Fact verification engine
AssigneeMicrosoft Corporation
InventorsSilviu-Petru Cucerzan, Chee Wee Leong
Life of this patent
Estimating the term…
Abstract
The described implementations relate to processing of electronic data. One implementation is manifested as a technique that can include receiving an input statement that includes a plurality of terms. The technique can also include providing, in response to the input statement, ranked supporting documents that support the input statement or ranked contradicting results that contradict the input statement.
Claims20
A method performed by at least one computing device, the method comprising:
- training a classifier to classify terms of input statements using training data, the training data including training statements having corresponding training references, wherein the training references identify supporting documents that support the training statements or contradictory documents that contradict the training statements;
- receiving a first input statement, the first input statement comprising a plurality of first terms;
- rewriting the first input statement to obtain one or more queries by classifying the first terms of the first input statement using the trained classifier;
- and
- using the one or more queries to obtain search results for the first input statement.
A system comprising:
- a verification engine;
- and
- one or more processing devices configured to execute the verification engine,
- wherein the verification engine is configured to receive an input statement and output a query,
- the verification engine comprising:
- training data comprising training statements and training references, the training references identifying supporting documents for the training statements,
- a classification component configured to learn, from the training statements and the supporting documents for the training statements, a function that assigns values to terms of the input statement, and
- a rewriting component configured to rewrite the input statement based on the values assigned to the terms to obtain the query.
One or more computer-readable storage media comprising instructions which, when executed by one or more processing devices, cause the one or more processing devices to perform acts comprising:
- obtaining a plurality of annotated documents;
- extracting training data from the plurality of annotated documents, the training data comprising training statements and training references that identify supporting documents or contradictory documents for the training statements;
- training a classifier to perform term evaluation using the training data;
- receiving an input statement comprising a plurality of terms;
- rewriting the input statement into at least one query comprising individual terms from the input statement or related terms that are related to the individual terms of the input statement by evaluating the plurality of terms of the input statement using the trained classifier;
- querying a search engine with the at least one query to obtain a plurality of search results identifying candidate supporting documents for the input statement;
- re-ranking the plurality of search results based on one or more ranking features;
- and
- providing the re-ranked search results in response to the input statement.
Dependencies are inferred from the claim text.