Sentiment Analysis of Social Media Content

People share their opinions and sentiments on various topics via social media. Automatic sentiment analysis of social media content is a requirement for assessing the pulse of any population around any topic. We try to explore a series of problems with respect to sentiment analysis of social media content, and these efforts will lead to deeper and broader understanding of people's sentiments.

• Sentiment expression extraction & construction of sentiment lexicon: Sentiment expression is a single word or a phrase that causes the sentiment on a target (e.g., a named entity) in the text. Identifying sentiment expressions is important in understanding the sentiment of the text and can be more useful than pure polarity results. For example, considering users' opinion on the movie Transformer 3, sentiment expressions would provide more fine-grained information about what people think of the movie, such as "must see", "rate 5 stars", "awesome", than the overall polarity (i.e., positive, negative or neutral) of the text only. Furthermore, extracting sentiment expressions from text collections is an important way to construct sentiment lexicon, which is essential to many sentiment analysis applications, such as sentiment classification, opinion summarization, etc.

• Joint recognition of entity and entity-level sentiment from event-centric data: Many discussions in social media are triggered by the events happening in the real world. People's attitudes, opinions and insights of different events are valuable social signals that are desired to be recognized from these discussions. People usually express their thought about different entities involved in the specific event during such discussions. In order to capture a comprehensive view of people's sentiments of the event, it is necessary to identify their sentiments toward these entities. For example, the Royal Wedding of Prince William and Kate Middleton gave rise to a world-wide discussion. People commented on not only the wedding itself, but also the bride, bridegroom, other celebrities attending the wedding and the wedding dress, etc. We try to leverage the semantic and syntactic relations between different entities involved in an event, sentiments with respect to these entities, and relations between these entities and sentiments to jointly recognize entities and entity-level sentiments from event-centric data.

• Sentiment tracking: This is an effort to understand whether and how people's sentiments changing over time. We are interested in analyzing not only the trend of the changes, but also how these changes happen, or in other words, how people's opinions are affected.

Publications

Lu Chen, Wenbo Wang, Meenakshi Nagarajan, Shaojun Wang and Amit P. Sheth. Extracting Diverse Sentiment Expressions with Target-dependent Polarity from Twitter. In Proceedings of the 6th International AAAI Conference on Weblogs and Social Media (ICWSM), 2012.

Datasets

(1) Domain-specific Twitter Datasets: 168K tweets on movie domain & 259K tweets on person domain.
(2) Labeled Twitter Datasets for Sentiment Analysis: 1500 tweets on movie domain & 1500 tweets on person domain
***** Currently we are trying to get the permission from Twitter to share the datasets on the web. Meanwhile, send us (chen@knoesis.org) an email if you want to obtain a copy.