JP2014006620A - Synonym estimation device, synonym estimation method, and synonym estimation program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a synonym estimation device, a synonym estimation method, and a synonym estimation program, capable of more accurately specifying a synonym similar to a query required for retrieval.SOLUTION: A synonym estimation device 10 generates a plurality of functions that calculate a new feature value by performing a weighting operation on a plurality of feature values to learn weighting of the plurality of feature values and the new feature value, and weighting of a combined feature value of the functions, the weighting enabling character strings registered in learning data 30 to be determined in a descending order of similarity. The synonym estimation device 10 calculates a new feature value for a plurality of synonym candidates retrieved from a query by using a function obtained after learning from the plurality of feature values to specify a synonym similar to the inputted query on the basis of a calculation result obtained by calculating the plurality of feature values and the calculated new feature value with the learned weighting.

Description

  The present invention relates to a synonym estimation device, a synonym estimation method, and a synonym estimation program.

  In a conventional web page search, when a query is input by a user, a search is performed by the search engine on the web page, and a search result including a plurality of URLs (Uniform Resource Locators) is presented to the user.

  By the way, although a constituent element of a query is dissimilar, there may be a synonym having a similar meaning to the query. In some cases, the search result is different depending on whether the search is performed using the input query or the synonym of the query. For example, when searching for A, there are 20000 search results, and when searching for A ', which is a synonym for A, the search results are 3000 and the search results may be different.

  If the system can understand that A and A 'are the same, the loss opportunity can be avoided by performing an OR search of two queries in advance.

  Therefore, the following conventional techniques for extracting synonyms of queries are known. For example, a conventional technique for searching for synonyms of an input query by label propagation using a search click-through log is known (for example, Patent Document 1). Also, in machine translation, a conventional technique is known in which feature sets are extracted from words in different languages and associated with words in different languages (see, for example, Patent Document 2).

JP 2012-79029 A JP 2010-198438 A

  However, the conventional technology may not be able to identify a synonym of a query for which a search is requested.

  The present application has been made in view of the above, and an object thereof is to provide a synonym estimation device, a synonym estimation method, and a synonym estimation program that can identify a synonym close to a query for which a search is requested with higher accuracy. And

  The synonym estimation device according to the present application includes a storage unit that stores learning data stored together with information indicating a meaning of a character string that can be regarded as a synonym, and a new feature value by weighting a plurality of feature values. Generating means for generating a plurality of nonlinear functions by changing the number of feature values and combinations of feature values, and the plurality of feature values capable of determining a character string in the order of meaning based on the learning data; and A learning means for learning weights for new feature values respectively calculated from a plurality of nonlinear functions, and weights for feature values combined with each nonlinear function, and a plurality of synonyms retrieved from a query for which a search is requested For each candidate, each feature value is weighted by weighting the combined feature values learned by the learning means from the plurality of feature values of the synonym candidates. A plurality of synonym candidates based on a calculation result obtained by calculating a new feature value using a shape function and calculating the plurality of feature values and the calculated new feature value by weighting learned by the learning unit And a specifying means for specifying a synonym having a meaning close to that of the query.

  The synonym estimation device, the synonym estimation method, and the synonym estimation program according to the present application can identify a synonym close to a query for which a search is requested with higher accuracy.

FIG. 1 is a diagram for explaining synonym search processing according to the embodiment. FIG. 2 is a diagram illustrating an example of features. FIG. 3 is a diagram illustrating an example of a new feature value. FIG. 4 is a graph showing the synonym candidates A and B in a graph of the features X and Y. FIG. 5 is a graph showing the synonym candidates A, B, C, and D using the feature X and Y graphs. FIG. 6 is a graph showing the synonym candidates A, B, C, and D in the feature Z graph. FIG. 7 is a diagram illustrating an example of a functional configuration of the synonym estimation device. FIG. 8 is a diagram illustrating an example of a data configuration of feature data. FIG. 9 is a diagram for explaining generation of a function. FIG. 10 is a flowchart showing the procedure of the learning process. FIG. 11 is a flowchart showing the procedure of the synonym specifying process.

