CN103235833B - Answer search method and device by the aid of statistical machine translation - Google Patents

Abstract

The invention discloses an answer retrieval method and device by means of statistical machine translation. Firstly, the candidate answers are translated into several other languages by using statistical machine translation tools, and several equivalent representations of the candidate answers are obtained. Then, the method of matrix decomposition is used to reduce the dimensionality of several equivalent representations of candidate answers to obtain low-dimensional implicit representations. Second, the query problem is transformed into a low-dimensional implicit representation using statistical machine translation and matrix factorization methods. Finally, the similarity between the query question and the candidate answer is calculated on the hidden space, and several candidate answers with the highest similarity are returned as the answer to the query question. The method proposed by the invention can effectively solve the problems of vocabulary mismatch and vocabulary ambiguity. Experiments prove that in a large-scale community question answering dataset, the performance of answer retrieval is improved by 29.36%.

Description

A method and device for answer retrieval by means of statistical machine translation

technical field

The invention relates to the technical field of natural language processing, and relates to an answer retrieval method and device by means of statistical machine translation.

Background technique

With the rapid development of Internet technologies, Internet services based on User-Generated Content (UGC) are becoming more and more popular. Community Q&A is a new information exchange and knowledge sharing system based on "question-answer" that emerged under this background, such as Yahoo! Answers, Baidu Zhizhi, etc. Different from the automatic question answering system, in the community question answering system, users can ask any type of questions, and can also answer any type of questions from other users. Answer retrieval is the basis of community Q&A analysis and occupies a very important position. The task of answer retrieval refers to retrieving semantically similar or similar answers to the query from a large-scale candidate answer library, and the user answers the query. Therefore, answer retrieval has important theoretical significance and practical value.

The main challenges in answer retrieval currently are the vocabulary mismatch between the query question and the candidate answers and the problem of lexical ambiguity. Vocabulary mismatch usually causes the answer retrieval model to retrieve many answers that do not match the user's query intent. The main reason is that the query questions and answers in the community Q&A are given by the user, and the user's query intent is highly diverse. For example, the word "interest" can mean either "curiosity" or "a charge for borrowing money", depending on the user. "Word ambiguity" is a common phenomenon between query questions and candidate answers. It is specifically manifested in the fact that many words do not appear many times in query questions and candidate answers, or even never appear in query questions or candidate answers. Use traditional term-based matching methods.

One way to solve the above-mentioned "lexical ambiguity" and "lexical gap" problems is to use statistical machine translation to represent ambiguous words in the original language and words with different literal meanings with their corresponding translations. The premise of the statistical machine translation method is to first establish a reasonable objective function to integrate the original language and its corresponding translation in a framework, then how to minimize the noise caused by statistical machine translation, and finally how to design a A fast solution method to solve the above objective function. However, directly adding the translated vocabulary to the original language will greatly reduce the accuracy of answer retrieval. The main reason is that directly adding the translated vocabulary to the original language will greatly increase the computational complexity, and at the same time, errors in machine translation will also bring problems. There is a lot of noise.

The task of answer retrieval is to retrieve the answer that can answer the query from the answer document collection for the query question input by the user. The main difficulty in answer retrieval is that user query questions and candidate answers use different word forms when expressing the same or similar meanings, which may easily lead to lexical mismatch and lexical ambiguity. Traditional methods mainly rely on mining word associations between monolinguals, ignoring the semantic associations between multilingual information.

Contents of the invention

In order to solve the above problems, the present invention first needs to design a reasonable objective function to effectively integrate the original language and its corresponding translation into a framework, and at the same time constrain the impact of machine translation noise on answer retrieval under this framework. Then, according to the established objective function and its constraints, a fast solution method is designed. By solving the objective function, the hidden representation of the original language and its corresponding translation is obtained, and finally the similarity between the user query and the candidate answer is calculated on the hidden space. According to the above ideas, the present invention mainly focuses on the two major difficulties in answer retrieval, and successfully introduces statistical machine translation into the process of answer retrieval. It is proved by experiments that this method effectively improves the accuracy of answer retrieval.

The basic idea of the present invention is to make full use of statistical machine translation to represent ambiguous words and words with different literal expressions in the original language with their corresponding translations, thereby improving the performance of answer retrieval.

The invention discloses

A method of answer retrieval by means of statistical machine translation, comprising the steps of:

Step 1. Translate all candidate answers expressed in the original language into other languages by means of statistical machine translation tools;

Step 2, integrating the candidate answers represented by each language including the original language into a framework based on non-negative matrix factorization;

Step 3, using the least squares fast gradient descent algorithm to solve the framework based on non-negative matrix factorization, and obtain the low-dimensional expressions of each language representation of all candidate answers;

Step 4. Translate the query expressed in the original language into other multiple languages by means of statistical machine translation tools;

Step 5, using the low-dimensional representations of each language representation of all candidate answers obtained in step 3, transforming the query question and other multilingual translations into a low-dimensional space;

Step 6. According to the query question and other multilingual translations, and the low-dimensional expression of candidate answers corresponding to the query question and other multilingual translations, calculate the query question and other multilingual translations and their corresponding candidates The similarity between the answers, and get the final retrieval result according to the similarity.

