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

Abstract

The invention discloses an answer search method and device by the aid of statistical machine translation. The method comprises the steps of firstly, translating candidate answers into a plurality of other languages by using a statistical machine translation tool to obtain a plurality of equivalent representations of the candidate answers; then, reducing dimensionalities of the plurality of equivalent representations of the candidate answers through a matrix decomposition method to obtain a low-dimension implication representation form; next, translating an inquired question into the low-dimension implication representation form through statistical machine translation and the matrix decomposition method; and finally, calculating the similarities between the inquired question and the candidate answers in implication space, and returning a plurality of candidate answers with the highest similarity as the answer of the inquired question. By means of the method, problems of vocabulary mismatching and ambiguity can be solved effectively, and tests prove that the answer research performance is improved by 29.36% in large-scale community question and answer data sets.

Description

Answer retrieval method and device by means of statistical machine translation

Technical Field

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

Background

With the rapid development of internet technology, User-Generated Content (UGC) -based internet services are becoming more and more popular. Community question-answering is a new "question-answer" based information exchange and knowledge sharing system that appears in this context, such as Yahoo! Answers, hundredths, etc. Different from an automatic question-answering system, on the community question-answering, a user can put forward any type of questions and can also answer any type of questions of other users. Answer retrieval is the basis of community question-answer analysis and occupies an important position. The task of answer retrieval is to retrieve answers semantically similar or similar to the query question, which the user answers, from a large-scale candidate answer library. Therefore, the answer retrieval has important theoretical significance and practical value.

The main challenges facing current answer retrieval are lexical mismatch between the query question and the candidate answer and lexical ambiguity problems. The word mismatch usually causes the answer retrieval model to retrieve many answers that do not match with the query intention of the user, mainly because the query question and the answer in the community question-answer are both given by the user, and the query intention of the user is highly diversified. For example, the word "interest" may refer to both "curiosity" and "a charge for billing money" depending on the user. The term ambiguity is a common phenomenon between a query question and a candidate answer, and is particularly shown in the following points that many terms appear in the query question and the candidate answer only a few times, even none of the terms appear in the query question or the candidate answer, and a traditional method based on term matching cannot be used.

One way to solve the above "lexical ambiguity" and "lexical gap" problem is to represent ambiguous words in the original language and words that are literally different from each other by their corresponding translations, by means of statistical machine translation. The method of statistical machine translation is characterized by firstly establishing a reasonable objective function, integrating the original language and the corresponding translation thereof in a frame, secondly reducing the noise caused by the statistical machine translation as much as possible, and finally designing a rapid solving method to solve the objective function. The obtained translation words are directly added into the original language, the accuracy of answer retrieval is greatly reduced, the main reason is that the calculation complexity is greatly increased when the translation words are directly added into the original language, and meanwhile, machine translation errors bring much noise.

The task of answer retrieval refers to retrieving answers capable of answering the query from a set of answer documents for a query question input by a user. The main difficulty faced in answer retrieval is that the user query question and the candidate answer use different wording forms when expressing the same or similar meanings, which easily results in the problems of word mismatching and word ambiguity. The traditional method mainly depends on mining word association between monolingues, and ignores semantic association between multilingual information.

Disclosure of Invention

In order to solve the above problems, the present invention firstly needs to design a reasonable objective function, effectively integrate the original language and its corresponding translation into a framework, and at the same time, constrain the influence of the noise of machine translation on the answer retrieval under the framework. Then, a quick solving method is designed according to the established objective function and the constraint thereof. And solving the target function to obtain the implicit expression of the original language and the corresponding translation of the original language, and finally calculating the similarity between the user query and the candidate answer in an implicit space. According to the thought, the method provided by the invention is mainly used for solving two problems existing in answer retrieval, and successfully introduces statistical machine translation into the answer retrieval process.

The basic idea of the invention is to represent ambiguous words in the original language and words which are literally represented differently by using their corresponding translations by fully using statistical machine translation, thereby improving the performance of answer retrieval.

The invention discloses

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.