  DESCRIPTION OF EMBODIMENTS Hereinafter, a form for implementing a synonym estimation apparatus, a synonym estimation method, and a synonym estimation program according to the present invention (hereinafter referred to as “embodiment”) will be described in detail with reference to the drawings. In addition, this invention is not limited by this embodiment.

[1. Synonym search processing]
First, synonym search processing performed by the synonym estimation device according to the embodiment will be described. FIG. 1 is a diagram for explaining synonym search processing according to the embodiment. In addition, the example of FIG. 1 has shown the case where the synonym of the input query is specified and the synonym nearest to a query is shown to a user as a suggestion query.

  In the synonym estimation device, a query to be searched is input by a user. The synonym estimation device searches for a plurality of synonym candidates for the input query. Then, the synonym estimation device ranks the plurality of searched synonym candidates as correction candidates, in the order similar to the input query, and presents the highest correction candidate to the user as a suggestion query.

  When ranking the correction candidates in the order of similarity, the synonym estimation device extracts features of the synonym candidates determined as the correction candidates in order to calculate the degree of association between the query and each correction candidate. FIG. 2 is a diagram illustrating an example of features. In the example of FIG. 2, when a synonym candidate is searched by label propagation, the score at the time of search by label propagation is extracted as a feature. In the example of FIG. 2, the likelihood of a query as a correction candidate is extracted as a feature based on the establishment calculated using a language model, TextRank, or the like. In the example of FIG. 2, the length of the correction candidate character string is extracted as a feature. In the example of FIG. 2, whether the query and the correction candidate match is extracted as a feature. In the example of FIG. 2, whether the query and the correction candidate have an Acronym relationship is extracted as a feature. Acronym is an abbreviation of only the first part of each word of the name, for example, “Nihon House Kyokai” is abbreviated as “NHK”. In the example of FIG. 2, the edit distance between the query and the correction candidate is extracted as a feature. The features are not limited to these, and may be other features.

  The synonym estimation device extracts each feature for each searched synonym candidate, and obtains a feature vector having the feature value of each feature of the synonym as an element for each synonym. Then, the synonym estimation device performs ranking of synonym candidates based on the feature vector.

  By the way, the retrieved synonym candidates have few clues in the extracted features, and the synonym candidates may not be sufficiently classified according to the degree of similarity with the query. In this way, when synonym candidates cannot be sufficiently classified, synonyms close to the input query cannot be accurately identified.

  Therefore, in the synonym estimation apparatus according to the present embodiment, for each synonym candidate, a combination of a plurality of feature values is calculated to obtain a new feature value. FIG. 3 is a diagram illustrating an example of a new feature value. In the example of FIG. 3, the feature values X_1, X_2, and X_3 are added to obtain a new feature value X'_1. In the example of FIG. 3, a new feature value X′_2 is obtained by multiplying the feature values X_1, X_2, and X_3.

  Here, the effect of calculating a new feature value will be described. Here, in order to simplify the description, the features will be described using two cases of X and Y. For example, the synonym candidate A is a synonym having a feature X of “1” and a feature Y of “0”, which is similar to a query. The synonym candidate B is assumed to be a synonym having a feature X of “0”, a feature Y of “1”, and dissimilar to the query. FIG. 4 is a graph showing the synonym candidates A and B in a graph of the features X and Y. In this case, for example, if the feature X is “1”, it is similar, and if the feature Y is “1”, it is determined that the synonym candidate is similar to the query.

