ConvFiT: Conversational Fine-Tuning of Pretrained Language Models

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 1151–1168 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 1151 CONVFIT: Conversational Fine-Tuning of Pretrained Language Models Ivan Vuli´c, Pei-Hao Su, Sam Coope, Daniela Gerz, Paweł Budzianowski, Iñigo Casanueva, Nikola Mrkši´c, and Tsung-Hsien Wen PolyAI Limited London, United Kingdom www.polyai.com Abstract Transformer-based language models (LMs) pretrained on large text collections are proven to store a wealth of semantic knowledge. However, 1) they are not effective as sen- tence encoders when used off-the-shelf, and 2) thus typically lag behind conversationally pretrained (e.g., via response selection) en- coders on conversational tasks such as intent detection (ID). In this work, we propose CON- VFIT, a simple and efficient two-stage proce- dure which turns any pretrained LM into a universal conversational encoder (after Stage 1 CONVFIT-ing) and task-specialised sentence encoder (after Stage 2). We demonstrate that 1) full-blown conversational pretraining is not required, and that LMs can be quickly trans- formed into effective conversational encoders with much smaller amounts of unannotated data; 2) pretrained LMs can be fine-tuned into task-specialised sentence encoders, opti- mised for the fine-grained semantics of a par- ticular task. Consequently, such specialised sentence encoders allow for treating ID as a simple semantic similarity task based on inter- pretable nearest neighbours retrieval. We vali- date the robustness and versatility of the CON- VFIT framework with such similarity-based inference on the standard ID evaluation sets: CONVFIT-ed LMs achieve state-of-the-art ID performance across the board, with particular gains in the most challenging, few-shot setups. 1 Introduction and Motivation Pretrained Transformer-based (masked) language models (LMs) such as BERT (Devlin et al., 2019) or RoBERTa (Liu et al., 2019b), coupled with task-specific fine-tuning, offer unmatched state-of- the-art performance in a wide array of standard language understanding and conversational tasks (Wang et al., 2019a; Mehri et al., 2020). However, pretrained LMs do not produce coherent and effec- tive sentence encodings off-the-shelf; their further adaptation is required, akin to standard task fine- tuning. For instance, Reimers and Gurevych (2019) transform monolingual English BERT with super- vised natural language inference and paraphrasing data (Williams et al., 2018; Wieting and Gimpel, 2018) into a sentence encoder which excels at sen- tence similarity and retrieval tasks (Marelli et al., 2014; Cer et al., 2017). This transformation pro- cess supports the creation of other similar universal sentence encoders in monolingual and multilingual settings (Chidambaram et al., 2019; Wieting et al., 2020; Feng et al., 2020), and is typically based on dual-encoder architectures. Another parallel research thread aims at learning conversational encoders: it validates the benefits of masked language modeling (MLM) pretraining on naturally conversational data (Wu et al., 2020; Mehri et al., 2021), as well as the benefits of trans- fer learning for conversational tasks which goes beyond MLM as the pretraining objective (Mehri et al., 2019; Coope et al., 2020; Henderson and Vuli´c, 2021, inter alia). In particular, response selection as a suitable pretraining task (Al-Rfou et al., 2016; Yang et al., 2018; Henderson et al., 2019b; …

Content truncated to 500 words