The invention also discloses an answer retrieval device by means of statistical machine translation, which includes:

A candidate answer translation module for translating the candidate answers into other languages;

a matrix factorization module that integrates candidate answers expressed in each language, including the original language, into a framework based on non-negative matrix factorization;

Optimizing the solution module, using the least squares fast gradient descent algorithm to solve the framework based on non-negative matrix factorization, and obtain the low-dimensional expression of each language representation of all candidate answers for each question;

Query question translation module for translating query questions into other languages;

A similarity calculation module based on a low-dimensional space, which is used to transform the query question into a low-dimensional space, and calculate the similarity between the query question and the candidate answer in the low-dimensional space;

The result ranking learning module is used to finally obtain the search answer according to the similarity calculated by the similarity calculation module.

The invention adopts the idea of statistical machine translation to improve the performance of answer retrieval. Using the statistical machine translation tool Google Translate, the ambiguous words and words with different literal expressions in the original language are represented by their corresponding translations, thereby improving the performance of answer retrieval.

Description of drawings

Fig. 1 is the answer retrieval method by means of statistical machine translation in the present invention.

Fig. 2 is a structural diagram of an answer retrieval device by means of statistical machine translation in the present invention.

detailed description

In order to make the object, technical solution and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with specific embodiments and with reference to the accompanying drawings.

The invention discloses an answer retrieval method and device by means of statistical machine translation. It can be divided into two parts: offline process and online process. The offline process is realized by three modules, namely the candidate answer translation module, the matrix decomposition module, and the optimization solution module. The online process is also divided into three modules, namely the query question translation module, the similarity calculation module based on low-dimensional space, and the result ranking learning module.

Fig. 1 shows an answer retrieval method by means of statistical machine translation proposed by the present invention. As shown in Figure 1, it includes two stages: offline part and online part. The offline process includes:

Step (1), use statistical machine translation tools to translate all candidate answers expressed in the original language l ₁ (such as English), and obtain equivalent representations in L-1 different languages {l ₁ , l ₂ , ..., l _{L -1} }, wherein L represents the number of all languages, and the statistical machine translation tool can be selected from Google Translate or the like.

Step (2), the set of candidate answers expressed in each language is represented as a word-document matrix of M _p ×N Among them, M _p represents all vocabulary in the candidate answer set represented by the p-th language, and N represents the number of answers in the candidate answer set.

Step (3), design a new objective function, use the method of non-negative matrix factorization to integrate the candidate answers expressed in P different languages into a unified framework, and use a regularization strategy to reduce the impact of statistical machine translation. noise.

Step (4), design a least squares-based fast gradient descent algorithm, and obtain low-dimensional representations of L different languages by solving the above objective function, that is, the coefficient matrix and the reconstruction matrix

The online process includes:

Step (1), using a statistical machine translation tool to translate the query expressed in the original language l ₁ (such as English) into equivalent representations in L-1 different languages. The statistical machine translation tool can be Google Translate or the like.

Step (2), using the coefficient matrix obtained in the above-mentioned offline process (4) Transform the query problem and its corresponding L-1 translation representations into a low-dimensional space. ;

Step (3), calculating the similarity between the query question and the candidate answer on the low-dimensional space representation.

Step (4), adopting the strategy of linear ranking learning, merging the similarities represented by L different languages in low-dimensional space, and returning several candidate answers with the highest scores as the final answers.

Fig. 2 shows the answer retrieval device by means of statistical machine translation proposed in the present invention. As shown in Figure 2, the retrieval device includes: a candidate answer translation module, a matrix decomposition module, an optimization solution module, a query translation module and a low-dimensional space-based similarity calculation module.

The candidate answer translation module is used to translate all the candidate answers expressed in the original language l ₁ (such as English) in the offline stage to obtain equivalent representations of L-1 different languages {l ₁ , l ₂ , ..., l _L-1 }, where L represents the number of all languages, that is, the candidate answer sets D ₂ , ..., D _L expressed in other L-1 languages are obtained by translating the candidate answer set D ₁ .

Candidate answer translation is one of the techniques of the present invention. In order to translate candidate answers from one language to another L-1 language, it is time-consuming and labor-intensive to use human translation, especially for the real task of community Q&A answer retrieval, it is obviously unrealistic to translate large-scale candidate answers. Fortunately, the current level of machine translation has been well developed in natural language processing, although the translation quality is not very satisfactory. At present, there are many public free translation tools that provide daily translation services. Adopt Google Translate in the preferred embodiment of the present invention, this translation tool utilizes statistical machine learning method to train translation model on the large-scale parallel corpus of construction, in the process that is translated into another language from one language, can consider rich context information, and has shown good translation performance in many translation tools. After translating the candidate answer set D ₁ , the candidate answer sets D ₂ , . . . , DL expressed in other _L -1 languages can be obtained.

The matrix decomposition module is used to express the set of candidate answers expressed in each language into an M _p ×N word-document matrix in the offline stage Among them, M _p represents all vocabulary in the candidate answer set represented by the p-th language, and N represents the number of answers in the candidate answer combination.

