Similarity Analysis of Contextual Word Representation Models

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4638–4655 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 4638 Similarity Analysis of Contextual Word Representation Models John M. Wu∗1 Yonatan Belinkov*12 Hassan Sajjad3 Nadir Durrani3 Fahim Dalvi3 James Glass1 1MIT Computer Science and Artificial Intelligence Laboratory, Cambridge, MA, USA 2Harvard John A. Paulson School of Engineering and Applied Sciences, Cambridge, MA, USA 3Qatar Computing Research Institute, HBKU Research Complex, Doha 5825, Qatar {johnmwu,belinkov,glass}@csail.mit.edu {hsajjad,ndurrani,faimaduddin}@qf.org.qa Abstract This paper investigates contextual word repre- sentation models from the lens of similarity analysis. Given a collection of trained mod- els, we measure the similarity of their inter- nal representations and attention. Critically, these models come from vastly different archi- tectures. We use existing and novel similarity measures that aim to gauge the level of local- ization of information in the deep models, and facilitate the investigation of which design fac- tors affect model similarity, without requiring any external linguistic annotation. The analy- sis reveals that models within the same fam- ily are more similar to one another, as may be expected. Surprisingly, different architectures have rather similar representations, but differ- ent individual neurons. We also observed dif- ferences in information localization in lower and higher layers and found that higher lay- ers are more affected by fine-tuning on down- stream tasks.1 1 Introduction Contextual word representations such as ELMo (Pe- ters et al., 2018a) and BERT (Devlin et al., 2019) have led to impressive improvements in a variety of tasks. With this progress in breaking the state of the art, interest in the community has expanded to analyzing such models in an effort to illuminate their inner workings. A number of studies have analyzed the internal representations in such mod- els and attempted to assess what linguistic proper- ties they capture. A prominent methodology for this is to train supervised classifiers based on the models’ learned representations, and predict var- ious linguistic properties. For instance, Liu et al. (2019a) train such classifiers on 16 linguistic tasks, including part-of-speech tagging, chunking, named ∗Equal contribution 1The code is available at https://github.com/ johnmwu/contextual-corr-analysis. entity recognition, and others. Such an approach may reveal how well representations from different models, and model layers, capture different proper- ties. This approach, known as analysis by probing classifiers, has been used in numerous other stud- ies (Belinkov and Glass, 2019). While the above approach yields compelling in- sights, its applicability is constrained by the avail- ability of linguistic annotations. In addition, com- parisons of different models are indirect, via the probing accuracy, making it difficult to comment on the similarities and differences of different mod- els. In this paper, we develop complementary meth- ods for analyzing contextual word representations based on their inter- and intra-similarity. While this similarity analysis does not tell us absolute facts about a model, it allows comparing representations without subscribing to one type of information. We consider several kinds of similarity measures based on different levels of localization/distributivity of information: from neuron-level …

Content truncated to 500 words