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

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

a matrix decomposition module for integrating 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 sorting learning module is used for obtaining the retrieval answer finally according to the similarity obtained by the similarity calculation module.

The invention adopts the idea of statistical machine translation to improve the performance of answer retrieval. And (3) utilizing a statistical machine translation tool Google Translate to represent ambiguous words in the original language and words with different literal representations by using the corresponding translations, thereby improving the performance of answer retrieval.

Drawings

FIG. 1 is a diagram of an answer retrieval method with statistical machine translation according to the present invention.

FIG. 2 is a block diagram of an answer retrieval device using statistical machine translation according to the present invention.

Detailed Description

In order that the objects, technical solutions and advantages of the present invention will become more apparent, the present invention will be further described in detail with reference to the accompanying drawings in conjunction with the following specific embodiments.

The invention discloses an answer retrieval method and device by means of statistical machine translation. It can be divided into an off-line process and an on-line process. The off-line process is realized by three modules, namely a candidate answer translation module, a matrix decomposition module and an optimization solving module. The online process is also carried out by three modules, namely a query question translation module, a similarity calculation module based on a low-dimensional space and a result sequencing learning module.

Fig. 1 shows an answer retrieval method by means of statistical machine translation according to the present invention. As shown in fig. 1, it includes two stages, an offline part and an online part. Wherein the off-line process comprises:

step (1), using statistical machine translation tool to use original language l₁All candidate answers in the representation (e.g., English) are translated to obtain L-1 equivalent representations in different languages, { L }₁，l₂，…，l_L-1L represents the number of all languages, and the statistical machine translation tool can be selected from Google Translate and the like.

Step (2), representing the candidate answer set represented by each language into M_p× N word-document matrixWherein M is_pAll words in the candidate answer set represented in the p language are represented, and N represents the number of answers in the candidate answer set.

And (3) designing a new objective function, integrating candidate answers expressed by P different languages into a unified frame by adopting a non-negative matrix decomposition method, and reducing noise caused by statistical machine translation by adopting a regularization strategy.

Step (4), designing a rapid gradient descent algorithm based on least square, and solving the objective function to obtain low-dimensional expression forms of L different languages, namely coefficient matrixAnd reconstructing the matrix

The online process comprises the following steps:

step (1) using statistical machine translation tool to convert original language l₁(e.g., English) representation of the query question into L-1 equivalent representations in different languages, and the statistical machine translation tool may be selected from Google Translate, and the like.

Step (2), coefficient matrix obtained by solving in the off-line process (4) is utilizedAnd converting the query question and the corresponding L-1 translation representations to a low-dimensional space. (ii) a

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

And (4) adopting a linear sequencing learning strategy to fuse the similarity expressed by the L different languages in the low-dimensional space, and returning a plurality of candidate answers with the highest score as final answers.

Fig. 2 shows an answer retrieval apparatus by statistical machine translation proposed in the present invention. As shown in fig. 2, the search device includes: the system comprises a candidate answer translation module, a matrix decomposition module, an optimization solving module, a query problem translation module and a similarity calculation module based on a low-dimensional space.

The candidate answer translation module is used for using the original language l in the off-line stage₁All candidate answers in the representation (e.g., English) are translated to obtain L-1 equivalent representations in different languages, { L }₁，l₂，…，l_L-1Where L represents the number of all languages, i.e. by applying to the candidate answer set D₁Translating to obtain candidate answer set D represented by other L-1 languages₂，…，D_L。

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, manual translation is time-consuming and labor-consuming, and particularly for the real task of searching community question and answer answers, translation of large-scale candidate answers is obviously unrealistic. Fortunately, the level of machine translation is currently well developed in natural language processing, although it is not yet fully satisfactory in terms of translation quality. There are many free translation tools that have been published to dateProviding daily translation services. In the preferred embodiment of the invention, Google Translate is adopted, and the translation tool trains a translation model on the constructed large-scale parallel corpus by utilizing a statistical machine learning method, so that rich context information can be considered in the process of translating from one language to another language, and good translation performance is expressed in a plurality of translation tools. By applying to the candidate answer set D₁After translation, a candidate answer set D represented by other L-1 languages can be obtained₂，…，D_L。

The matrix decomposition module is used for representing a candidate answer set represented by each language into M in an off-line stage_p× N word-document matrixWherein M is_pAll words in the candidate answer set represented in the p language are represented, and N represents the number of answers in the candidate answer set.