The matrix decomposition module is one of the key technologies of the present invention. Definition {l ₁ , l ₂ , ..., l _L } denotes the set of languages used in the present invention, where L denotes the number of languages, l ₁ denotes the original language (e.g., English), l ₂ ...l _p denotes another L-1 languages. definition Represents the set of candidate answers based on _l1 language expression. define candidate answers Can be expressed as a M _p -dimensional vector where the vector Each element in corresponds to a word, which represents the importance of the word in the i-th candidate answer; the vector It can be calculated with tf-idf, which is a statistical method to evaluate the importance of a word for a file set or a part of a data set. D _p can be represented as a word-document matrix of M _p ×N dimensions In this matrix, each row represents a different word, and each column represents a candidate answer, where M _p represents the number of non-repeated words in D _p , and N represents the number of candidate answers in D _p .

Intuitively speaking, the words in the candidate answer set D ₂ ,...,D _L in another L-1 languages obtained after translation can be directly added to the original candidate answer set D ₁ , which will result in D ₁ corresponding matrix The dimensionality of increases from M ₁ ×N to However, this approach has two disadvantages: (1) it causes data sparsity; (2) translation errors in statistical machine translation will cause noise problems. In order to solve the above problems, the present invention adopts the method of matrix decomposition.

hypothesis matrix can be decomposed into two low-dimensional matrices and Also consider the matrix Independent of the The following objective function can be obtained:

$\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; u</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; ,</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; V</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\underset{{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; u</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; \&Greater\; Equal;</span>\mathrm{<span\; class="notranslate">\; 0</span>}<span\; class="notranslate">\; ,</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; V</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; \&Greater\; Equal;</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\mathrm{<span\; class="notranslate">\; min</span>}}{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; D.</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; -</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; u</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}}{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; V</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; f</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}$

Among them, ||·|| _F represents the norm of the matrix, where Represents the coefficient matrix obtained after decomposition, Represents the reconstruction matrix obtained after decomposition, and K represents the dimensionality of the hidden space.

In order to reduce the noise problem caused by statistical machine translation errors, the present invention assumes from the matrix The reconstruction matrix obtained by (p∈[2,L]) should be the same as from the matrix The obtained reconstruction matrix The closer the better. Therefore, the present invention proposes to minimize the reconstruction matrix (p∈[2,L]) and reconstruction matrix Previous distance:

${\mathrm{<span\; class="notranslate">\; f</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; (</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; V</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\underset{{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; V</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; \&Greater\; Equal;</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\mathrm{<span\; class="notranslate">\; min</span>}}\underset{\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 2</span>}}{\overset{\mathrm{<span\; class="notranslate">\; L</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; V</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; -</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; V</span>}}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; f</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}$

Combining the above two objective functions, the following objective function can be obtained:

${\mathrm{<span\; class="notranslate">\; f</span>}}^{<span\; class="notranslate">\; \&prime;</span><span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; (</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; u</span>}}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; ,</span><span\; class="notranslate">\; \&CenterDot;</span><span\; class="notranslate">\; \&CenterDot;</span><span\; class="notranslate">\; \&Center\; Dot;</span><span\; class="notranslate">\; ,</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; u</span>}}}_{\mathrm{<span\; class="notranslate">\; L</span>}}<span\; class="notranslate">\; ;</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; V</span>}}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; ,</span><span\; class="notranslate">\; \&Center\; Dot;</span><span\; class="notranslate">\; \&CenterDot;</span><span\; class="notranslate">\; \&Center\; Dot;</span><span\; class="notranslate">\; ,</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; V</span>}}}_{\mathrm{<span\; class="notranslate">\; L</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; L</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; D.</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; -</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; u</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}}{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; V</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; f</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span>\underset{\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 2</span>}}{\overset{\mathrm{<span\; class="notranslate">\; L</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; \&lambda;</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; V</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; -</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; V</span>}}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; f</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}$

Among them, the parameter λ _p (p∈[2, L]) is used to adjust the relative weight of the two parts. If a small value is set for the parameter λ _p , the above objective function Similar to the traditional non-negative matrix (Non-negative Matrix Factorization), if a larger value is set for the parameter λ _p , the above objective function Greater emphasis on errors introduced by statistical machine translation.

The optimization solving module is used to solve the parameters in the above-mentioned matrix decomposition module, that is, the coefficient matrix and the reconstruction matrix Through this optimization solution module, the coefficient matrix is obtained and the reconstruction matrix The local optimal representation of is the input result of the offline part.

The optimization solution module is one of the core technologies of the present invention. The above objective function Considering both data sparsity and statistical machine translation errors, there are 2L pairs of optimization objects in the objective function, when considering and When , it is difficult to find an algorithm to solve the above minimization problem. The present invention proposes a fast gradient descent algorithm based on the least square method, which is used to find a local optimal solution, and when optimizing a certain target object, keep other 2L-1 objects unchanged.

