US 10,762,438 B1Granted

Extracting questions and answers

AssigneeFacebook, Inc.

InventorsYing Zhang, Irina-Elena Veliche, Benoit F. Dumoulin, Aram Grigoryan, Wenhai Yang

Life of this patent

Estimating the term…

Loading patent…

Abstract

A system for answering user questions can provide answers from a knowledge base that stores question/answer pairs. These pairs can be associated with characteristics of the asking user so that, when subsequent users ask similar questions, answers can be selected that have been identified as most relevant to that type of user. The question/answer pairs in the knowledge base can be identified from social media posts where the original post contains a question and one or more comments on the post provide an answer. Posts can be identified as containing a question using a question classification model. A post comment can be identified as an answer based on: whether the question poster responded positively to the comment, whether the comment has similar keywords to the question, whether the comment has the characteristics of an answer, and how often a similar answer has been provided for similar questions.

Claims19

Dependent claims

A method for identifying a question and answer pair, comprising:

determining, by a computer processor, that a social media post provided by a first user includes a question, wherein the social media post is associated with comments provided by at least one second user other than the first user;
identifying a second question similar to the question based on a similarity between a question vector for the question and a question vector for the second question;
selecting, from the comments associated with the social media post, an answer to the question, wherein selecting the answer is based on a combined score computed for the answer, wherein the combined score is based on:
an answer-type score computed for the answer;
a question similarity score computed for the answer;
a comment score computed for the answer;
and
a frequency score computed for the answer based on a similarity between an answer vector for the answer and an answer vector for a stored answer to the second question;
determining a context associated with the social media post or author characteristics associated with the first user;
determining a type of the question by a question classifier, wherein the type comprises one of a recommendation type or a factual type;
storing a version of the question and the selected answer in a knowledge base, wherein the stored version is associated with the determined context or the determined author characteristics and the determined type of the question;
and
generating abstracted versions of the question and the selected answer by mapping the question and the selected answer into a vector space based on a root node of a dependency structure for the question and a root node of a dependency structure for the selected answer, wherein the dependency structures comprise a respective plurality of nodes, wherein generating the abstracted versions of the question and selected answer further comprises:
computing, for each node of the plurality of nodes, a respective vector in the vector space;
selecting, as the vector in the vector space for the selected answer, the vector computed for the root node of the plurality of nodes for the dependency structure for the selected answer;
and
selecting, as the vector in the vector space for the question, the vector computed for the root node of the plurality of nodes for the dependency structure for the question;
receiving, by the processor, a new question;
and
returning a question and answer pair from the knowledge base as a response to the new question.

17 more elements

A non-transitory computer-readable storage medium storing instructions that, when executed by a computing system, cause the computing system to perform operations for selecting an answer to a question, the operations comprising:

identifying a second question similar to a question in a social media post based on a similarity between a question vector for the question and a question vector for the second question;
selecting, from comments to the social media post, an answer to the question, wherein selecting the answer is based on a combined score computed for the answer based on a combination of:
a question similarity score computed for the answer;
a frequency score computed for the answer based on a similarity between an answer vector for the answer and an answer vector for a stored answer to the second question;
and
a comment score computed for the answer indicating how an author of the question responded to the answer and how multiple other users have responded to the answer;
determining, by a question classifier, whether a type of the question comprises a recommendation type or a factual type;
storing the selected answer and the determined type of the question in a knowledge base with an association to the question;
generating abstracted versions of the question and the selected answer by mapping the question and the selected answer into a vector space based on a root node of a dependency structure for the question and a root node of a dependency structure for the selected answer, wherein the dependency structures comprise a respective plurality of nodes, wherein generating the abstracted versions of the question and selected answer further comprises:
computing, for each node of the plurality of nodes, a respective vector in the vector space;
selecting, as the vector in the vector space for the selected answer, the vector computed for the root node of the plurality of nodes for the dependency structure for the selected answer;
and
selecting, as the vector in the vector space for the question, the vector computed for the root node of the plurality of nodes for the dependency structure for the question;
receiving a new question;
and
returning a question and answer pair from the knowledge base as a response to the new question.

13 more elements

A automated question response system, comprising:

a memory;
one or more processors;
a knowledge base comprising multiple question and answer pairs, wherein at least some of the question and answer pairs were selected by:
using a question classifier trained to identify questions in social media posts associated with comments;
identifying a second question similar to a question based on a similarity between a question vector for the question and a question vector for the second question;
selecting, from the comments associated with the corresponding social media post, an answer to the question, wherein selecting the answer is based on a combined score computed for the answer, wherein the combined score is based on:
a comment score computed for the answer, a frequency score computed for the answer, and a similarity score computed for the answer based on a similarity between an answer vector for the answer and an answer vector for a stored answer to the second question;
determining a type of the question by the question classifier, wherein the type comprises one of a recommendation type or a factual type;
storing a version of the question, the selected answer, a context of the corresponding social media post, and the determined type of the question in the knowledge base;
and
generating abstracted versions of the question and the selected answer by mapping the question and the selected answer into a vector space based on a root node of a dependency structure for the question and a root node of a dependency structure for the selected answer, wherein the dependency structures comprise a respective plurality of nodes, wherein generating the abstracted versions of the question and selected answer further comprises:
computing, for each node of the plurality of nodes, a respective vector in the vector space;
selecting, as the vector in the vector space for the selected answer, the vector computed for the root node of the plurality of nodes for the dependency structure for the selected answer;
and
selecting, as the vector in the vector space for the question, the vector computed for the root node of the plurality of nodes for the dependency structure for the question;
and
an interface configured to receive an identification of a new question;
wherein the automated question response system is configured to respond to the new question by:
identifying a question and answer pair from the knowledge base, wherein the question of the identified question and answer pair is within a threshold level of similarity to the new question;
and
providing, through the interface, the answer of the identified question and answer pair as a response to the new question.

18 more elements

Dependencies are inferred from the claim text.