On Continual Model Refinement in Out-of-Distribution Data Streams

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 3128 - 3139 May 22-27, 2022 c⃝2022 Association for Computational Linguistics On Continual Model Refinement in Out-of-Distribution Data Streams Bill Yuchen Lin†∗ Sida Wang‡ Xi Victoria Lin‡ Robin Jia† Lin Xiao‡ Xiang Ren† Wen-tau Yih‡ ‡ Facebook AI Research †University of Southern California {yuchen.lin,robinjia,xiangren}@usc.edu,{sida,victorialin,linx,scottyih}@fb.com Abstract Real-world natural language processing (NLP) models need to be continually updated to fix the prediction errors in out-of-distribution (OOD) data streams while overcoming catastrophic forgetting. However, existing continual learn- ing (CL) problem setups cannot cover such a realistic and complex scenario. In response to this, we propose a new CL problem for- mulation dubbed continual model refinement (CMR). Compared to prior CL settings, CMR is more practical and introduces unique chal- lenges (boundary-agnostic and non-stationary distribution shift, diverse mixtures of multiple OOD data clusters, error-centric streams, etc.). We extend several existing CL approaches to the CMR setting and evaluate them extensively. For benchmarking and analysis, we propose a general sampling algorithm to obtain dy- namic OOD data streams with controllable non- stationarity, as well as a suite of metrics mea- suring various aspects of online performance. Our experiments and detailed analysis reveal the promise and challenges of the CMR prob- lem, supporting that studying CMR in dynamic OOD streams can benefit the longevity of de- ployed NLP models in production. 1 1 Introduction Fine-tuning large pre-trained language models (LMs) has become the de facto standard for train- ing models of a variety of tasks in natural language processing (NLP). These success stories are usu- ally in places where the training and testing data are drawn from the same distribution. However, in real-world scenarios, a deployed model (e.g., a question answering service) often encounters exam- ples that are out of the training distribution (i.e., out- of-distribution, OOD). Such distribution shift often leads to a high error rate. In practice, it is highly ∗The work was done when Bill was an intern at FAIR. 1Our code and data are available at the project website — https://cmr-nlp.github.io/. preferred to continually refine deployed models whenever new errors are reported and annotated, in order to reduce their further negative impacts. In spite of its importance, the challenge of con- tinually refining a model over OOD data streams has been underexplored. Prior work in contin- ual learning (CL) has primarily focused on task- incremental settings with boundary-aware data streams. These CL methods are usually evaluated on simple models and data (e.g., image classifica- tion with MNIST) (Aljundi et al., 2019). It is not clear to what extent they can efficiently refine a model in boundary-agnostic streams for a complex language task (e.g., reading comprehen- sion) with modern LMs. In addition, there is no existing evaluation protocol for comprehensively comparing the collection of applicable methods for such a practical and complex problem. Traditional CL paradigms mainly focus on incrementally learn- ing a model from a data stream with a sequence of distinct tasks with explicit …

Content truncated to 500 words