Keep and Invariant, for the coefficient matrix The iterative update of the above objective function can be Transformed into the following optimization problem:

$\underset{{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; u</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; \&Greater\; Equal;</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\mathrm{<span\; class="notranslate">\; min</span>}}{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; D.</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; -</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; u</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}}{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; V</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; f</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}$

definition Represents a column vector, representing a matrix All elements of row i; Represents a column vector, representing the coefficient matrix All elements in row i. Therefore, the above optimization problem can be decomposed into M _p independent sub-optimization problems, and each sub-optimization problem corresponds to the coefficient matrix A line of:

$\underset{{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; u</span>}}}_{\mathrm{<span\; class="notranslate">\; i</span>}}^{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; \&Greater\; Equal;</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\mathrm{<span\; class="notranslate">\; min</span>}}{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; d</span>}}}_{\mathrm{<span\; class="notranslate">\; i</span>}}^{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; -</span>{{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; V</span>}}}^{\mathrm{<span\; class="notranslate">\; T</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}}{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; u</span>}}}_{\mathrm{<span\; class="notranslate">\; i</span>}}^{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}$

Subscript i=1,..., M _p , where M _p represents the number of non-repeated words in D _p .

The above sub-optimization problem is a standard least squares problem, and its numerical solution is:

${\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; u</span>}}}_{\mathrm{<span\; class="notranslate">\; i</span>}}^{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; =</span>{<span\; class="notranslate">\; (</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; V</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}}{{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; V</span>}}}^{\mathrm{<span\; class="notranslate">\; T</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; )</span>}^{<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; V</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}}{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; d</span>}}}_{\mathrm{<span\; class="notranslate">\; i</span>}}^{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; )</span>}$

Keep the coefficient matrix and the reconstruction matrix Invariant to the reconstruction matrix The iterative update of the above objective function can be Transform into the following two types of optimization problems:

When p ∈ [2, L], can be transformed into the following objective function:

$\underset{{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; V</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; \&Greater\; Equal;</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\mathrm{<span\; class="notranslate">\; min</span>}}{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; D.</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; -</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; u</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}}{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; V</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; f</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; \&lambda;</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; V</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; -</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; V</span>}}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; f</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}$

When p=1, can be transformed into the following objective function:

$\underset{{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; V</span>}}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; \&Greater\; Equal;</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\mathrm{<span\; class="notranslate">\; min</span>}}{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; D.</span>}}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; -</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; u</span>}}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; V</span>}}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; f</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; \&lambda;</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; V</span>}}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; f</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}$

For the objective function in the first case above, define is the matrix The jth column vector in , represents the reconstruction matrix The jth column vector in . Therefore, the objective function in the first case above can be decomposed into N mutually independent sub-optimization problems, and each sub-optimization problem corresponds to the reconstruction matrix A column of:

$\underset{{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; v</span>}}}_{\mathrm{<span\; class="notranslate">\; j</span>}}^{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; \&Greater\; Equal;</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\mathrm{<span\; class="notranslate">\; min</span>}}{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; d</span>}}}_{\mathrm{<span\; class="notranslate">\; j</span>}}^{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; -</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; u</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}}{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; v</span>}}}_{\mathrm{<span\; class="notranslate">\; j</span>}}^{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; \&lambda;</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; v</span>}}}_{\mathrm{<span\; class="notranslate">\; j</span>}}^{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; -</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; v</span>}}}_{\mathrm{<span\; class="notranslate">\; j</span>}}^{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}$

Wherein the subscript j=1,...,N, N represents the number of candidate answers in the set _Dp .

The above sub-optimization problem is a standard least squares problem based _on L2 regularization, then its numerical solution is:

${\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; v</span>}}}_{\mathrm{<span\; class="notranslate">\; j</span>}}^{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; =</span>{<span\; class="notranslate">\; (</span>{{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; u</span>}}}^{\mathrm{<span\; class="notranslate">\; T</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}}{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; u</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; \&lambda;</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; I</span>}}<span\; class="notranslate">\; )</span>}^{<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; (</span>{{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; u</span>}}}^{\mathrm{<span\; class="notranslate">\; T</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}}{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; d</span>}}}_{\mathrm{<span\; class="notranslate">\; j</span>}}^{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; \&lambda;</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; v</span>}}}_{\mathrm{<span\; class="notranslate">\; j</span>}}^{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; )</span>$

where p ∈ [2, L] represents the translated p-th language, represents the identity matrix.

Similarly, the objective function of the second case above can be solved by a similar method, and its numerical solution is:

${\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; v</span>}}}_{\mathrm{<span\; class="notranslate">\; j</span>}}^{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; =</span>{<span\; class="notranslate">\; (</span>{{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; u</span>}}}^{\mathrm{<span\; class="notranslate">\; T</span>}}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; u</span>}}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; \&lambda;</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; I</span>}}<span\; class="notranslate">\; )</span>}^{<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; u</span>}}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; d</span>}}}_{\mathrm{<span\; class="notranslate">\; j</span>}}^{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>}$

The query question translation module is used to translate the query question into L-1 equivalent representations in different languages using a statistical machine translation tool during the online phase. The statistical machine translation tool can be Google Translate or the like.

