Don't Cry for me, Google: Semantic Analysis of Travel Reviews to Understand Feelings

 

Boris Galitsky

Conventional search engines deliver documents which include significant occurrences of keywords in user queries. Additionally, Google is good at selecting those documents which has satisfied users in previous similar searches. While this works well enough for general searches, "vertical" searches for specific kinds of products can do much better.

To recommend a product, it is very convincing to refer to the experience of those who used it before, and to provide argumentation for this product based on feeling of these users. To do that, a search engine must not only "understand" the features of products such as "a hotel close to outdooring activities", but also feeling of people about these products like "not impressed with a view but nice for guys' getaway".

Obviously, Google search engine cannot provide recommendation by finding documents which include keywords "hotel+close+outdoor[ing]">+activities+not+impres[sed]...". Neither can many other Travel sites, which possesses necessary data for vertical search but uses keyword match for search. To handle recommendation queries, a search engine must know that hotels are characterized with locations, sometimes good locations are those which are close to activities, in particular, outdoor activities. Furthermore, search engine must know that 'views' are important considerations while staying in hotels, expression "not impressed" refers to a negative feeling, which is nevertheless combined with positive reference to the category such as "guys getaway". It is also necessary to understand sarcastic expressions like "I would not let my dog stay here"='not clean' versus "They would not let my dog stay here"='dogs are not allowed'. Notice how similar these sentences are, and how different are the meanings.

To enable such feeling-based recommendation, we built a knowledge base and reasoning engine which operates with entities of the travel domain.  We also constructed a formal model of human sentiments and feelings to be extracted from text and to serve as a basis for providing recommendations.