Personality-dependent Neural Text Summarization

Proceedings of Recent Advances in Natural Language Processing, pages 205–212, Varna, Bulgaria, Sep 2–4, 2019. https://doi.org/10.26615/978-954-452-056-4_024 205 Personality-dependent Neural Text Summarization Pablo Botton da Costa University of S˜ao Paulo S˜ao Paulo, Brazil pablo.botton.costa@gmail.com Ivandr´e Paraboni University of S˜ao Paulo S˜ao Paulo, Brazil ivandre@usp.br Abstract In Natural Language Generation (NLG) systems, personalization strategies - i.e., the use of information about a target au- thor to generate text that (more) closely re- sembles human-produced language - have long been applied to improve results. The present work addresses one such strategy - namely, the use of Big Five person- ality information about the target author - applied to the case of abstractive text summarization using neural sequence-to- sequence models. Initial results suggest that having access to personality informa- tion does lead to more accurate (or human- like) text summaries, and paves the way for more robust systems of this kind. 1 Introduction Computational approaches to text summarization may be divided into two general categories: ab- stractive and extractive summarization. Extrac- tive summarization consists of selecting relevant pieces of text to compose a subset of the original sentences, whereas the more complex abstractive summarization involves interpreting the input text and rewriting its main ideas in a new, shorter ver- sion. Both strategies may be modelled as a ma- chine learning problem by making use of unsuper- vised (Ren et al., 2017), graph-based and neural methods (Wan and Yang, 2006; Cao et al., 2015), among others. The present work focuses on the is- sue of neural abstractive summarization, address- ing the issue of personalized text generation in sys- tems of this kind. Text-generating systems may in principle pro- duce always the same fixed output from a given in- put representation. In order to generate more nat- ural (or ‘human-like’) output, however, systems of this kind will often implement a range of stylis- tic variation strategies. Among these, the use of computational models of human personality has emerged as a popular alternative, and it is com- monly associated with the rise of the Big Five model of human personality (Goldberg, 1990) in many related fields. The Big Five model is based on the assump- tion that differences in personality are reflected in natural language use, and comprises five fun- damental dimensions of personality: Extraver- sion, Agreeableness, Conscientiousness, Neuroti- cism, and Openness to experience. Given its lin- guistic motivation, the Big Five personality traits have been addressed in a wide range of studies in both natural language understanding and gen- eration alike. Thus, for instance, the work in Mairesse and Walker (2007) introduces PERSON- AGE, a fully-functional NLG system that pro- duces restaurant recommendations. PERSON- AGE and many of its subsequent extensions sup- port multiple stylistic variations that are controlled by personality information provided as an input. The use of personality information for text sum- marization, by contrast, seems to be far less com- mon, and we are not aware of any existing work that addresses the issue of personality-dependent neural text summarization. Based on these ob- …

Content truncated to 500 words