Similar to the candidate answer translation module, in order to translate the query from one language to another L-1 languages, the present invention uses a statistical machine translation tool—Google Translate. For a given query question q, other query questions q ₂ ,...,q _L expressed in other L-1 languages are obtained after translation.

The low-dimensional space-based similarity calculation module is used to calculate the similarity between the query question and the candidate answer on the representation of the low-dimensional space.

The similarity calculation module based on low-dimensional space is one of the key technologies of the present invention. For a given query question q and its corresponding translations q ₂ ,...,q _L in L-1 languages, it needs to be transformed into a low-dimensional space. For ease of expression, the query question q expressed in the original language is replaced by the symbol q ₁ , that is, q=q ₁ . Therefore, the following formula can be used to transform q ₁ into a low-dimensional space:

${\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; v</span>}}}_{{\mathrm{<span\; class="notranslate">\; q</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; arg</span>}\underset{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; v</span>}}<span\; class="notranslate">\; \&Greater\; Equal;</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\mathrm{<span\; class="notranslate">\; min</span>}}{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; q</span>}}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; -</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; u</span>}}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; v</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; \&lambda;</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; v</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}$

in, is the vector representation of the query question q ₁ , is the vector representation of query question q ₁ in low-dimensional space, that is, the reconstruction matrix; where Represents the coefficient matrix corresponding to the original language obtained by the optimization solution module. However, for the candidate answer d ₁ , the conversion result obtained after low-dimensional conversion by the matrix factorization module can be directly used, namely The similarity between query question q ₁ and candidate answer d ₁ in low-dimensional space can be expressed by cosine similarity:

$\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; q</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; d</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{<span\; class="notranslate">\; <</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; v</span>}}}_{{\mathrm{<span\; class="notranslate">\; q</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}}<span\; class="notranslate">\; ,</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; v</span>}}}_{{\mathrm{<span\; class="notranslate">\; d</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}}<span\; class="notranslate">\; ></span>}{{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; v</span>}}}_{{\mathrm{<span\; class="notranslate">\; q</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; \&CenterDot;</span>{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; v</span>}}}_{{\mathrm{<span\; class="notranslate">\; d</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; 2</span>}}}$

Among them, s(q ₁ , d ₁ ) represents the similarity between the query question q ₁ and the candidate answer d ₁ in low-dimensional space.

For the translation q _i (i∈[2, L]) corresponding to q ₁ , the following formula can be used to express it in a low-dimensional space:

${\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; v</span>}}}_{{\mathrm{<span\; class="notranslate">\; q</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; arg</span>}\underset{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; v</span>}}<span\; class="notranslate">\; \&Greater\; Equal;</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\mathrm{<span\; class="notranslate">\; min</span>}}{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; q</span>}}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; -</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; u</span>}}}_{\mathrm{<span\; class="notranslate">\; i</span>}}\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; v</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; \&lambda;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; v</span>}}<span\; class="notranslate">\; -</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; v</span>}}}_{{\mathrm{<span\; class="notranslate">\; q</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}$

in, is the vector representation of the query question q _i . Similarly, for the translation d _i (i∈[2, L]) corresponding to the candidate answer d ₁ , the result obtained after the low-dimensional space transformation by the matrix factorization module can be directly used The similarity between the translation q _i corresponding to the query question q ₁ and the translation d _i corresponding to the candidate answer d ₁ in the low-dimensional space can be calculated using the above-mentioned similar cosine similarity calculation method.

The result sorting learning module is used to fuse the similarities represented by L different languages in the low-dimensional space, and return several candidate answers with the highest scores as the final answers. For a given query question q ₁ and candidate answer d ₁ , the present invention designs a ranking learning function as follows:

$\mathrm{<span\; class="notranslate">\; Score</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; q</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; d</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; \&theta;</span>}}<span\; class="notranslate">\; \&Center\; Dot;</span>\mathrm{<span\; class="notranslate">\; \&Phi;</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; q</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; d</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; )</span>$

Among them, Score(q ₁ , d ₁ ) represents the final score of query question q ₁ and candidate answer d ₁ , Represents the weight of the feature vector, Φ(q ₁ , d ₁ )={s(q ₁ , d ₁ ), s(q ₂ , d ₂ ),…, s(q _L , d _L )} represents the feature vector, corresponding to The similarity between the query question q ₁ and the candidate answer d ₁ expressed in L different languages in low-dimensional space. Among them, the parameter The best value is obtained using the cross-validation strategy most commonly used in statistical machine learning. Finally, sort according to the high and low of Score(q ₁ , d ₁ ), and return several candidate answers with the highest scores as the final answers.

In order to illustrate the performance of the device, the present invention uses experiments to verify the improvement of the performance of the answer retrieval system by means of statistical machine translation methods.

