Questions disambiguation using generative evidence fusion and round-trip prediction
Techniques for predicting an answer to a question using a machine learning model are described. In some examples, the model predicts one or more answers to the question by: predicting at least two answers to the question using a first component of the question-answer model from a set of passages, generating, using a second component of the question-answer model, at least one question for each of the predicted at least two answers, and performing roundtrip predictions until each generated question only has one answer.
1. A computer-implemented method comprising:
receiving a request to predict an answer to a question using a question-answer model;
predicting one or more answers to the question by:
predicting at least two answers to the question from a set of ranked passages using a first encoder/decoder model of the question-answer model, wherein an input to the first encoder/decoder model comprises a concatenation of the question and each passage, and wherein the first encoder/decoder model processes each passage of the set of ranked passages independently in an encoder of the first encoder/decoder model and jointly in a decoder of the encoder/decoder model,
generating, using a second encoder/decoder model of the question-answer model, at least one question for each of the predicted at least two answers,
performing roundtrip question-answer pair prediction, wherein performing roundtrip question-answer pair prediction comprises:
feeding at least a proper subset of the generated questions back to the first encoder/decoder model of the question answer model, wherein the at least a proper subset of the generated questions comprises less than all or all of the generated questions,
predicting at least one answer to each question of the at least proper subset of the generated questions using the first encoder/decoder model of the question-answer model from the set of ranked passages, and
generating, when there are multiple predicted answers to any question of the at least proper subset of generated questions, using the second encoder/decoder model of the question-answer model, at least one question for each of the predicted multiple answers, wherein each generated question that has only one predicted answer and the only one predicted answer form a question-answer pair to potentially be presented as a predicted result, and
continuing performing roundtrip question-answer pair prediction until each generated question fed back to the first encoder/decoder model has only one predicted answer; and
providing at least a proper subset of the question-answer pairs.
2. The computer-implemented method of claim 1 , wherein the request includes at least the question and one or more of an indication of a data source to use to answer the question, a maximum number of answers to receive, or an indication of how the answer is to be presented.
3. A computer-implemented method comprising:
receiving a request to predict an answer to a question using a question-answer model;
predicting one or more answers to the question by:
predicting at least two answers to the question from a set of passages using a first component of the question-answer model, wherein an input to the first component comprises a concatenation of the question and each passage of the set of passages, and wherein the first component processes each passage of the set of passages independently in an encoder of the first component and jointly in a decoder of the first component,
generating, using a second component of the question-answer model, at least one question for each of the predicted at least two answers,
performing roundtrip question-answer pair prediction, wherein performing roundtrip question-answer pair prediction comprises:
feeding at least a proper subset of the generated questions back to the first component of the question answer model,
predicting at least one answer to each question of the at least proper subset of the generated questions using the first component of the question-answer model from the set of passages, and
generating, when there are multiple predicted answers to any question of the at least proper subset of the generated questions, using the second component of the question-answer model, at least one question for each of the multiple predicted answers, wherein each generated question that has only one predicted answer and the only one predicted answer form a question-answer pair to potentially be presented as a predicted result, and
continuing the performing roundtrip question-answer pair prediction until each generated question fed back to the first component of the question answer model has only one predicted answer; and
providing at least a proper subset of the question-answer pairs.
4. The computer-implemented method of claim 3 , wherein the set of passages is a subset of passages retrieved from storage that have been ranked.
5. The computer-implemented method of claim 4 , wherein the ranking is performed by an encoder/decoder based model.
6. The computer-implemented method of claim 3 , further comprising:
verifying and filtering the generated questions prior to feeding the at least a proper subset of the generated questions to the first component of the question answer model.
7. The computer-implemented method of claim 6 , wherein the verifying and filtering is performed using language model verification.
8. The computer-implemented method of claim 6 , wherein the verifying and filtering is performed using exact match verification.
9. The computer-implemented method of claim 3 , wherein an input to the first component of the question-answer model comprises one of a combination of a question and each passage or the question and a list of retrieved passages.
10. The computer-implemented method of claim 3 , wherein an input to the second component of the question-answer model comprises a concatenation of a question, each passage, and a predicted answer.
11. The computer-implemented method of claim 3 , wherein the second component utilizes a bidirectional encoder that feeds an autoregressive decoder.
12. The computer-implemented method of claim 3 , wherein providing at least a proper subset of the question-answer pairs comprises highlighting each answer in a relevant passage.
13. The computer-implemented method of claim 3 , wherein the request includes at least the question and one or more of an indication of a data source to use to answer the question, a maximum number of answers to receive, or an indication of how the answer is to be presented.
14. A system comprising:
a first one or more electronic devices to implement a storage service in a multi-tenant provider network to store passages from documents; and
a second one or more electronic devices to implement a search service in the multi-tenant provider network, the search service including instructions that upon execution cause the search service to:
receive a request to predict an answer to a question using a question-answer model;
predict one or more answers to the question by:
predicting at least two answers to the question from a set of stored passages using a first component of the question-answer model, wherein an input to the first component comprises a concatenation of the question and each passage of the set of stored passages, and wherein the first component processes each passage of the set of stored passages independently in an encoder of the first component and jointly in a decoder of the first component,
generating, using a second component of the question-answer model, at least one question for each of the predicted at least two answers,
performing roundtrip question-answer pair prediction, wherein performing roundtrip question-answer pair prediction comprises:
feeding at least a proper subset of the generated questions back to the first component of the question answer model,
predicting at least one answer to each question of the at least proper subset of the generated questions using the first component of the question-answer model from the set of passages, and
generating, when there are multiple predicted answers to a question of the at least proper subset of the generated questions, using the second component of the question-answer model, at least one question for each of the multiple predicted answers, wherein each generated question that has only one predicted answer and the only one predicted answer form a question-answer pair to potentially be presented as a predicted result, and
continuing the performing roundtrip question-answer pair prediction until each generated question fed back to the first component of the question answer model has only one predicted answer; and
providing at least a proper subset of the question-answer pairs.
15. The system of claim 14 , wherein the set of passages is a subset of passages retrieved from storage that have been ranked.
16. The system of claim 15 , wherein the ranking is performed by an encoder/decoder based model.
17. The system of claim 14 , wherein the search service is further to verify and filter the generated questions prior to feeding the at least a proper subset of the generated questions to the first component of the question answer model.
18. The system of claim 14 , wherein an input to the first component of the question-answer model comprises one of a combination of a question and each passage or the question and a list of retrieved passages.
19. The system of claim 14 , wherein an input to the second component of the question-answer model comprises a concatenation of a question, each passage, and a predicted answer.