Graph Based Sentiment Aggregation using ConceptNet Ontology

Proceedings of the The 8th International Joint Conference on Natural Language Processing, pages 525–535, Taipei, Taiwan, November 27 – December 1, 2017 c⃝2017 AFNLP Graph Based Sentiment Aggregation using ConceptNet Ontology Srikanth G Tamilselvam and Seema Nagar and Abhijit Mishra and Kuntal Dey IBM Research, Bangalore and New Delhi, India {srikanth.tamilselvam, senagar3, abhijimi, kuntadey}@in.ibm.com Abstract The sentiment aggregation problem ac- counts for analyzing the sentiment of a user towards various aspects/features of a product, and meaningfully assimilating the pragmatic significance of these fea- tures/aspects from an opinionated text. The current paper addresses the senti- ment aggregation problem, by assigning weights to each aspect appearing in the user-generated content, that are propor- tionate to the strategic importance of the aspect in the pragmatic domain. The nov- elty of this paper is in computing the prag- matic significance (weight) of each aspect, using graph centrality measures (applied on domain specific ontology-graphs ex- tracted from ConceptNet), and deeply in- graining these weights while aggregating the sentiments from opinionated text. We experiment over multiple real-life product review data. Our system consistently out- performs the state of the art - by as much as a F-score of 20.39% in one case. 1 Introduction User-generated content accounts for a large frac- tion of the online content that is available today. The advent of web-platforms such as on online so- cial networks (e.g., Facebook, Twitter etc.), blogs, discussion forums and product portals (e.g., Ama- zon) has resulted in creation of a plethora of user- generated opinionated-content. The task of senti- ment analysis focuses on analyzing such text, and deciphering user sentiment towards given products or their features, often referred to as aspects. The task of sentiment aggregation builds upon sentiment analysis processes in general. While sentiment analysis (SA) aims to classify an opinion into positive or negative or neutral categories (in case of coarse-grained SA) or into more intricate categories (in case of fine- grained/dimensional SA), often enough, it con- siders the opinion in its entirety, and is agnos- tic of the aspect-specific sentiments expressed. However, in practical settings, opinions expressed by users (such as product reviews) often tend to focus on multiple aspects, not just one. The diversity of the aspects, does not allow the aspect-specific sentiment-polarity values to be just naively summed up for the purpose of obtaining an overall aggregated sentiment of a given user to- wards a given product. As an example, let us consider the following text regarding a software application: This app has a beautiful interface. It is not bug free though. The first sentence here is a positive feel to it from the interface feature (aspect). The second sentence has a negative feel to it from the accuracy (bug) feature (aspect). Thus, a simple linear aggregation (sum of the individual polarities) of the features, will yield a overall neutral review polarity score. This is not necessarily accurate. While an inter- face is a necessary enabler for users to use a soft- ware, it is of imperative necessity that the software …

Content truncated to 500 words