Investigating BERT’s Knowledge of Language: Five Analysis Methods with NPIs

Though state-of-the-art sentence representation models can perform tasks requiring significant knowledge of grammar, it is an open question how best to evaluate their grammatical knowledge. We explore five experimental methods inspired by prior work evaluating pretrained sentence representation models. We use a single linguistic phenomenon, negative polarity item (NPI) licensing in English, as a case study for our experiments. NPIs like "any" are grammatical only if they appear in a licensing environment like negation ("Sue doesn't have any cats" vs. "Sue has any cats"). This phenomenon is challenging because of the variety of NPI licensing environments that exist. We introduce an artificially generated dataset that manipulates key features of NPI licensing for the experiments. We find that BERT has significant knowledge of these features, but its success varies widely across different experimental methods. We conclude that a variety of methods is necessary to reveal all relevant aspects of a model's grammatical knowledge in a given domain.

[1]  Jeffrey Pennington,et al.  GloVe: Global Vectors for Word Representation , 2014, EMNLP.

[2]  Dieuwke Hupkes,et al.  Do Language Models Understand Anything? On the Ability of LSTMs to Understand Negative Polarity Items , 2018, BlackboxNLP@EMNLP.

[3]  Tal Linzen,et al.  Targeted Syntactic Evaluation of Language Models , 2018, EMNLP 2018.

[4]  Noam Chomsky,et al.  Lectures on Government and Binding , 1981 .

[5]  F. Wilcoxon.  Individual Comparisons by Ranking Methods , 1945 .

[6]  Anastasia Giannakidou,et al.  Polarity sensitivity as (non) veridical dependency , 2000 .

[7]  G. Chierchia,et al.  Logic in Grammar: Polarity, Free Choice, and Intervention , 2013 .

[8]  Samuel R. Bowman,et al.  Grammatical Analysis of Pretrained Sentence Encoders with Acceptability Judgments , 2019, ArXiv.

[9]  Samuel R. Bowman,et al.  Sentence Encoders on STILTs: Supplementary Training on Intermediate Labeled-data Tasks , 2018, ArXiv.

[10]  Allyson Ettinger,et al.  Assessing Composition in Sentence Vector Representations , 2018, COLING.

[11]  Emmanuel Dupoux,et al.  Assessing the Ability of LSTMs to Learn Syntax-Sensitive Dependencies , 2016, TACL.

[12]  Samuel R. Bowman,et al.  Linguistic Analysis of Pretrained Sentence Encoders with Acceptability Judgments , 2019 .

[13]  Alexander Clark,et al.  Grammaticality, Acceptability, and Probability: A Probabilistic View of Linguistic Knowledge , 2017, Cogn. Sci..

[14]  F. Zwarts.  Three Types of Polarity , 1997 .

[15]  B. Matthews.  Comparison of the predicted and observed secondary structure of T4 phage lysozyme. , 1975, Biochimica et biophysica acta.

[16]  Xiaodong Liu,et al.  Multi-Task Deep Neural Networks for Natural Language Understanding , 2019, ACL.

[17]  Xing Shi,et al.  Does String-Based Neural MT Learn Source Syntax? , 2016, EMNLP.

[18]  Noam Chomsky,et al.  वाक्यविन्यास का सैद्धान्तिक पक्ष = Aspects of the theory of syntax , 1965 .

[19]  Allyson Ettinger,et al.  Probing for semantic evidence of composition by means of simple classification tasks , 2016, RepEval@ACL.

[20]  M. Linebarger.  The grammar of negative polarity , 1981 .

[21]  Florian Mohnert,et al.  Under the Hood: Using Diagnostic Classifiers to Investigate and Improve how Language Models Track Agreement Information , 2018, BlackboxNLP@EMNLP.

[22]  Samuel R. Bowman,et al.  Neural Network Acceptability Judgments , 2018, Transactions of the Association for Computational Linguistics.

[23]  Omer Levy,et al.  GLUE: A Multi-Task Benchmark and Analysis Platform for Natural Language Understanding , 2018, BlackboxNLP@EMNLP.

[24]  M. Krifka,et al.  The Semantics and Pragmatics of Polarity Items , 2003 .

[25]  Iain Murray,et al.  BERT and PALs: Projected Attention Layers for Efficient Adaptation in Multi-Task Learning , 2019, ICML.

[26]  Mark Steedman,et al.  CCGbank: A Corpus of CCG Derivations and Dependency Structures Extracted from the Penn Treebank , 2007, CL.

[27]  Alex Wang,et al.  What do you learn from context? Probing for sentence structure in contextualized word representations , 2019, ICLR.

[28]  Ming-Wei Chang,et al.  BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding , 2019, NAACL.

[29]  Noam Chomsky,et al.  The Minimalist Program , 1992 .

[30]  Samuel R. Bowman,et al.  A Broad-Coverage Challenge Corpus for Sentence Understanding through Inference , 2017, NAACL.

[31]  Yonatan Belinkov,et al.  Fine-grained Analysis of Sentence Embeddings Using Auxiliary Prediction Tasks , 2016, ICLR.

[32]  Gilles Fauconnier,et al.  Polarity and the Scale Principle , 1975 .

[33]  Luke S. Zettlemoyer,et al.  AllenNLP: A Deep Semantic Natural Language Processing Platform , 2018, ArXiv.

[34]  Roger Levy,et al.  Structural Supervision Improves Learning of Non-Local Grammatical Dependencies , 2019, NAACL.

[35]  Holger Schwenk,et al.  Supervised Learning of Universal Sentence Representations from Natural Language Inference Data , 2017, EMNLP.

[36]  Guillaume Lample,et al.  What you can cram into a single $&!#* vector: Probing sentence embeddings for linguistic properties , 2018, ACL.

[37]  Samuel R. Bowman,et al.  Verb Argument Structure Alternations in Word and Sentence Embeddings , 2018, ArXiv.

[38]  Noah D. Goodman,et al.  Evaluating Compositionality in Sentence Embeddings , 2018, CogSci.

[39]  Alex Wang,et al.  jiant: A Software Toolkit for Research on General-Purpose Text Understanding Models , 2020, ACL.

[40]  Beatrice Santorini,et al.  Building a Large Annotated Corpus of English: The Penn Treebank , 1993, CL.

[41]  Alex Wang,et al.  Looking for ELMo's friends: Sentence-Level Pretraining Beyond Language Modeling , 2018, ArXiv.