The experimental data of the present invention comes from Yahoo! The Answers community question answering system, in these historical question sets, each question is mainly composed of four parts: the title of the question, the category of the question, the description of the question, and the answer to the question. The dataset we adopt contains 1232 user category labels and 2,288,607 question-answer pairs. In order to evaluate the effectiveness of the inventive method, we additionally selected 252 query questions as a test data set. For each query question in the test dataset, we use the language model to retrieve the best 20 results, and then let two annotators manually annotate them. If the returned candidate answer is similar to the query question, it is marked as "relevant", otherwise it is marked as "irrelevant". If two annotators have conflicting annotation structures, let a third person make the final decision. In the process of judging whether the candidate answer is similar to the query question, the annotator only knows the question itself.

In the present invention, setting parameter L=5 means that English needs to be translated into other 4 languages (Chinese, French, Italian, German).

Assuming that Q _t represents a test problem set, the present invention adopts the following two evaluation indicators:

Average Accuracy Rate (MAP): Its calculation formula is as follows:

$\mathrm{<span\; class="notranslate">\; MAP</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; Q</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{{\mathrm{<span\; class="notranslate">\; Q</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}}\underset{\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; \&Element;</span>{\mathrm{<span\; class="notranslate">\; Q</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; q</span>}}}\underset{\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; q</span>}}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; Precision</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; )</span>$

Among them, m _q is the number of questions related to the query question q, R _k is the collection of the kth question and all previous questions in the retrieval results, and Precision(R _k ) is the proportion of questions related to R _k and q. This indicator reflects the overall average of test results.

Precision@n(P@n): Defined as the accuracy rate of the first n results returned by the system for query questions. The Precision@n of the entire test set is the average of the Precision@n of all questions in the test set, and its calculation formula is as follows:

$\mathrm{<span\; class="notranslate">\; P</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; @</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; k</span>}}{\mathrm{<span\; class="notranslate">\; no</span>}}$

Among them, k represents the number of related questions among the first k questions returned by the retrieval system, and n represents the total number of questions returned by the retrieval system. therefore,

$\mathrm{<span\; class="notranslate">\; P</span>}<span\; class="notranslate">\; @</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}^{{\mathrm{<span\; class="notranslate">\; Q</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}}\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; @</span>\mathrm{<span\; class="notranslate">\; no</span>}}{{\mathrm{<span\; class="notranslate">\; Q</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}}$

Considering that users often want to find the information they need in the first few results when viewing the retrieval results, n=10 is often set.

By means of statistical machine translation, the present invention uses translated words to represent the "lexical ambiguity" and "lexical gap" between query questions and candidate answers, which can effectively solve the above two problems. Table 1 presents experiments on answer retrieval performance with the aid of statistical machine translation.

search method MAP P@10 TRLM 0.436 0.261 SMT 0.564 (↑29.36%) 0.291 (↑11.49%)

Table 1: Experiments on answer retrieval performance with the help of statistical machine translation

As shown in Table 1, TRLM represents the traditional answer retrieval method based on monolingual translation; SMT represents the answer retrieval method proposed by the present invention with the help of statistical machine translation. Through the comparison in Table 1, it can be seen that the method of the present invention significantly improves the performance of answer retrieval. For example, MAP increased by 29.36%, and P@10 increased by 11.49%. Experimental results prove that the present invention can better improve the performance of answer retrieval.

From the experimental results in Table 1 above, it can be seen that the answer retrieval method with the help of statistical machine translation has achieved good results in performance, and this method has been proved to be effective.

The specific embodiments described above have further described the purpose, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above descriptions are only specific embodiments of the present invention, and are not intended to limit the present invention. Within the spirit and principles of the present invention, any modifications, equivalent replacements, improvements, etc., shall be included in the protection scope of the present invention.

Claims (12)

1. An answer retrieval method by means of statistical machine translation, comprising the steps of:

step 1, translating all candidate answers represented by an original language into other languages by means of a statistical machine translation tool;

step 2, integrating candidate answers expressed by each language including the original language into a framework based on non-negative matrix decomposition;

step 3, solving the non-negative matrix decomposition-based framework by using a least square method fast gradient descent algorithm to obtain low-dimensional expression of each language expression of all candidate answers;

step 4, translating the query problem expressed by the original language into other multi-language translations by means of a statistical machine translation tool;

step 5, translating the query question and other multi-language translations into a low-dimensional space by using the low-dimensional expression of each language expression of all the candidate answers obtained in the step 3;

and 6, calculating the similarity between the query question and the candidate answers corresponding to the query question and the other multi-language translations according to the query question and the other multi-language translations and the low-dimensional expression of the candidate answers corresponding to the query question and the other multi-language translations, and obtaining a final retrieval result according to the similarity.

