We are concerned with the task of Aspect Based Sentiment Analysis (ABSA), where the goal is to identify the aspects of given target entities and the sentiment expressed towards each aspect in restaurant reviews. Also, we identify the categories of the aspect term and the category polarity for each review. Our approach is rule-based and we have come up with various heuristics to solve this non-trivial problem. When evaluated using SemEval-2014 dataset, we obtain decent results for aspect based sentiment analysis.

Present Day Scenarios...

In recent years, the explosion of social networking sites, blogs and review sites provide a lot of information. Mil- lions of people express uninhibited opinions about various product features and their nuances. This forms an active feedback which is of importance not only to the companies developing the products, but also to their rivals and several other potential customers. Sentiment Analysis is the task of tapping this goldmine of information.

Workarounds...

The sentiment regarding a particular product in a review is seldom explicitly positive or negative; rather people tend to have a mixed opinion about various features, some positive and some negative. Thus the feature specific opinion matters more than the overall opinion.

Consider a review: The food was good but servicing was a bit slow. This sentence has a mixed emotion. The emotion regarding food is
positive whereas that regarding servicing life is negative.
Hence, it is of utmost importance to extract only those opinions relevant to a particular feature (like food or service) and classify them, instead of taking the complete sentence and the overall sentiment. We present a rule-based approach that exploits common-sense knowledge and sentence dependency trees to detect both aspects, their categories, similarity and polarity score.

Brief Tour of the Project

** Aspect Term Extraction **

** Aspect Category Detection and Polarity Results **

Conclusion

In this project, we developed a system that extracts potential features from a review and clusters opinion expressions describing each of the features. Eventually, it retrieves the opinion expression describing the user specified feature. The drawback of the system is that it cannot evaluate domain dependent implicit sentiment as it does not train on any domain specific data. Thus the system will not be able to distinguish between: The story is unpredictable(positive sentiment) and The steering wheel is unpredictable(negative sentiment).
This is due to the usage of a generic sentiment lexicon, in the final stage, in rule based classification. Supervised classification can distinguish between the two sentiments but it needs tagged data and separate training for every domain. This system can be further improved upon by:

• Discovering more rules for aspect term extraction.
• Combining existing rules for complex aspect extraction.
• To extract aspect categories dictionaries created can be made more noise free.
• Hybrid approach or supervised classification can be used for better performance.