A Sentiment Analysis of Men’s and Women’s Speech in the BNC64

A Sentiment Analysis of Men’s and Women’s Speech in the BNC64 Yong-hun Lee Ji-Hye Kim Chungnam National University Hannam University 99 Daehak-ro, Yuseong-gu 70 Hannamro, Daedeok-gu Daejeon 34134, Korea Daejeon 34430, Korea yleeuiuc@hanmail.net jihye_0612@naver.com Abstract This study applied sentiment analysis to the corpus data and analyzes the men’s and women’s speech. For the analysis, the BNC64 corpus was used and three different types of analyses were employed: dictionary- based analysis, GRU-based analysis, and BERT-based analysis. When the data were analyzed with the dictionary-based analysis, there was no significant difference in the use of sentiment words between men and women. When the data were analyzed with the GRU-based and BERT-based analysis, it was observed that even though men and women used a similar proportion of sentiment words, women used more positive words. The tendency became much clearer in the BERT-based analysis. This study implied that the claims in previous studies were supported by the authentic corpus data and that it showed how the gender differences became clearer as the method developed from the dictionary- based method to the BERT-based analysis. 1 Introduction It is known that men’s language use is different from women’s, and there have been a lot of studies on the gender differences in language use (Holmes and Meyrehoff, 2003; Baker, 2014). These studies investigated both similarities and differences between men and women. As deep learning technology develops (Ian et al., 2016), there have been a few trials to apply the technology to examine the gender differences in language use, and one of the frequently-used methods is sentiment analysis. Since the advent of the Bidirectional Encoder Representations from Transformers (BERT; Devlin et al., 2019), it is possible to use a pre-trained language model to conduct sentiment analysis, because it belongs to one of the sub-tasks that can be analyzed with the BERT. This study employed one of the traditional methods in sentiment analysis (i.e., dictionary-based methods) and two deep-learning methods (GRU and BERT). The BNC64 corpus was used in the analysis, which was constructed by extracting the men’s and women’s speech from the British National Corpus (BNC). The data were analyzed with three different types of sentiment analyses, and their analysis results were compared. This study has importance in that the claims in previous studies are supported by the authentic corpus data and that it demonstrates how the gender differences became clearer as the analysis method develops from the dictionary-based method to the BERT-based analysis. This study is also important in that it is one of a few studies which investigate gender differences with the BERT. 2 Previous Studies 2.1 Language and Gender The study on language and gender started from the early academic approach which was called ‘gender difference paradigm’. Lakoff (1975) introduced a ‘male dominance’ theory of language use, which claimed that males used language to dominate females. Fisherman (1977) developed the theory further and proposed that women engaged in the so-called ‘interactional shitwork’, which involved using questions and hedges to force responses from men in order …

Content truncated to 500 words