Lifelong Knowledge-Enriched Social Event Representation Learning

Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics, pages 3624–3635 April 19 - 23, 2021. ©2021 Association for Computational Linguistics 3624 Lifelong Knowledge-Enriched Social Event Representation Learning Prashanth Vijayaraghavan MIT Media Lab Cambridge, MA, USA pralav@mit.edu Deb Roy MIT Media Lab Cambridge, MA, USA dkroy@media.mit.edu Abstract The ability of humans to symbolically rep- resent social events and situations is crucial for various interactions in everyday life. Sev- eral studies in cognitive psychology have es- tablished the role of mental state attributions in effectively representing variable aspects of these social events. In the past, NLP research on learning event representations often focuses on construing syntactic and semantic informa- tion from language. However, they fail to consider the importance of pragmatic aspects and the need to consistently update new so- cial situational information without forgetting the accumulated experiences. In this work, we propose a representation learning frame- work to directly address these shortcomings by integrating social commonsense knowledge with recent advancements in the space of life- long language learning. First, we investigate methods to incorporate pragmatic aspects into our social event embeddings by leveraging so- cial commonsense knowledge. Next, we in- troduce continual learning strategies that allow for incremental consolidation of new knowl- edge while retaining and promoting efficient usage of prior knowledge. Experimental re- sults on event similarity, reasoning, and para- phrase detection tasks prove the efficacy of our social event embeddings. 1 Introduction Everyday life comprises the ways in which peo- ple typically act, think, and feel on a daily basis. Our life experiences unfold naturally into tempo- rally extended daily events. The event descriptions can be packaged in various ways depending on several factors like speaker’s perspective or the re- lated domain. Interpretation of event descriptions will be incomplete without understanding multi- ple entities involved in the events and even more so when the focus is primarily on “social events”, S1: Student goes to class S2: Student goes to wedding S3: Student takes course S4: Teacher takes course S5: Professor teaches subject x S1 x S2 x S3 x S4 x S5 x S1 x S2 x S3 x S4 x S5 ConceptNet ConceptNet ATOMIC ConceptNet ATOMIC SB-SCK x S1 x S2 x S3 x S4 x S5 x S1 x S2 x S3 x S4 x S5 (a) (b) (c) (d) Figure 1: Illustration of functioning of our lifelong rep- resentation learning approach that produces incremen- tally richer social event representations. Event texts are given in the top green box. With more knowledge, so- cial event embeddings move beyond high lexical over- lap [shown in (a)] and learn to integrate semantic and pragmatic properties [shown in (b), (c)] of event texts along with social role information [shown in (d)]. i.e., events explaining social situations and inter- actions. Therefore, a social event representation model must capture the semantic properties from the event text description and embed salient knowl- edge that encompasses the implicit pragmatic abil- ities. Early definitions of pragmatic aspects refer …

Content truncated to 500 words