Assessing Phrasal Representation and Composition in Transformers

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 4896–4907, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 4896 Assessing Phrasal Representation 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 Deep transformer models have pushed perfor- mance on NLP tasks to new limits, suggesting sophisticated treatment of complex linguistic inputs, such as phrases. However, we have lim- ited understanding of how these models han- dle representation of phrases, and whether this reflects sophisticated composition of phrase meaning like that done by humans. In this pa- per, we present systematic analysis of phrasal representations in state-of-the-art pre-trained transformers. We use tests leveraging human judgments of phrase similarity and meaning shift, and compare results before and after con- trol of word overlap, to tease apart lexical ef- fects versus composition effects. We find that phrase representation in these models relies heavily on word content, with little evidence of nuanced composition. We also identify vari- ations in phrase representation quality across models, layers, and representation types, and make corresponding recommendations for us- age of representations from these models. 1 Introduction A fundamental component of language understand- ing is the capacity to combine meaning units into larger units—a phenomenon known as composi- tion—and to do so in a way that reflects the nuances of meaning as understood by humans. Transform- ers (Vaswani et al., 2017) have shown impressive performance in NLP, particularly transformers us- ing pre-training, like BERT (Devlin et al., 2019) and GPT (Radford et al., 2018, 2019), suggesting that these models may be succeeding at composi- tion of complex meanings. However, because trans- formers (like other contextual embedding models) typically maintain representations for every token, it is unclear how and at what points they might be combining word meanings into phrase mean- ings. This contrasts with models that incorporate explicit phrasal composition into their architecture, e.g. RNNG (Dyer et al., 2016; Kim et al., 2019), recursive models for semantic composition (Socher et al., 2013), or transformers with attention-based composition modules (Yin et al., 2020). In this paper we take steps to clarify the nature of phrasal representation in transformers. We fo- cus on representation of two-word phrases, and we prioritize identifying and teasing apart two impor- tant but distinct notions: how faithfully the mod- els are representing information about the words that make up the phrase, and how faithfully the models are representing the nuances of the com- posed phrase meaning itself, over and above a sim- ple account of the component words. To do this, we begin with existing methods for testing how well representations align with human judgments of meaning similarity: similarity correlations and paraphrase classification. We then introduce con- trolled variants of these datasets, removing cues of word overlap, in order to distinguish effects of word content from effects of more sophisticated composition. We complement these phrase simi- larity analyses with classic sense selection tests of phrasal composition (Kintsch, …

Content truncated to 500 words