Analyzing the Forgetting Problem in Pretrain-Finetuning of Open-domain Dialogue Response Models

By Myle Ott. 2021

61 citations recorded by the source corpus.

In this work, we study how the finetuning stage in the pretrain-finetune framework changes the behavior of a pretrained neural language generator. We focus on the transformer encoder-decoder model for the open-domain dialogue response generation task. Our major finding is that after standard finetuning, the model forgets some of the important language generation skills acquired during large-scale pretraining. We demonstrate the forgetting phenomenon through a set of detailed behavior analysis from the perspectives of knowledge transfer, context sensitivity, and function space projection. As a preliminary attempt to alleviate the forgetting problem, we propose an intuitive finetuning strategy named “mix-review”. We find that mix-review effectively regularizes the finetuning process, and the forgetting problem is alleviated to some extent. Finally, we discuss interesting behavior of the resulting dialogue model and its implications.

Subfields: Artificial Intelligence

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

Citations in the corpus

Towards Continual Knowledge Learning of Language Models (2021, methodology): “Mix-Review (He et al., 2021) falls into the category of rehearsal methods, which assumes access to the initial pretraining corpus and mixes in random subsets of the initial pretraining data during continued pretraining, depending on the mix-ratio at the current time step.”
Overcoming Catastrophic Forgetting During Domain Adaptation of Seq2seq Language Generation (2022, neutral): “Recent work in this area includes (Mi et al., 2020; Madotto et al., 2020; He et al., 2021; Shin et al., 2017).”
Exposure Bias versus Self-Recovery: Are Distortions Really Incremental for Autoregressive Text Generation? (2019, background): “As shown by recent studies, there exists important problems other than exposure bias for the current NLG models, such as the likelihood trap (Holtzman et al., 2020), factuality (Massarelli et al., 2019; He et al., 2021), or robustness (Cheng et al., 2018; Kassner and Schütze, 2019), etc.”
Perturbation Augmentation for Fairer NLP (2022, background): “However, there is always a worry that intrinsic metrics may not be predictive of what happens downstream, as finetuning can overwrite some of what is learned in pretraining (Zhao and Bethard, 2020; He et al., 2021), a situation we call “catastrophic forgetting”, when something we like gets overwritt”
Continual Learning of Large Language Models: A Comprehensive Survey (2024, unknown): “The bottom layer, CFT, targets models for final downstream tasks on the consumer side (Section 4.3), where the model needs to be updated after deployment for the specified task.”
Knowledge Conflicts for LLMs: A Survey (2024, unknown): “CL suffers from the issue of catastrophic forgetting and demands significant computational resources (De Lange et al., 2021; He et al., 2021; Wang et al., 2023d).”
TemporalWiki: A Lifelong Benchmark for Training and Evaluating Ever-Evolving Language Models (2022, methodology): “Among continual learning meth-ods, K-A DAPTER and L O RA shows higher overall perplexities than D IFF while M IX - REVIEW and R EC A DAM shows similar perplexity.”