Processing Narratives by Means of Action Languages

In this work we design a narrative understanding system Text2ALM that can be used in Question Answering domains. System Text2ALM utilizes an action language ℒℳ to perform inferences on complex interactions of events described in narratives. The methodology that Text2ALM follows in its implementation was originally outlined by Yuliya Lierler, Daniela Inclezan, and Michael Gelfond in 2017 via a manual process, and this work serves as a proof of concept in a large-scale environment. Our system automates the conversion of a narrative to an ℒℳ model containing facts about the narrative. We make use of the VerbNet lexicon that we annotated with interpretable semantics in ℒℳ. Text2ALM also utilizes Text2DRS system developed by Gang Ling at UNO in 2018. These resources are used to produce an ℒℳ program with a system description containing information on the narrative’s entities, events, and their relations, as well as a history of the narrative’s events. The ℒℳ logic is used in tandem with a basic commonsense library of ℒℳ modules to generate a formal structure capturing the narrative’s properties. The CALM system designed by researchers at Texas Tech in 2018 and is used by Text2ALM to process the ℒℳ program. The effectiveness of this approach is measured by the system’s ability to correctly answer questions from the QA bAbI tasks published by Facebook Research in 2015. The Text2ALM system matched or exceeded the performance of state-of-the-art machine learning methods in six of the seven tested tasks