On the Interplay Between Fine-tuning and Composition in Transformers

Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 2279–2293 August 1–6, 2021. ©2021 Association for Computational Linguistics 2279 On the Interplay Between Fine-tuning and Composition in Transformers Lang Yu Deptartment of Computer Science University of Chicago langyu@uchicago.edu Allyson Ettinger Department of Linguistics University of Chicago aettinger@uchicago.edu Abstract Pre-trained transformer language models have shown remarkable performance on a variety of NLP tasks. However, recent research has suggested that phrase-level representations in these models reflect heavy influences of lexi- cal content, but lack evidence of sophisticated, compositional phrase information (Yu and Et- tinger, 2020). Here we investigate the impact of fine-tuning on the capacity of contextual- ized embeddings to capture phrase meaning information beyond lexical content. Specifi- cally, we fine-tune models on an adversarial paraphrase classification task with high lexical overlap, and on a sentiment classification task. After fine-tuning, we analyze phrasal represen- tations in controlled settings following prior work. We find that fine-tuning largely fails to benefit compositionality in these represen- tations, though training on sentiment yields a small, localized benefit for certain models. In follow-up analyses, we identify confounding cues in the paraphrase dataset that may explain the lack of composition benefits from that task, and we discuss potential factors underlying the localized benefits from sentiment training. 1 Introduction Transformer language models like BERT (Devlin et al., 2019), GPT (Radford et al., 2018, 2019) and XLNet (Yang et al., 2019b), have improved the state-of-art in many NLP tasks since their introduc- tion. The versatility of these pre-trained models suggests that they may acquire fairly robust linguis- tic knowledge and capacity for natural language “understanding”. However, an emerging body of analysis demonstrates a level of superficiality in these models’ handling of language (Niven and Kao, 2019; Kim and Linzen, 2020; McCoy et al., 2019; Ettinger, 2020; Yu and Ettinger, 2020). In particular, although composition—a model’s capacity to combine meaning units into more com- plex units reflecting phrase meanings—is an indis- pensable component of language understanding, when testing for composition in pre-trained trans- former representations, Yu and Ettinger (2020) re- port that these representations reflect word content of phrases, but don’t show signs of more sophisti- cated humanlike composition beyond word content. In the present paper we perform a direct follow- up of that study, asking whether models will show better evidence of composition after fine-tuning on tasks that are good candidates for requiring composition: 1) the Quora Question Pairs dataset in Paraphrase Adversaries from Word Scrambling (PAWS-QQP) (Zhang et al., 2019a), an adversar- ial paraphrase dataset forcing models to classify paraphrases with high lexical overlap, and 2) the Stanford Sentiment Treebank (Socher et al., 2013), a sentiment dataset with fine-grained phrase labels to promote composition. We base our analysis on the tests proposed by Yu and Ettinger (2020), which rely on alignment with human judgments of phrase pair similarities, and which leverage control of lexi- cal overlap to target compositionality. We fine-tune and evaluate the same models and representation types tested in that paper, for optimal comparison. We find …

Content truncated to 500 words