The matrix decomposition module is one of the key technologies of the invention. Definition l₁，l₂，…，l_LDenotes a set of languages used in the present invention, where L denotes the number of languages, L₁Representing the original language (e.g. English), l₂…l_pRepresenting another L-1 language. Definition ofRepresentation is based on₁A set of candidate answers expressed in language. Defining candidate answersCan be expressed as an M_pVector of dimensionsWherein the vectorEach element in (a) corresponds to a word representing the degree of importance of the word in the ith candidate answer(ii) a The vectorTf-idf, a statistical method to assess how important a word is to one of a set of files or a set of data, can be used. D_pCan be expressed as an M_p× N-dimensional word-document matrixIn the matrix, each row represents a different word and each column represents a candidate answer, where M_pRepresents D_pNumber of non-repeated words, N represents D_pThe number of candidate answers in (1).

Intuitively, the translated candidate answer set D represented by other L-1 languages₂，…，D_LIs directly added to the original candidate answer set D₁This will result in D₁Corresponding matrixDimension of from M₁× N is increased toHowever, this approach has two disadvantages: (1) causing data sparsity; (2) statistical machine translation errors will cause noise problems. In order to solve the above problem, the present invention adopts a matrix decomposition method.

Hypothesis matrixCan be decomposed into two low-dimensional matricesAndtaking into account matrices simultaneouslyIndependently ofThe following objective function can be obtained:

$F({\stackrel{\‾}{U}}_p,{\stackrel{\‾}{V}}_p)=\underset{{\stackrel{\‾}{U}}_p\≥0,{\stackrel{\‾}{V}}_p\≥0}{\min }{\left|\right|{\stackrel{\‾}{D}}_p-{\stackrel{\‾}{U}}_p{\stackrel{\‾}{V}}_p\left|\right|}_F^2$

wherein | · | purple sweet_FRepresenting the norm of a matrix, whereinA coefficient matrix obtained after the decomposition is represented,and expressing a reconstruction matrix obtained after decomposition, wherein K expresses the dimension of the hidden space.

In order to reduce the noise problem caused by statistical machine translation errors, the invention assumes that the slave matrix(p∈[2，L]) Obtained reconstruction matrixShould be matched with the slave matrixObtained reconstruction matrixThe closer the better. Therefore, the present invention proposes to minimize the reconstruction matrix(p∈[2，L]) And reconstruction matrixDistance before:

$F^{\′}\left({\stackrel{\‾}{V}}_p\right)=\underset{{\stackrel{\‾}{V}}_p\≥0}{\min }\underset{p=2}{\overset{L}{\mathrm{\Σ}}}{\left|\right|{\stackrel{\‾}{V}}_p-{\stackrel{\‾}{V}}_1\left|\right|}_F^2$

combining the two objective functions, the following objective functions can be obtained:

$F^{\′\′}({\stackrel{\‾}{U}}_1,\·\·\·,{\stackrel{\‾}{U}}_L;{\stackrel{\‾}{V}}_1,\·\·\·,{\stackrel{\‾}{V}}_L)=\underset{p=1}{\overset{L}{\mathrm{\Σ}}}{\left|\right|{\stackrel{\‾}{D}}_p-{\stackrel{\‾}{U}}_p{\stackrel{\‾}{V}}_p\left|\right|}_F^2+\underset{p=2}{\overset{L}{\mathrm{\Σ}}}{\mathrm{\λ}}_p{\left|\right|{\stackrel{\‾}{V}}_p-{\stackrel{\‾}{V}}_1\left|\right|}_F^2$

wherein the parameter lambda_p(p∈[2，L]) For adjusting the relative weights of the two parts. If for the parameter lambda_pSetting a smaller value, the above-mentioned objective functionSimilar to a conventional Non-negative matrix (Non-negative matrix), if for the parameter λ_pSetting a larger value, the above-mentioned objective functionWith greater emphasis on statistical machinesErrors introduced by translation.

