TESA: A Task in Entity Semantic Aggregation for Abstractive Summarization

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 8031–8050, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 8031 TESA: A Task in Entity Semantic Aggregation for Abstractive Summarization Cl´ement Jumel McGill University, Mila clement.jumel@mail.mcgill.ca Annie Louis Google Research, London annielouis@google.com Jackie C. K. Cheung McGill University, Mila jcheung@cs.mcgill.ca Abstract Human-written texts contain frequent general- izations and semantic aggregation of content. In a document, they may refer to a pair of named entities such as ‘London’ and ‘Paris’ with different expressions: “the major cities”, “the capital cities” and “two European cities”. Yet generation, especially, abstractive summa- rization systems have so far focused heavily on paraphrasing and simplifying the source con- tent, to the exclusion of such semantic abstrac- tion capabilities. In this paper, we present a new dataset and task aimed at the seman- tic aggregation of entities. TESA contains a dataset of 5.3K crowd-sourced entity aggrega- tions of PERSON, ORGANIZATION, and LO- CATION named entities.1 The aggregations are document-appropriate, meaning that they are produced by annotators to match the sit- uational context of a given news article from the New York Times. We then build baseline models for generating aggregations given a tu- ple of entities and document context. We fine- tune on TESA an encoder-decoder language model and compare it with simpler classifica- tion methods based on linguistically informed features. Our quantitative and qualitative eval- uations show reasonable performance in mak- ing a choice from a given list of expressions, but free-form expressions are understandably harder to generate and evaluate. 1 Introduction Abstractly speaking, abstraction can be defined as the process of deriving general concepts from spe- cific instances. In automatic summarization, how- ever, “abstractive” summarization often means any type of rewriting of words in some source docu- ment into an output summary. Concretely, recent summarization datasets including XSum (Narayan 1TESA’s code and access to its dataset can be found at: https://github.com/clementjumel/tesa Input Entities Franc¸ois Bayrou, Nicolas Sarkozy, S´egol`ene Royal Document Context Franc¸ois Bayrou, Nicolas Sarkozy, and S´egol`ene Royal are the main contenders in the French presidential elections. Possible Aggregations • the French politicians • the French presidential can- didates • the politicians Table 1: An example of semantic entity aggregation. The input consists of a tuple of named entities, a sit- uational (document) context, and background informa- tion about the entities (not shown here). The expected output is an aggregation of the tuple of entities. et al., 2018) and NEWSROOM (Grusky et al., 2018) quantify the degree of abstractiveness of a summary in terms of its novel N-grams. While such a surface-level definition of abstrac- tiveness is certainly useful and convenient, it is nev- ertheless only a proxy for abstraction in the broader sense which concerns semantic generalization. We argue that it is important to also focus explicitly on semantic abstraction, as this capability is required for more difficult types of summarization which are out of reach of current methods. For example, generating a plot summary of a novel might require describing sequences …

Content truncated to 500 words