Entity-Level Sentiment Analysis (ELSA): An Exploratory Task Survey

Proceedings of the 29th International Conference on Computational Linguistics, pages 6773–6783 October 12–17, 2022. 6773 Entity-Level Sentiment Analysis (ELSA): An exploratory task survey Egil Rønningstad and Erik Velldal and Lilja Øvrelid University of Oslo Department of Informatics {egilron,erikve,liljao}@ifi.uio.no Abstract This paper explores the task of identifying the overall sentiment expressed towards voli- tional entities (persons and organizations) in a document – what we refer to as Entity-Level Sentiment Analysis (ELSA). While identify- ing sentiment conveyed towards an entity is well researched for shorter texts like tweets, we find little to no research on this specific task for longer texts with multiple mentions and opinions towards the same entity. This lack of research would be understandable if ELSA can be derived from existing tasks and models. To assess this, we annotate a set of pro- fessional reviews for their overall sentiment to- wards each volitional entity in the text. We sam- ple from data already annotated for document- level, sentence-level, and target-level sentiment in a multi-domain review corpus, and our re- sults indicate that there is no single proxy task that provides this overall sentiment we seek for the entities at a satisfactory level of per- formance. We present a suite of experiments aiming to assess the contribution towards ELSA provided by document-, sentence-, and target- level sentiment analysis, and provide a discus- sion of their shortcomings. We show that senti- ment in our dataset is expressed not only with an entity mention as target, but also towards targets with a sentiment-relevant relation to a volitional entity. In our data, these relations extend beyond anaphoric coreference resolu- tion, and our findings call for further research of the topic. Finally, we also present a survey of previous relevant work. 1 Introduction Over the course of the last two decades, the field of NLP has generated a vast body of research on sentiment analysis (SA), i.e. the task of identifying opinions expressed in text. Prior work has focused on a range of different levels of analysis; from document- or sentence-level polarity classification to more fine-grained prediction of various compo- nents of opinions, like source/holder expressions, polar expressions, target expressions and aspect- based sentiment classification. However, we observe that a more aggregated level of analysis, what we here dub "entity-level sentiment analysis" (or ELSA for short), remains under-explored. For our purposes, we will de- fine ELSA as the task of determining the overall (i.e. document-level) polarity (positive/negative) expressed towards an entity in a text. Moreover, for the current paper we will restrict the discus- sion to volitional entities, like persons (PER) or organizations (ORG). A given text might make reference to multi- ple distinct entities, each of which might be men- tioned multiple times, both directly and indirectly, and also have multiple opinions directed towards them. Hence, solving the task of ELSA in its full complexity may potentially involve several different sub-tasks, including (but not necessar- ily limited to) named entity recognition (NER), resolution of entity mentions, coreference- and anaphora-resolution, identification of sentiment tar- …

Content truncated to 500 words