March 17 | Erasmus Studio Lunch Lecture: "Detecting Implicit Aspects in Online Reviews for Sentiment Analysis"

With more and more people getting access to the Web, consumer reviews are now a mainstream part of the online shopping experience. As a source of information, these reviews are valuable for consumers as well as producers. Since the number of reviews is beyond manual processing, there is a wide interest in automatic tools for analysing review texts, for example in order to determine on the overall sentiment among reviewers. This so-called sentiment analysis can also be applied on other text types than consumer reviews. 

For the purpose of sentiment analysis the first step is to trace the textual elements in a review expressing the reviewer’s sentiment with respect to the entity that is being reviewed. The next step in sentiment analysis is to go beyond the entity level, and to detect the sentiment at the aspect level by assigning separate sentiment scores to the various aspects of the entity being reviewed (e.g., for a product, this could be the price, the availability, customer service, its aesthetics, etc.). This second step yields a more refined analysis, since many people have a positive opinion regarding one aspect (e.g., price), but a negative opinion about another (e.g., quality). A complication for this second step is that while in most cases aspects are literally mentioned in the text, but sometimes they are implied.

The talk will address the problem of finding these implicit aspects, by using the words that are present in the sentence as pointers to the implicit aspect. When a certain implicit aspect often appears alongside certain words (in annotated training data), these words can be used as cues to detect implicit aspects for unseen data. The method to be presented is tested on reviews for products and restaurants and shows increased performance compared to other methods.