Quantifying the Contextualization of Word Representations with Semantic Class Probing

Findings of the Association for Computational Linguistics: EMNLP 2020, pages 1219–1234 November 16 - 20, 2020. c⃝2020 Association for Computational Linguistics 1219 Quantifying the Contextualization of Word Representations with Semantic Class Probing Mengjie Zhao†, Philipp Dufter†, Yadollah Yaghoobzadeh‡, Hinrich Schütze† † CIS, LMU Munich, Germany ‡ Microsoft Turing, Montréal, Canada mzhao@cis.lmu.de Abstract Pretrained language models achieve state-of- the-art results on many NLP tasks, but there are still many open questions about how and why they work so well. We investigate the con- textualization of words in BERT. We quantify the amount of contextualization, i.e., how well words are interpreted in context, by studying the extent to which semantic classes of a word can be inferred from its contextualized em- bedding. Quantifying contextualization helps in understanding and utilizing pretrained lan- guage models. We show that the top layer rep- resentations support highly accurate inference of semantic classes; that the strongest contex- tualization effects occur in the lower layers; that local context is mostly sufficient for con- textualizing words; and that top layer repre- sentations are more task-specific after finetun- ing while lower layer representations are more transferable. Finetuning uncovers task-related features, but pretrained knowledge about con- textualization is still well preserved. 1 Introduction Pretrained language models like ELMo (Peters et al., 2018a), BERT (Devlin et al., 2019), and XL- Net (Yang et al., 2019) are top performers in NLP because they learn contextualized representations, i.e., representations that reflect the interpretation of a word in context as opposed to its general mean- ing, which is less helpful in solving NLP tasks. As stated, pretrained language models contextualize words, is clear qualitatively; there has been lit- tle work on investigating contextualization, i.e., to which extent a word can be interpreted in context, quantitatively. We use BERT (Devlin et al., 2019) as our pre- trained language model and quantify contextual- ization by investigating how well BERT infers se- mantic classes (s-classes) of a word in context, e.g., the s-class organization for “Apple" in “Apple stock rises" vs. the s-class food in “Apple juice is healthy". We use s-class inference as a proxy for contextualization since accurate s-class inference reflects a successful contextualization of a word: an effective interpretation of the word in context. We adopt the methodology of probing (Adi et al., 2016; Shi et al., 2016; Belinkov et al., 2017; Liu et al., 2019; Tenney et al., 2019b; Be- linkov and Glass, 2019; Hewitt and Liang, 2019; Yaghoobzadeh et al., 2019): diagnostic classifiers are applied to pretrained language model embed- dings to determine whether they encode desired syntactic or semantic features. By probing for s-classes we quantify directly where and how contextualization happens in BERT. E.g., we find that the strongest contextual inter- pretation effects occur in the lower layers and that the top two layers contribute little to contextual- ization. We also investigate how the amount of context available affects contextualization. In addition, since pretrained language models in practice need to be finetuned on downstream tasks (Devlin et al., 2019; Peters et …

Content truncated to 500 words