The optimization solving module is used for solving the parameters, namely the coefficient matrix, in the matrix decomposition moduleAnd reconstructing the matrixObtaining a coefficient matrix through the optimization solving moduleAnd reconstructing the matrixI.e. the input result of the offline part.

The optimization solving module is one of the core technologies of the invention. The above objective functionSimultaneously considering the problems of data sparsity and statistical machine translation errors, the objective function has 2L paired optimization objects, and when simultaneously consideringAndit is difficult to find an algorithm to solve the minimization problem. The invention provides a fast gradient descent algorithm based on a least square method, which is used for finding a local optimal solution, and other 2L-1 objects are kept unchanged when a certain target object is optimized.

HoldingAndthe temperature of the molten steel is not changed,for coefficient matrixMay update the objective functionThe following optimization problems are changed:

$\underset{{\stackrel{\‾}{U}}_p\≥0}{\min }{\left|\right|{\stackrel{\‾}{D}}_p-{\stackrel{\‾}{U}}_p{\stackrel{\‾}{V}}_p\left|\right|}_F^2$

definition ofRepresenting 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. Thus, the optimization problem described above can be decomposed into M_pEach sub-optimization problem corresponds to a coefficient matrixOne row of (c):

$\underset{{\stackrel{\‾}{u}}_i^{\left(p\right)}\≥0}{\min }{\left|\right|{\stackrel{\‾}{d}}_i^{\left(p\right)}-{{\stackrel{\‾}{V}}^T}_p{\stackrel{\‾}{u}}_i^{\left(p\right)}\left|\right|}_2^2$

subscript i ═ 1, …, M_pWherein M is_pRepresents D_pThe number of non-repeated words in (1).

The sub-optimization problem is a standard least squares problem whose numerical solution is:

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

matrix of hold coefficientsAnd reconstructing the matrixInvariant, to reconstruction matrixMay update the objective functionThe optimization problem is changed into the following two types:

when p ∈ [2, L ]]，Can be converted to the following objective function:

$\underset{{\stackrel{\‾}{V}}_p\≥0}{\min }{\left|\right|{\stackrel{\‾}{D}}_p-{\stackrel{\‾}{U}}_p{\stackrel{\‾}{V}}_p\left|\right|}_F^2+{\mathrm{\λ}}_p{\left|\right|{\stackrel{\‾}{V}}_p-{\stackrel{\‾}{V}}_1\left|\right|}_F^2$

when p is equal to 1, the compound is,can be converted to the following objective function:

$\underset{{\stackrel{\‾}{V}}_1\≥0}{\min }{\left|\right|{\stackrel{\‾}{D}}_1-{\stackrel{\‾}{U}}_1{\stackrel{\‾}{V}}_1\left|\right|}_F^2+{\mathrm{\λ}}_1{\left|\right|{\stackrel{\‾}{V}}_1\left|\right|}_F^2$

for the objective function of the first case mentioned above, the definitionIs a matrixThe j-th column vector of (1),representing a reconstruction matrixJ-th column vector of (1). Therefore, the objective function of the first case can be decomposed into N independent sub-optimization problems, each of which corresponds to a reconstruction matrixOne column of (c):

$\underset{{\stackrel{\‾}{v}}_j^{\left(p\right)}\≥0}{\min }{\left|\right|{\stackrel{\‾}{d}}_j^{\left(p\right)}-{\stackrel{\‾}{U}}_p{\stackrel{\‾}{v}}_j^{\left(p\right)}\left|\right|}_2^2+{\mathrm{\λ}}_p{\left|\right|{\stackrel{\‾}{v}}_j^{\left(p\right)}-{\stackrel{\‾}{v}}_j^{\left(1\right)}\left|\right|}_2^2$

where the subscript j ═ 1, …, N, denotes the set D_pThe number of candidate answers in (1).

