Tweet Review Mining focusing on Celebrities by Machine Reading Comprehension based on BERT

Tweet Review Mining focusing on Celebrities by Machine Reading Comprehension based on BERT Yuta Nozaki, Kotoe Sugawara, Yuki Zenimoto, Takehito Utsuro Degree Programs in Systems and Information Engineering, Graduate School of Science and Technology, University of Tsukuba, 1-1-1, Tennodai, Tsukuba, Ibaraki, 305-8573, Japan {s2020776,s2120736,s2220753} @ s.tsukuba.ac.jp, utsuro @ .iit.tsukuba.ac.jp Abstract In this paper, we propose a method of min- ing tweets that represent reviews on celebrities by the machine reading comprehension model based on BERT. The purpose of this paper is to collect and aggregate reviews on tweets for celebrities in order to support the activi- ties of fans of celebrities searching for infor- mation on celebrities’ criticisms and impres- sion trends on the Web. Specifically, we fo- cus on the celebrities’ names and adjective ex- pressions that appear in tweets, and determine whether or not there is a relationship such as review and impression between the two words using the BERT machine reading comprehen- sion framework. In the machine reading com- prehension framework, the tweet is the con- text, the celebrity’s name is the question, and an adjective that has a relationship such as review and impression with the celebrity is the answer. As the result of the evaluation experiment, the machine reading comprehen- sion framework achieved fairly reliable per- formance. 1 Introduction The purpose of this paper is to support the process by which TV drama viewers and celebrity fans search for information on critiques and interest trends in celebrities on the Web. In this paper, we propose a method for mining reviews of popular celebrities using tweets as the information source. Usually, when an event or an incident that is related to a pop- ular celebrity occurs, a large number of tweets are posted, where, in those tweets, people express their own thoughts in the way they like. In Twitter, how- ever, there exist no rule on the grammatical correct- ness of the posted tweets. Thus, this makes it un- expectedly difficult to correctly identify what people actually intend to express in their tweets, mainly due to the lack of grammatical correctness of tweet sen- tences (Sanguinetti et al., 2020). And, in the NLP community, in the case of other applications such as tweet sentiment analysis, it is quite common to avoid grammatically parsing tweet sentences, but to directly analyze sentiment of tweets (e.g, Nakov et al. (2016)). Considering such a background, in this paper, we apply the BERT framework of machine reading comprehension (Figure 1) to tweet sentences, so as to detect the span of impression relations from text showing the review relations of celebrities. Specifi- cally, first we collect tweets in which the name of the celebrity is included. We develop a dataset of tweets that are annotated with the span of adjectives indi- cating review relationships for the celebrity names co-occurring in the tweets. Next, in a machine read- ing comprehension framework, we train the BERT model (Devlin et al., 2019) to predict the spans that indicate reviews on previously unobserved celebri- ties. Machine …

Content truncated to 500 words