Learning Relationship between Authors’ Activity and Sentiments: A case study of online medical forums

Proceedings of Recent Advances in Natural Language Processing, pages 604–610, Hissar, Bulgaria, Sep 7–9 2015. Learning Relationship between Authors’ Activity and Sentiments: A case study of online medical forums Marina Sokolova IBDA and University of Ottawa sokolova@uottawa.ca Victoria Bobicev Technical University of Moldova vika@rol.md Abstract Our current work analyses relations between sentiments and activity of authors of online In- Vitro Fertilization forums. We focus on two types of active authors: those who start new discussions and those who post significantly more messages than other authors. By incor- porating authors‟ activity information into a domain-specific lexical representation of mes- sages, we were able to improve multi-class classification of sentiments by 9% for Support Vector Machines and by 15.3 % for Condi- tional Random Fields. 1 Introduction User-friendly information and communications technologies and easily available access to the Internet were critical in development of Social Web, a socio-technical phenomenon that enables people to connect, support and learn from each other (Ho et al, 2014). The world-wide social media helped to create a digital resource of texts written by the general public. Those texts aggre- gate sentiments expressed by millions of people in relations to consumer goods, political cam- paigns, climate change and other matters of so- cial importance. However, not all participants in online communities contribute equally to that resource: there are visitors who only read the posted texts, authors posting occasional messag- es and a small group of active authors whose on- line contributions significantly overweigh contri- butions of other authors. Those most active par- ticipants significantly influence online discus- sions (Tan et al., 2011; Zafarani et al., 2010). Our current work studies relations between au- thors‟ activity levels and expressed sentiments in an online IVF forum. The forum is a public plat- form for discussion of In-Vitro Fertilization (IVF) treatment. It has been shown that senti- ments on forums dedicated to specific health conditions dependent on the topic of discussion (Ali et al, 2013). We use a set of sentiment and factual categories tailored for on-line IVF discus- sions: encouragement, gratitude, confusion, en- dorsement and facts. We are interested in two types of active au- thors: a) those who start discussions (a.k.a. first authors), b) those who post significantly more messages than other authors (a.k.a. prolific au- thors). The remaining authors usually post one- two messages, and their contributions are rather sporadic. We have found that distribution of sen- timents appeared in text written by different types of authors differs considerably. For exam- ple, the authors who start new topics and actively post in the following discussion usually express more gratitude: 26% of messages posted by the first authors vs. 9% of messages for all the au- thors. We wanted to confirm that information about the author‟s activity has practical implication and can enhance sentiment and subjectivity lexicons. We used automated prediction of sentiments, where messages were first represented through a domain-specific subjectivity lexicon and then authors‟ activity information was added to the representation. This enhancement helped to im- prove …

Content truncated to 500 words