Q2: Evaluating Factual Consistency in Knowledge-Grounded Dialogues via Question Generation and Question Answering

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 792–806 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 792 Exploring the Role of BERT Token Representations to Explain Sentence Probing Results Hosein Mohebbi⋆♥ Ali Modarressi⋆♥ Mohammad Taher Pilehvar♠ ♥ Iran University of Science and Technology, Iran ♠ Tehran Institute for Advanced Studies, Khatam University, Iran {hosein_mohebbi, m_modarressi}@comp.iust.ac.ir mp792@cam.ac.uk Abstract Several studies have been carried out on re- vealing linguistic features captured by BERT. This is usually achieved by training a diagnos- tic classifier on the representations obtained from different layers of BERT. The subsequent classification accuracy is then interpreted as the ability of the model in encoding the corre- sponding linguistic property. Despite provid- ing insights, these studies have left out the po- tential role of token representations. In this pa- per, we provide a more in-depth analysis on the representation space of BERT in search for distinct and meaningful subspaces that can explain the reasons behind these probing re- sults. Based on a set of probing tasks and with the help of attribution methods we show that BERT tends to encode meaningful knowl- edge in specific token representations (which are often ignored in standard classification se- tups), allowing the model to detect syntac- tic and semantic abnormalities, and to distinc- tively separate grammatical number and tense subspaces.1 1 Introduction Recent years have seen a surge of interest in pre- trained language models, highlighted by extensive research around BERT (Devlin et al., 2019) and its derivatives. One strand of research has focused on enhancing existing models with the primary ob- jective of improving downstream performance on various NLP tasks (Liu et al., 2019b; Lan et al., 2019; Yang et al., 2019). Another strand analyzes the behaviour of these models with the hope of get- ting better insights for further developments (Clark et al., 2019; Kovaleva et al., 2019; Jawahar et al., 2019; Tenney et al., 2019; Lin et al., 2019). Probing is one of the popular analysis methods, often used for investigating the encoded knowledge Authors marked with a star (⋆) contributed equally. 1Code is available at https://github.com/ hmohebbi/explain-probing-results in language models (Conneau et al., 2018; Ten- ney et al., 2018). This is typically carried out by training a set of diagnostic classifiers that predict a specific linguistic property based on the representa- tions obtained from different layers. Recent works in probing language models demonstrate that ini- tial layers are responsible for encoding low-level linguistic information, such as part of speech and positional information, whereas intermediate layers are better at syntactic phenomena, such as syntac- tic tree depth or subject-verb agreement, while in general semantic information is spread across the entire model (Lin et al., 2019; Peters et al., 2018; Liu et al., 2019a; Hewitt and Manning, 2019; Ten- ney et al., 2019). Despite elucidating the type of knowledge encoded in various layers, these studies do not go further to investigate the reasons behind the layer-wise behavior and the role played by to- ken representations. Analyzing …

Content truncated to 500 words