Answer searching method and device based on deep question and answer

1. An answer searching method based on deep question and answer, comprising:

receiving a query statement;

obtaining webpages containing candidate answers by searching according to the query statement;

obtaining a first candidate answer, a second candidate answer and a third candidate answer based on a webpage analysis method, a deep learning method and a template matching method respectively;

scoring and ranking the first candidate answer, the second candidate answer and the third candidate answer;

obtaining a candidate answer having a highest score from the first candidate answer, the second candidate answer and the third candidate answer, and judging whether the highest score is larger than a preset score;

when the highest score is larger than the preset score, determining the candidate answer having the largest score to be a final answer; and displaying the final answer;

wherein obtaining the first candidate answer based on the webpage analysis method comprises:

performing a structure analysis and a content analysis on the webpages containing candidate answers, to obtain structure information and content information of sentences in the webpages; and

selecting the first candidate answer from the sentences in the webpages according to the structure information, the content information and the characteristics of the query statement;

wherein obtaining the second candidate answer based on the deep learning method comprises:

obtaining a sentence vector of the query statement, sentence vectors of the candidate answers contained in the webpages and structural knowledge vectors in a structural knowledge base;

inputting the sentence vector of the query statement, sentence vectors of the candidate answers contained in the webpages and structural knowledge vectors in the structural knowledge base to a convolutional neural network, to generate a middle information vector; and

inputting the middle information vector to a recurrent neural network, to generate the second candidate answer;

wherein obtaining the third candidate answer based on the template matching method comprises:

obtaining the candidate answers in the webpages containing candidate answers;

performing a semantic analysis on the candidate answers to extract word characteristics of the candidate answers;

obtaining candidate templates from an answer template base according to the word characteristics;

obtaining a correlation between each of the candidate templates and the query statement, and determining an answer template from the candidate templates according to the correlation; and

generating the third candidate answer according to the answer template.

2. The method according to claim 1 , wherein obtaining webpages containing candidate answers by searching according to the query statement comprises:

performing an analysis processing on the query statement, and generating characteristics of the query statement, where, the analysis processing comprises at least one of word segmentation, part-of-speech tagging, named entity recognition, semantic role labeling, syntax dependency parsing, and semantic dependency analysis;

obtaining the webpages containing candidate answers by searching according to the characteristics.

3. The method according to claim 1 , wherein displaying the final answer comprises:

directly displaying the final answer at a preset position in a search result providing page.

4. An answer searching device based on deep question and answer, comprising:

a processor; and

a memory, configured to store one or more software modules executable by the processor,

wherein the one or more software modules comprises:

a receiving module, configured to receive a query statement;

a search module, configured to obtain webpages containing candidate answers by searching according to the query statement;

an obtaining module, configured to obtain a first candidate answer, a second candidate answer and a third candidate answer based on a webpage analysis method, a deep learning method and a template matching method respectively;

a determining module, configured to determine a final answer according to the first candidate answer, the second candidate answer and the third candidate answer; and

a displaying module, configured to display the final answer;

wherein the determining module comprise: a ranking unit, configured to score and rank the first candidate answer, the second candidate answer and the third candidate answer; a second processing unit, configured to obtain a candidate answer having a highest score, and judging whether the highest score is larger than a preset score; and a determining unit, configured to determine the candidate answer having the largest score to be the final answer when the highest score is larger than the preset score;

wherein the obtaining module is further configured to:

perform a structure analysis and a content analysis on the webpages containing candidate answers, to obtain structure information and content information of sentences in the webpages; and

select the first candidate answer from the sentences in the webpages according to the structure information, the content information and the characteristics of the query statement;

wherein the obtaining module is further configured to:

obtain a sentence vector of the query statement, sentence vectors of the candidate answers contained in the webpages and structural knowledge vectors in a structural knowledge base;

input the sentence vector of the query statement, sentence vectors of the candidate answers contained in the webpages and structural knowledge vectors in the structural knowledge base to a convolutional neural network, to generate a middle information vector; and

input the middle information vector to a recurrent neural network, to generate the second candidate answer;

wherein the obtaining module is further configured to:

obtain the candidate answers in the webpages containing candidate answers;

perform a semantic analysis on the candidate answers to extract word characteristics of the candidate answers;

obtain candidate templates from an answer template base according to the word characteristics;

obtain a correlation between each of the candidate templates and the query statement, and determine an answer template from the candidate templates according to the correlation; and

generate the third candidate answer according to the answer template.

5. The device according to claim 4 , wherein the search module comprises:

a first processing unit, configured to perform an analysis processing on the query statement, and to generate characteristics of the query statement, where, the analysis processing comprises at least one of word segmentation, part-of-speech tagging, named entity recognition, semantic role labeling, syntax dependency parsing, and semantic dependency analysis;

an obtaining unit, configured to obtain the webpages containing candidate answers by searching according to the characteristics.

6. The device according to claim 4 , wherein the displaying module is configured to:

directly display the final answer at a preset position in a search result providing page.

7. A non-transitory computer-readable storage medium, configured to store instructions that, when executed by a processor of a terminal, cause the terminal to perform an answer searching method based on deep question and answer, the method comprising:

receiving a query statement;

obtaining webpages containing candidate answers by searching according to the query statement;

obtaining a first candidate answer, a second candidate answer and a third candidate answer based on a webpage analysis method, a deep learning method and a template matching method respectively;

scoring and ranking the first candidate answer, the second candidate answer and the third candidate answer;

obtaining a candidate answer having a highest score from the first candidate answer, the second candidate answer and the third candidate answer, and judging whether the highest score is larger than a preset score;

when the highest score is larger than the preset score, determining the candidate answer having the largest score to be a final answer; and

displaying the final answer;

wherein obtaining the first candidate answer based on the webpage analysis method comprises:

performing a structure analysis and a content analysis on the webpages containing candidate answers, to obtain structure information and content information of sentences in the webpages; and

selecting the first candidate answer from the sentences in the webpages according to the structure information, the content information and the characteristics of the query statement;

wherein obtaining the second candidate answer based on the deep learning method comprises:

obtaining a sentence vector of the query statement, sentence vectors of the candidate answers contained in the webpages and structural knowledge vectors in a structural knowledge base;

inputting the sentence vector of the query statement, sentence vectors of the candidate answers contained in the webpages and structural knowledge vectors in the structural knowledge base to a convolutional neural network, to generate a middle information vector; and

inputting the middle information vector to a recurrent neural network, to generate the second candidate answer;

wherein obtaining the third candidate answer based on the template matching method comprises:

obtaining the candidate answers in the webpages containing candidate answers;

performing a semantic analysis on the candidate answers to extract word characteristics of the candidate answers;

obtaining candidate templates from an answer template base according to the word characteristics;

obtaining a correlation between each of the candidate templates and the query statement, and determining an answer template from the candidate templates according to the correlation; and

generating the third candidate answer according to the answer template.