Entity-Level Sentiment: More than the Sum of Its Parts

  • 2024-09-17 10:07:01
  • Egil Rønningstad, Roman Klinger, Lilja Øvrelid, Erik Velldal
  • 0

Abstract

In sentiment analysis of longer texts, there may be a variety of topicsdiscussed, of entities mentioned, and of sentiments expressed regarding eachentity. We find a lack of studies exploring how such texts express theirsentiment towards each entity of interest, and how these sentiments can bemodelled. In order to better understand how sentiment regarding persons andorganizations (each entity in our scope) is expressed in longer texts, we havecollected a dataset of expert annotations where the overall sentiment regardingeach entity is identified, together with the sentence-level sentiment for theseentities separately. We show that the reader's perceived sentiment regarding anentity often differs from an arithmetic aggregation of sentiments at thesentence level. Only 70\% of the positive and 55\% of the negative entitiesreceive a correct overall sentiment label when we aggregate the(human-annotated) sentiment labels for the sentences where the entity ismentioned. Our dataset reveals the complexity of entity-specific sentiment inlonger texts, and allows for more precise modelling and evaluation of suchsentiment expressions.

 

Quick Read (beta)

loading the full paper ...