benchmark (Wang et al., 2018) restricted devices. To accelerate inference and the challenging multi-hop reasoning task (Ding and reduce model size while maintaining et al., 2019)). However, PLMs usually have a accuracy, we first propose a novel Trans- former distillation method that is specially de- large number of parameters and take long infer- signed for knowledge distillation (KD) of the ence time, which are difficult to be deployed on Transformer-based models. By leveraging this edge devices such as mobile phones. Recent stud- new KD method, the plenty of knowledge en- ies (Kovaleva et al., 2019; Michel et al., 2019; Voita coded in a large “teacher” BERT can be ef- et al., 2019) demonstrate that there is redundancy fectively transferred to a small “student” Tiny- in PLMs.
    [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]
  • Word Sense Disambiguation with Transformer Models
    Word Sense Disambiguation with Transformer Models Pierre-Yves Vandenbussche Tony Scerri Ron Daniel Jr. Elsevier Labs Elsevier Labs Elsevier Labs Radarweg 29, 1 Appold Street, 230 Park Avenue, Amsterdam 1043 NX, London EC2A 2UT, UK New York, NY, 10169, Netherlands [email protected] USA [email protected] [email protected] Abstract given hypernym. In Subtask 3 the system can use both the sentence and the hypernym in making the In this paper, we tackle the task of Word decision. Sense Disambiguation (WSD). We present our system submitted to the Word-in-Context Tar- The dataset provided with the WiC-TSV chal- get Sense Verification challenge, part of the lenge has relatively few sense annotated examples SemDeep workshop at IJCAI 2020 (Breit et al., (< 4; 000) and with a single target sense per word. 2020). That challenge asks participants to pre- This makes pre-trained Transformer models well dict if a specific mention of a word in a text suited for the task since the small amount of data matches a pre-defined sense. Our approach would limit the learning ability of a typical super- uses pre-trained transformer models such as vised model trained from scratch. BERT that are fine-tuned on the task using different architecture strategies. Our model Thanks to the recent advances made in language achieves the best accuracy and precision on models such as BERT (Devlin et al., 2018) or XL- Subtask 1 – make use of definitions for decid- Net (Yang et al., 2019) trained on large corpora, ing whether the target word in context corre- neural language models have established the state- sponds to the given sense or not.
    [Show full text]
  • Improving the Prosody of RNN-Based English Text-To-Speech Synthesis by Incorporating a BERT Model
    INTERSPEECH 2020 October 25–29, 2020, Shanghai, China Improving the Prosody of RNN-based English Text-To-Speech Synthesis by Incorporating a BERT model Tom Kenter, Manish Sharma, Rob Clark Google UK ftomkenter, skmanish, [email protected] Abstract the pretrained model, whereas it is typically hard for standard parsing methods to resolve this ambiguity. Sentences that are The prosody of currently available speech synthesis systems challenging linguistically (e.g., because they are long or other- can be unnatural due to the systems only having access to the wise difficult to parse) can benefit from the new approach based text, possibly enriched by linguistic information such as part- on BERT representations rather than explicit features, as any of-speech tags and parse trees. We show that incorporating errors in the parse information no longer impair performance. a BERT model in an RNN-based speech synthesis model — where the BERT model is pretrained on large amounts of un- We show that human raters favour CHiVE-BERT over a labeled data, and fine-tuned to the speech domain — improves current state-of-the-art model on three datasets designed to prosody. Additionally, we propose a way of handling arbitrar- highlight different prosodic phenomena. Additionally, we pro- ily long sequences with BERT. Our findings indicate that small pose a way of handling long sequences with BERT (which has BERT models work better than big ones, and that fine-tuning fixed-length inputs and outputs), so it can deal with arbitrarily the BERT part of the model is pivotal for getting good results. long input.
    [Show full text]
  • Recombining Frames and Roles in Frame-Semantic Parsing
    Breeding Fillmore’s Chickens and Hatching the Eggs: Recombining Frames and Roles in Frame-Semantic Parsing Gosse Minnema and Malvina Nissim Center for Language and Cognition University of Groningen, The Netherlands fg.f.minnema, [email protected] Abstract (given a predicate-frame pair, find and label its argu- ments). Some recent systems, such as the LSTM- Frame-semantic parsers traditionally predict based Open-SESAME and (Swayamdipta et al., predicates, frames, and semantic roles in a 2017) or the classical-statistical SEMAFOR (Das fixed order. This paper explores the ‘chicken- et al., 2014), implement the full pipeline, but with a or-egg’ problem of interdependencies between these components theoretically and practically. strong focus specifically on argID. Other models We introduce a flexible BERT-based sequence implement some subset of the components (Tan, labeling architecture that allows for predict- 2007; Hartmann et al., 2017; Yang and Mitchell, ing frames and roles independently from each 2017; Peng et al., 2018), while still implicitly adopt- other or combining them in several ways. ing the pipeline’s philosophy.1 However, little fo- Our results show that our setups can approxi- cus has been given to frame-semantic parsing as mate more complex traditional models’ perfor- an end-to-end task, which entails not only imple- mance, while allowing for a clearer view of the interdependencies between the pipeline’s com- menting the separate components of the pipeline, ponents, and of how frame and role prediction but also looking at their interdependencies. models make different use of BERT’s layers. We highlight such interdependencies from a the- oretical perspective, and investigate them empiri- 1 Introduction cally.
    [Show full text]
  • A Lite BERT for Self-Supervised Learning of Audio Representation
    Audio ALBERT: A Lite BERT for Self-supervised Learning of Audio Representation Po-Han Chi 1 Pei-Hung Chung 1 Tsung-Han Wu 1 Chun-Cheng Hsieh 1 Shang-Wen Li 2 3 Hung-yi Lee 1 Abstract model learns the robust speech representations for speech For self-supervised speech processing, it is crucial processing tasks, for example, ASR and speaker recogni- to use pretrained models as speech representation tion, with the self-supervised learning approaches (Liu et al., extractors. In recent works, increasing the size 2019a; Jiang et al., 2019; Ling et al., 2019; Baskar et al., of the model has been utilized in acoustic model 2019; Schneider et al., 2019). However, since the size of training in order to achieve better performance. In the pretraining models, no matter the text or speech ver- this paper, we propose Audio ALBERT, a lite ver- sions is usually prohibitively large, they require a significant sion of the self-supervised speech representation amount of memory for computation, even at the fine-tuning model. We use the representations with two down- stage. The requirement hinders the application of pretrained stream tasks, speaker identification, and phoneme models from different downstream tasks. classification. We show that Audio ALBERT is ca- ALBERT (Lan et al., 2019) is a lite version of BERT for pable of achieving competitive performance with text by sharing one layer parameters across all layers and those huge models in the downstream tasks while factorizing the embedding matrix to reduce most parame- utilizing 91% fewer parameters. Moreover, we ters. Although the number of parameters is reduced, the use some simple probing models to measure how representations learned in ALBERT are still robust and task much the information of the speaker and phoneme agnostic, such that ALBERT can achieve similar perfor- is encoded in latent representations.
    [Show full text]