Corpus Building and Evaluation of Aspect-based Opinion Summaries from Tweets in Spanish

Corpus Building and Evaluation of Aspect-based Opinion Summaries from Tweets in Spanish Daniel Pe˜naloza, Rodrigo L´opez, Juanjos´e Tenorio, H´ector G´omez, Arturo Oncevay-Marcos, Marco A. Sobrevilla Cabezudo Research Group on Artificial Intelligence (IA-PUCP) Departamento de Ingenier´ıa, Pontificia Universidad Cat´olica del Per´u, Lima, Peru {a20110645,a20112387,juanjose.tenorio,hector.gomez,arturo.oncevay,msobrevilla}@pucp.edu.pe Abstract This project involves the presentation and analysis of a corpus of Spanish extractive and abstractive summaries of opinions. The purpose of this work is to display a corpus of diverse summaries that could be used as a reference for academic research as we have not found one for the Spanish language as far as we know. We have analyzed the summaries based on the agreement between them as this shows how different they are written between each other and on aspect coverage and sentiment orientation as this proves the difference between the content that each summary tries to express. After the experimentation, we have found that even if each annotator uses a different expression to summarize a text, all of them contain similar messages. Furthermore, when writing, all of them prioritize on common aspects that are more representative of the corpus. Keywords: aspect-based opinion summarization, corpora building, Spanish 1. Introduction Automatic summarization is one of the most challenging problems in Natural Language Processing (NLP). The task is even more difficult if the summarized content is gener- ated entirely by the program, and it does not present any extract of the original text. Theres are two kinds of sum- maries: abstractive and extractive (Labb´e and Portet, 2012). In extractive summarization, we try to select the principal ideas or sentences and grouping them in the summary. In abstractive summarization, we try to write a text which con- tains all ideas. The relevance of the summarization task increases if the context of the high variety and volume of information is considered. Social networks, such as micro-blogging sites, nowadays are offering the possibility to exploit any kind of publications, where user opinions are greatly valuable. Opinion mining (OM) is the NLP subfield advocated to this problem. The main goal is to extract a subjective value from the text, that might include a certain polarity degree, and it could be focused on a particular entity or aspect (Liu, 2012). Therefore, it is likely to be a special difficulty in summa- rizing user opinions from social media, as there is not an specific order among all the opinion posts grouped by the same topic. Besides, each post may address a different as- pect or entity related to the main topic, so there is a need to identify them first. After that, the subjectivity of each aspect must be measured, such as a positive or negative po- larity (Wu et al., 2016). There is also a quantitative scope by considering the distribution of different positions regard- ing an opinion, as no one should be treated as more relevant than others (Liu, 2012). Aspect-based opinion summarization is an attractive task in this context. However, the primary research is central- ized in some specific languages …

Content truncated to 500 words