2. The method of claim 1, wherein the non-negative matrix factorization based framework is specifically represented as follows:

$F^{\&prime;\&prime;}({\stackrel{\&OverBar;}{U}}_1,...,{\stackrel{\&OverBar;}{U}}_L;{\stackrel{\&OverBar;}{V}}_1,...,{\stackrel{\&OverBar;}{V}}_L)=\underset{p=1}{\overset{L}{\&Sigma;}}\left|\right|{\stackrel{\&OverBar;}{D}}_p-{\stackrel{\&OverBar;}{U}}_p{\stackrel{\&OverBar;}{V}}_p|{|}_F^2+\underset{p=2}{\overset{L}{\&Sigma;}}{\mathrm{\&lambda;}}_p||{\stackrel{\&OverBar;}{V}}_p-{\stackrel{\&OverBar;}{V}}_1|{|}_F^2$

wherein,an objective function representing the framework; l represents the number of all languages within the original language;representing an M corresponding to the p-th language_p× N-dimensional word-document matrix, M_pRepresenting the number of non-repeated words in all candidate answer sets, N representing the number of all candidate answers, vectorEach element in the first candidate answer corresponds to a word in the ith candidate answer, and the element value of each element represents the importance degree of the word in the ith candidate answer;to representThe coefficient matrix obtained after the decomposition is obtained,to representObtaining a reconstruction matrix after decomposition; i | · | purple wind_FExpressing the norm of the matrix, parameter λ_pIs used to adjust the relative weights of the two parts,representing the reconstruction matrix corresponding to the original language.

3. The method according to claim 2, wherein the non-negative matrix factorization based framework is solved, in particular found, using the least squares based fast gradient descent algorithmAnda local optimal solution of; wherein, when optimizing the p coefficient matrixAt the same time, keepAndinvariant, to coefficient matricesPerforming iterative update to the objective functionThe following optimization problems are changed:

$\underset{{\stackrel{\&OverBar;}{U}}_p\&GreaterEqual;0}{min}\left|\right|{\stackrel{\&OverBar;}{D}}_p-{\stackrel{\&OverBar;}{U}}_p{\stackrel{\&OverBar;}{V}}_p|{|}_F^2.$

4. the method of claim 3, wherein the p-th reconstruction matrix is optimized when optimizing the p-th reconstruction matrixTime-keeping coefficient matrixAnd reconstructing the matrixInvariant, to reconstruction matrixPerforming iterative update to the objective functionThe optimization problem is changed into the following two types:

the first optimization problem is when p ∈ [2, L ]]，The following objective function was converted:

$\underset{{\stackrel{\&OverBar;}{V}}_p\&GreaterEqual;0}{min}\left|\right|{\stackrel{\&OverBar;}{D}}_p-{\stackrel{\&OverBar;}{U}}_p{\stackrel{\&OverBar;}{V}}_p|{|}_F^2+{\mathrm{\&lambda;}}_p||{\stackrel{\&OverBar;}{V}}_p-{\stackrel{\&OverBar;}{V}}_1|{|}_F^2$

the second type of optimization problem: when p is equal to 1, the compound is,the following objective function was converted:

$\underset{{\stackrel{\&OverBar;}{V}}_1\&GreaterEqual;0}{min}\left|\right|{\stackrel{\&OverBar;}{D}}_1-{\stackrel{\&OverBar;}{U}}_1{\stackrel{\&OverBar;}{V}}_1|{|}_F^2+{\mathrm{\&lambda;}}_1|\left|{\stackrel{\&OverBar;}{V}}_1\right|{|}_F^2.$

5. the method of claim 3, wherein the coefficient matrix is aligned withWhen iterative updating is carried out, the optimization problem of the objective function is decomposed into M_pEach sub-optimization problem corresponds to a coefficient matrixOne row of (c):

$\underset{{\stackrel{\&OverBar;}{u}}_i^{\left(p\right)}\&GreaterEqual;0}{min}\left|\right|{\stackrel{\&OverBar;}{d}}_i^{\left(p\right)}-{\stackrel{\&OverBar;}{V}}_P^T{\stackrel{\&OverBar;}{u}}_i^{\left(p\right)}|{|}_2^2$

wherein,representing a column vector, representing a matrixAll elements of row i of (1);representing a column vector, representing a matrix of coefficientsAll elements of row i.

6. The method of claim 4, wherein the reconstruction matrix is reconstructedWhen iterative updating is carried out, the first-class optimization problem is decomposed into N sub-optimization problems which are independent from each other, and each sub-optimization problem corresponds to a reconstruction matrixOne column of (c):

$\underset{{\stackrel{\&OverBar;}{v}}_j^{\left(p\right)}\&GreaterEqual;0}{min}\left|\right||{\stackrel{\&OverBar;}{d}}_j^{\left(p\right)}-{\stackrel{\&OverBar;}{U}}_p{\stackrel{\&OverBar;}{v}}_j^{\left(p\right)}|{|}_2^2+{\mathrm{\&lambda;}}_p\left|\right|{\stackrel{\&OverBar;}{L}}_j^{\left(p\right)}-{\stackrel{\&OverBar;}{v}}_j^{\left(1\right)}|{|}_2^2$

wherein, defineIs a matrixThe j-th column vector of (1),representing a reconstruction matrixThe jth column vector of (1);

likewise, the second type of optimization problem can be solved in the same way as the first type of optimization problem.

