Opinion Learning from Medical Forums

Proceedings of Recent Advances in Natural Language Processing, pages 18–24, Hissar, Bulgaria, 7-13 September 2013. Opinion Learning from Medical Forums Tanveer Ali University of Ottawa tali028@uottawa.ca Marina Sokolova University of Ottawa & CHEO sokolova@uottawa.ca David Schramm University of Ottawa & CHEO dschramm@ottawahospital.on.ca Diana Inkpen University of Ottawa Diana.Inkpen@uottawa.ca Abstract Our study focuses on opinion mining of several medi- cal forums dedicated to Hearing Loss (HL). Surgeries related to HL are the most common surgeries in North America; thus, they affect many patients and their families. We have extracted the opinions of people from these forums related to stigma of HL, conse- quences of HL surgeries, living with HL, failures of HL loss treatments, etc. We performed a manual an- notation first with two annotators and have 93% over- all agreement with kappa 0.78 and then applied Ma- chine Learning methods to classify the data into opin- ionated and non-opinionated messages. Using our feature set, we achieved best F-score 0.577 and 0.585 with SVM and logistic-R classifier respectively. 1 Introduction The development of the Internet and of the user- friendly Web technologies profoundly changed the ways the general public can express their opinions on a multitude of topics. In order to make informed decisions, there is a necessity to develop methods that adequately – efficiently and effectively – extract new knowledge from the online messages (Bobicev et al., 2012). Opin- ions depend on individual's personality, culture and expectations of the society. Thus, opinions are challenging for independent external evalua- tion and categorization. Natural language statements can be divided into two categories: facts and opinions. Facts can be expressed with topic keywords, while opin- ions are more difficult to express with a few keywords. They are the words of mouth on the web, e.g., Factual Sentence: Most things come in somewhere between 40 and 105, depending on the frequency. Opinionated Sentence: I don't think you will find anyone who this level of amplification is undamaging, but the option is to not hear. In this work, we have performed opinion mining of message posted on medical forums dedicated to Hearing Loss. Surgeries related to HL are the most common surgeries in North America; thus, they affect many patients and their families. Our current work aims to provide a tool that can extract opinions expressed by the general public. Understanding of what people think about the surgeries and their consequences helps health care providers to develop better health care poli- cies and the general public outreach. We collected data from web forums and we invited two annotators to manually annotate texts gathered from medical forums. We obtained the overall agreement of 93% and kappa was 0.78. Then we used a subjectivity lexicon and machine learning algorithms to automatically classify the posts. Our experiments with different combina- tions of features using different classifiers, i.e., Naïve Bayes, SVM and Logistics-R have shown significant improvement in F-score performance (55.7%, 56.8% and 57.8%, respectively) over the majority class baseline, which was 47.6%. 2 Related Work A very limited …

Content truncated to 500 words