The sub-optimization problem is a standard L-based₂A regularized least squares problem, then its numerical solution is:

${\stackrel{\‾}{v}}_j^{\left(p\right)}={({{\stackrel{\‾}{U}}^T}_p{\stackrel{\‾}{U}}_p+{\mathrm{\λ}}_p\stackrel{\‾}{I})}^{-1}({{\stackrel{\‾}{U}}^T}_p{\stackrel{\‾}{d}}_j^{\left(p\right)}+{\mathrm{\λ}}_p{\stackrel{\‾}{v}}_j^{\left(1\right)})$

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

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

${\stackrel{\‾}{v}}_j^{\left(1\right)}={({{\stackrel{\‾}{U}}^T}_1{\stackrel{\‾}{U}}_1+{\mathrm{\λ}}_1\stackrel{\‾}{I})}^{-1}{\stackrel{\‾}{U}}_1^T{\stackrel{\‾}{d}}_j^{\left(1\right)}$

the query question translation module is used for translating the query question into equivalent expressions of L-1 different languages by utilizing a statistical machine translation tool in an online stage, wherein the statistical machine translation tool can select Google Translate and the like.

Similar to the candidate answer translation module, to Translate a query question from one language to another L-1 language, the present invention relies on the statistical machine translation tool, Google Translate. For a given query question q, the query question q is translated to obtain other L-1 language expressions₂，…，q_L。

And the similarity calculation module based on the low-dimensional space is used for calculating the similarity of the query question and the candidate answer on the low-dimensional space representation.

The similarity calculation module based on the low-dimensional space is one of the key technologies of the invention. For a given query question q and its corresponding translation q in L-1 languages₂，…，q_LIt needs to be transformed to a low dimensional space. For the sake of convenience of presentation, the symbol q is used₁Substituting the query question q in the original language, i.e. q ═ q₁. Therefore, q can be expressed by the following formula₁Conversion to a low dimensional space:

${\stackrel{\‾}{v}}_{q_1}=\arg \underset{\stackrel{\‾}{v}\≥0}{\min }{\left|\right|{\stackrel{\‾}{q}}_1-{\stackrel{\‾}{U}}_1\stackrel{\‾}{v}\left|\right|}_2^2+{\mathrm{\λ}}_1{\left|\right|\stackrel{\‾}{v}\left|\right|}_2^2$

wherein,is a query question q₁Is used to represent the vector of (a),is a query question q₁Vector representation on a low-dimensional space, namely a reconstruction matrix; whereinAnd expressing a coefficient matrix corresponding to the original language obtained by the optimization solving module. However for candidate answer d₁The conversion result obtained after the low-dimensional conversion can be directly performed by using the matrix decomposition module, namelyQuery question q₁And candidate answer d₁The similarity in the low-dimensional space can be represented by cosine similarity:

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

wherein, s (q)₁，d₁) Representing query questions q₁And candidate answer d₁Similarity in low dimensional space.

For q₁Corresponding translation q_i(i∈[2，L]) In other words, it can be expressed to a low-dimensional space using the following formula:

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

wherein,is a query question q_iIs represented by a vector of (a). Class ISimilarly, for candidate answer d₁Corresponding translation d_i(i∈[2，L]) In other words, the matrix decomposition module can be directly utilized to obtain the result after low-dimensional space conversionQuery question q₁Corresponding translation q_iAnd candidate answer d₁Corresponding translation d_iThe similarity in the low-dimensional space may employ a similar cosine similarity calculation method as described above.

