Understanding Tables on the Web

Overview | Understanding Process | Application Snapshot | Experiment

Query: American politicians birthday

The figure demonstrates our understanding of the tables. After correctly parsed "American politicians" as a concept and "birthday" as an an attribute, the system is able to find tables which are actually talking about politicians, not "states" or "party" as in the second result. Besides, our system can understand "birthdate" and "date of birth" are synonyms of "birthday", then return ranked valid statements as result.

Query: tech companies revenue

Besides showing our understanding of tables as in the above example, this figure exhibits another side of why searching table is valuable. For attributes like "revenue" which is related to other information (like year), searching and conserving the original table structure can help user to better understand the result. In the first table, the year can be found in Wikipedia's "title" and "section". In the second table, the table contains a column of year itself. That's something hard to achieve if we only mining from web text. Plus, tables can provide more information than the user's intension, for example, the "profits", "headquarters", and "revenue growth %", which is another advantage.

Query: tech companies business type

The figure shows our understanding on synonyms and locating entities of the queried concept.

 Query: airlines airport