Online users always seek experiences. For example, travelers might look for a hotel with clean rooms and a lively bar, while dating couples might search for a restaurant with a romantic ambiance.
Close to 70% of experiential attributes are subjective. Moreover, these preferences are based on personal beliefs or feelings, not facts. Everyone is different.
However, e-commerce search engines often support queries with objective attributes, such as location and prices. If users want to learn about the subjective qualities of an experience, their best chance is to read text reviews, sometimes hundreds of them.
Systems should support online users in their search for subjective experiences, such as “quiet room,” “vibrant ambiance,” or “close to transit.” To this end, we built OpineDB, a subjective database engine that supports answering experiential queries.
Here’s how it works:
First, OpineDB extracts key experiences from a large set of customer review texts. The extraction process uses BERT and opinion mining techniques. Then, OpineDB maps those experiences to a domain-specific schema to convert them into histogram chart summaries. These summaries are stored in the database. Next, when a user searches for an experience, OpineDB maps the search phrase to the domain-specific schema. This could be a challenging task. For example, a user might search for “LA hotels with clean rooms that are romantic getaways.” “Clean rooms” clearly maps to “cleanliness.” However, “romantic getaways” maps to something ambiguous, because “romantic” is not in the hotel domain schema.
To solve this problem, OpineDB uses a method based on word co-occurrence. In this case, users frequently mention the “staff” and “style” attributes in the reviews for romantic occasions, like honeymoons and anniversaries. Therefore, these two attributes are useful mappings for “romantic getaway.”
Finally, OpineDB outputs to the users not only the search results ranked by a combination of “cleanliness,” “staff,” and “style,” but also the snippets of the reviews elaborating the searched terms.
Don’t these search results from OpineDB answer the users’ experiential needs more effectively? Using real data in the hotel and restaurant domain, OpineDB shows an 8% improvement in ranking quality compared to the traditional IR-based search.