And the result sorting learning module is used for fusing the similarity expressed in the low-dimensional space by the L different languages, and returning a plurality of candidate answers with the highest score as final answers. For a given query question q₁And candidate answer d₁The invention designs a sort learning function as follows:

$\mathrm{Score}(q_1,d_1)=\stackrel{\&OverBar;}{\mathrm{\&theta;}}\&CenterDot;\mathrm{\&Phi;}(q_1,d_1)$

wherein, Score (q)₁，d₁) Representing query questions q₁And candidate answer d₁The final score is then calculated based on the results of the scoring,weight, phi (q), representing the feature vector₁，d₁)＝{s(q₁，d₁)，s(q₂，d₂)，…，s(q_L，d_L) Represents a feature vector corresponding to the query question q₁And candidate answer d₁The similarity of the L different languages in the low-dimensional space is represented. Wherein the parametersAnd obtaining the optimal value by adopting the most common cross-validation strategy in the statistical machine learning. Finally, according to Score (q)₁，d₁) The candidate answers with the highest scores are returned as final answers.

To illustrate the performance of the device, the present invention experimentally verified the improvement in answer retrieval system performance by statistical machine translation methods.

The experimental data of the present invention are derived from Yahoo! The Answers community question-answering system, in these historical problem sets, each problem 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 used contains 1232 user category labels, 2,288,607 question-and-answer pairs. To evaluate the effectiveness of the inventive method, we additionally selected 252 query questions as the test data set. For each query question in the test dataset, we used the language model to retrieve the best 20 results and then let both annotators label manually. If the returned candidate answer is similar to the query question, it is labeled "relevant", otherwise it is labeled "irrelevant". If the annotation structures of two annotators conflict, a third person is allowed to make a final decision. In determining whether the candidate answer is similar to the query question, the annotator only knows the question itself.

In the present invention, the parameter L is set to 5, i.e. english needs to be translated into other 4 languages (chinese, french, italian, german).

Suppose Q_tRepresenting a test problem set, the invention adopts the following two evaluation indexes:

average correct rate (MAP): the calculation formula is as follows:

$\mathrm{MAP}\left(Q_t\right)=\frac{1}{Q_t}\underset{q\&Element;Q_t}{\mathrm{\&Sigma;}}\frac{1}{m_q}\underset{k=1}{\overset{m_q}{\mathrm{\&Sigma;}}}\mathrm{Precision}\left(R_k\right)$

wherein m is_qIs the number of questions, R, associated with the query question q_kIs the set of the kth question and all questions before it in the search results, Precision (R)_k) Is R_kProblem ratio related to q. This index reflects the average level of the test results as a whole.

Precision @ n (P @ n): defined as the accuracy of the first n results returned by the system for the query question. Precision @ n of the whole test set is the average value of Precision @ n of all problems in the test set, and the calculation formula is as follows:

$P\left(q\right)@n=\frac{k}{n}$

wherein k represents the number of related questions in the first k questions returned by the retrieval system, and n represents the total number of questions returned by the retrieval system. Therefore, the temperature of the molten metal is controlled,

$P@n=\frac{{\mathrm{\&Sigma;}}_{q=1}^{Q_t}p\left(q\right)@n}{Q_t}$

considering that a user often wants to find information needed by the user in the first few results when viewing a search result, n is often set to 10.

The invention uses the translated words to represent the 'vocabulary ambiguity' and 'vocabulary gap' problems between the query question and the candidate answer by means of statistical machine translation, and can effectively solve the two problems. Table 1 gives the experiment of answer search performance by statistical machine translation.

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

Table 1: experiment of answer search performance by statistical machine translation

As shown in table 1, TRLM represents a conventional answer retrieval method based on a single language translation; SMT represents the answer retrieval method proposed by the present invention with statistical machine translation. By comparing table 1, it can be seen that the method of the present invention significantly improves the performance of answer retrieval. For example, the MAP is improved by 29.36 percent, and the P @10 is improved by 11.49 percent. Experimental results prove that the method can better improve the answer retrieval performance.

As can be seen from the experimental results of table 1 above, the answer retrieval method by means of statistical machine translation has achieved good performance, and this method has proved to be effective.

The above-mentioned embodiments are intended to illustrate the objects, technical solutions and advantages of the present invention in further detail, and it should be understood that the above-mentioned embodiments are only exemplary embodiments of the present invention and are not intended to limit the present invention, and any modifications, equivalents, improvements and the like made within the spirit and principle of the present invention should 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.