7. The method of claim 5, wherein M is_pThe numerical solution corresponding to each of the independent sub-optimization problems is:

${\stackrel{\&OverBar;}{u}}_i^{\left(p\right)}={\left({\stackrel{\&OverBar;}{V}}_p{\stackrel{\&OverBar;}{V}}_p^T\right)}^{-1}{\stackrel{\&OverBar;}{V}}_p{\stackrel{\&OverBar;}{d}}_i^{\left(p\right)}$

8. the method of claim 6, wherein the first type of optimization problem corresponds to a numerical solution of:

${\stackrel{\&OverBar;}{v}}_j^{\left(p\right)}={({\stackrel{\&OverBar;}{U}}_p^T{\stackrel{\&OverBar;}{U}}_p+{\mathrm{\&lambda;}}_p\stackrel{\&OverBar;}{I})}^{-1}({\stackrel{\&OverBar;}{U}}_p^T{\stackrel{\&OverBar;}{d}}_j^{\left(p\right)}+{\mathrm{\&lambda;}}_p{\stackrel{\&OverBar;}{v}}_j^{\left(1\right)})$

wherein, p ∈ [2, L]Representing the p-th language after translation,representing an identity matrix;

the numerical solution corresponding to the second type of optimization problem is as follows:

${\stackrel{\&OverBar;}{v}}_j^{\left(1\right)}={({\stackrel{\&OverBar;}{U}}_1^T{\stackrel{\&OverBar;}{U}}_1+{\mathrm{\&lambda;}}_1\stackrel{\&OverBar;}{I})}^{-1}{\stackrel{\&OverBar;}{U}}_1^T{\stackrel{\&OverBar;}{d}}_j^{\left(1\right)}.$

9. the method of claim 2, wherein step 5 transforms the query question into a low-dimensional space using a low-dimensional representation of said each linguistic representation of said all candidate answers, as calculated by:

${\stackrel{\&OverBar;}{v}}_{q_1}=\arg \underset{\stackrel{\&OverBar;}{v}\&GreaterEqual;0}{min}\left|\right|{\stackrel{\&OverBar;}{q}}_1-{\stackrel{\&OverBar;}{U}}_1\stackrel{\&OverBar;}{v}|{|}_2^2+{\mathrm{\&lambda;}}_1|\left|\stackrel{\&OverBar;}{v}\right|{|}_2^2$

wherein,is a query question q₁Is used to represent the vector of (a),is a query question q₁The vector representation in the low-dimensional space,a coefficient matrix corresponding to the original language is represented,representing query questions q₁A low-dimensional vector representation of (2), parameter lambda₁For adjusting the relative weights of the two parts.

10. The method of claim 2, wherein step 5 uses said low-dimensional representation of each language representation of all candidate answers to translate other multi-language translations into a low-dimensional space, as follows:

${\stackrel{\&OverBar;}{v}}_{q_i}=\arg \underset{\stackrel{\&OverBar;}{v}\&GreaterEqual;0}{min}\left|\right|{\stackrel{\&OverBar;}{q}}_i-{\stackrel{\&OverBar;}{U}}_i\stackrel{\&OverBar;}{v}|{|}_2^2+{\mathrm{\&lambda;}}_i||\stackrel{\&OverBar;}{v}-{\stackrel{\&OverBar;}{v}}_{q_1}|{|}_2^2$

wherein,other multi-language translations q that are query questions_iIs used to represent the vector of (a),other multi-language translations q representing correspondences to query questions_iA corresponding coefficient matrix;representing query questions q₁Corresponding translation q_iA low-dimensional vector representation of (a),representing query questions q₁Is represented by an optimal low-dimensional vector of (a) parameter λ_iFor adjusting the relative weights of the two parts.

11. The method of claim 1, wherein the query question q₁And candidate answer d₁The similarity in the low-dimensional space is calculated as follows:

$s(q_1,d_1)=\frac{<{\stackrel{\&OverBar;}{v}}_{q_1},{\stackrel{\&OverBar;}{v}}_{d_1}>}{\left|\right|{\stackrel{\&OverBar;}{v}}_{q_1}|{|}_2\&CenterDot;|\left|{\stackrel{\&OverBar;}{v}}_{d_1}\right|{|}_2}$

wherein, s (q)₁，d₁) Representing query questions q₁And candidate answer d₁The similarity in the low-dimensional space is,andseparately representing query questions q₁And candidate answer d₁Vector representation in a low-dimensional space;

likewise, question q is queried₁Corresponding translation q_iAnd candidate answer d₁Corresponding translation d_iThe similarity in the low-dimensional space is calculated by the same method.

12. An answer retrieval device with statistical machine translation, comprising:

the candidate answer translation module is used for translating the candidate answers into other languages;

the matrix decomposition module integrates candidate answers expressed by each language including the original language into a framework based on non-negative matrix decomposition;

the optimization solving module is used for solving the framework based on the non-negative matrix decomposition by utilizing a least square method fast gradient descent algorithm to obtain the low-dimensional expression of each language expression of all candidate answers of each question;

the query question translation module is used for translating the query question into other languages;

the similarity calculation module is used for converting the query question to the low-dimensional space and calculating the similarity of the query question and the candidate answers in the low-dimensional space;

and the result sequencing learning module is used for obtaining the retrieval answer finally according to the similarity obtained by the similarity calculation module.