Coalescing Global and Local Information for Procedural Text Understanding

Proceedings of the 29th International Conference on Computational Linguistics, pages 1534–1545 October 12–17, 2022. 1534 Coalescing Global and Local Information for Procedural Text Understanding Kaixin Ma†, Filip Ilievski§, Jonathan Francis†¶, Eric Nyberg†, Alessandro Oltramari¶ †Language Technologies Institute, Carnegie Mellon University §Information Sciences Institute, University of Southern California ¶Human-Machine Collaboration, Bosch Research Pittsburgh {kaixinm, jmf1, ehn}@cs.cmu.edu, ilievski@isi.edu, alessandro.oltramari@us.bosch.com Abstract Procedural text understanding is a challenging language reasoning task that requires models to track entity states across the development of a narrative. A complete procedural under- standing solution should combine three core aspects: local and global views of the inputs, and global view of outputs. Prior methods con- sidered a subset of these aspects, resulting in either low precision or low recall. In this pa- per, we propose Coalescing Global and Local Information (CGLI), a new model that builds entity- and timestep-aware input representa- tions (local input) considering the whole con- text (global input), and we jointly model the en- tity states with a structured prediction objective (global output). Thus, CGLI simultaneously optimizes for both precision and recall. We extend CGLI with additional output layers and integrate it into a story reasoning framework. Extensive experiments on a popular procedu- ral text understanding dataset show that our model achieves state-of-the-art results; experi- ments on a story reasoning benchmark show the positive impact of our model on downstream reasoning. We release our code here: https: //github.com/Mayer123/CGLI 1 Introduction Understanding the causal links of events in proce- dures is a key aspect of intelligence, facilitating human narration and dialogue. For instance, under- standing why story B is plausible and why story A is not (Figure 1) requires procedural understanding of the causes of John leaving his notebook at home, as opposed to him taking out his notebook from his bag: writing in a notebook is counterfactual in the former case, and intuitive in the latter. Un- derstanding stories requires procedural models that can reason consistently about event implications, and do so at different granularities. For a model to decide whether a story is plausible, it has to track the entity states over time, understand the effects of Figure 1: An example story of understanding task. Given two stories, the task is to judge which story is more plausible, find the conflicting sentence pair in the implausible story, and predict entity states at every step. the described actions (green arrows), and consider the preconditions for a given action (pink arrows). Meanwhile, the model must reconcile the causes and effects of all events described in the story, to provide a globally consistent interpretation. While procedural reasoning research reports steady progress in recent years (Rajaby Faghihi and Kordjamshidi, 2021; Gupta and Durrett, 2019a; Zhang et al., 2021), story understanding and pro- cedural reasoning have rarely been considered to- gether (Storks et al., 2021). Works have attended only to complementary aspects of the procedural reasoning problem, e.g., Gupta and Durrett (2019a) build entity-centric context representations and ig- noring timestep-wise representation modeling; and Rajaby Faghihi and Kordjamshidi (2021) later pro- …

Content truncated to 500 words