Breakpoint Transformers for Modeling and Tracking Intermediate Beliefs

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 9703 - 9719 December 7-11, 2022 ©2022 Association for Computational Linguistics Breakpoint Transformers for Modeling and Tracking Intermediate Beliefs Kyle Richardson2† Ronen Tamari1†∗ Oren Sultan1 Reut Tsarfaty2,3 Dafna Shahaf1 Ashish Sabharwal2 1The Hebrew University of Jerusalem 2Allen Institute for AI 3Bar-Ilan University {ronent,orens,dshahaf}@cs.huji.ac.il, {kyler,reutt,ashishs}@allenai.org Abstract Can we teach natural language understanding models to track their beliefs through intermedi- ate points in text? We propose a representation learning framework called breakpoint modeling that allows for learning of this type. Given any text encoder and data marked with intermediate states (breakpoints) along with corresponding textual queries viewed as true/false proposi- tions (i.e., the candidate beliefs of a model, con- sisting of information changing through time) our approach trains models in an efficient and end-to-end fashion to build intermediate rep- resentations that facilitate teaching and direct querying of beliefs at arbitrary points alongside solving other end tasks. To show the benefit of our approach, we experiment with a diverse set of NLU tasks including relational reasoning on CLUTRR and narrative understanding on bAbI. Using novel belief prediction tasks for both tasks, we show the benefit of our main breakpoint transformer, based on T5, over con- ventional representation learning approaches in terms of processing efficiency, prediction accuracy and prediction consistency, all with minimal to no effect on corresponding QA end- tasks. To show the feasibility of incorporating our belief tracker into more complex reasoning pipelines, we also obtain SOTA performance on the three-tiered reasoning challenge for the TRIP benchmark (around 23-32% absolute im- provement on Tasks 2-3).1 1 Introduction Despite considerable progress made recently in nat- ural language understanding (NLU), driven largely by advances in language model pre-training (De- vlin et al., 2019; Raffel et al., 2020) and the de- velopment of large-scale NLU benchmarks (Wang et al., 2018), understanding the behavior of mod- els remains a formidable and highly consequential ∗Work begun during an internship at the Allen Institute. †Equal contribution. 1Project code available at https://github.com/ allenai/situation_modeling. b=[1,2,3,4,5],... c=5, b=[1,2,3,4],... (a) Breakpoints in a formal language 4 5 (b) "Breakpoints" for natural language Variable States Program Text 1 2 3 4 The apple is in the park The apple is in the kitchen Mary is at the park John is at the kitchen b=[ ? , ? , ? , T ] Belief Queries/Variables b=[ F , T , ? , T ] b=[ F , T , F , T ] b=[ T , F , T , T ] (c) Use breakpoints for model debugging (joint model) Question: Where is the apple? constraint checking Does my model know: Objects that don't move stay in their location? Feed answer as query to belief model, check consistency Sys. Answer: The apple is in the park Post-hoc Probing Figure 1: Deep narrative understanding in natural lan- guage (bottom) involves the ability to answer queries about arbitrary intermediate points in a given story. We liken this task to breakpoints in programming (top), …

Content truncated to 500 words