  On the other hand, for example, the synonym candidate A is a synonym having a feature X of “1” and a feature Y of “1”, which is similar to a query. The synonym candidate B is a synonym that has a feature X of “−1”, a feature Y of “1”, and is dissimilar to the query. Further, the synonym candidate C is a synonym similar to the query, with the feature X being “−1” and the feature Y being “−1”. Further, the synonym candidate D is a synonym having a feature X of “1”, a feature Y of “−1”, and dissimilar to the query. FIG. 5 is a graph showing the synonym candidates A, B, C, and D using the feature X and Y graphs. In this case, the feature value of the feature X is “1” for both the synonym candidate A similar to the query and the synonym candidate D dissimilar to the query, and the synonym candidate B and query dissimilar to the query. Are both “−1” in candidate C of synonyms similar to. Also, the feature value of the feature Y is “1” for both the synonym candidate A similar to the query and the synonym candidate B dissimilar to the query, and the synonym candidate C and query similar to the query are not “1”. Both of the similar synonym candidates D are “−1”. Therefore, it cannot be determined from the feature values of the features X and Y whether the synonym candidate is similar to the query.

  Therefore, for example, for the synonym candidates A to D, the feature value of the feature X and the feature value of the feature Y are respectively multiplied to calculate a new feature value of the feature Z. In this case, in the synonym candidate A, the feature X is “1” and the feature Y is “1”, so the feature Z is “1”. The synonym candidate B has a feature X of “−1” and a feature Y of “1”, so the feature Z is “−1”. In the synonym candidate C, the feature X is “−1” and the feature Y is “−1”, so the feature Z is “1”. The synonym candidate D has the feature X of “1” and the feature Y of “−1”, so the feature Z is “−1”. FIG. 6 is a graph showing the synonym candidates A, B, C, and D in the feature Z graph. In this new feature Z, the values of the synonym candidates A and C similar to the query and the synonym candidates B and D dissimilar to the query are separated. Therefore, for example, it is possible to determine whether or not a synonym candidate is similar to a query by determining similarity if the feature Z is “1” and dissimilarity if the feature Z is “−1”. In this way, by obtaining a new feature, it becomes possible to discriminate characteristics that could not be discriminated, so that synonyms that could not be discriminated can be specified.

  By the way, in order to find a function that is effective for specifying synonyms, it is necessary to consider various combinations of features. For this reason, there is a limit in finding effective combinations of features in manual work by an administrator or the like. Moreover, when the feature combined with the function is not effective, the accuracy of the specified synonym may be reduced. Therefore, the synonym estimation device according to the present embodiment obtains a function effective for identifying synonyms by performing learning using learning data registered in association with character strings that can be regarded as synonyms having similar meanings. The synonym is specified by the obtained function.

[2. Configuration of synonym estimation device]
Hereinafter, the synonym estimation device 10 according to the present embodiment will be described in more detail. FIG. 7 is a diagram illustrating an example of a functional configuration of the synonym estimation device. As illustrated in FIG. 7, the synonym estimation device 10 includes a communication I / F (interface) unit 20, a storage unit 21, and a control unit 22.

  The communication I / F unit 20 is an interface such as a NIC (Network Interface Card). The communication I / F unit 20 transmits and receives various data to and from other devices via the network 11. Other devices such as a client terminal 12 operated by a user and a management terminal 13 operated by an administrator are communicably connected to the network 11.

  The client terminal 12 is an information processing device used by a user. For example, the client terminal 12 is a desktop PC (Personal Computer), a tablet PC, a notebook PC, a mobile phone, a PDA (Personal Digital Assistant), or the like.

  The client terminal 12 receives the web page provided by the synonym estimation device 10 and displays the received web page on a predetermined display unit (display). Such a web page is provided with an input area in which a character string to be searched can be input. When performing a search, the user operates the client terminal 12 to input a character string to be searched for in the input area of the web page displayed on the display unit, and instructs execution of the search. When the client terminal 12 is instructed to execute the search, the client terminal 12 outputs the character string input in the input area of the web page to the synonym estimation device 10 as a query.

  The management terminal 13 is an information processing device used when the administrator manages the synonym estimation device 10. For example, the management terminal 13 is a desktop PC, a tablet PC, a notebook PC, or the like. The administrator operates the management terminal 13 to perform registration of various data and various instructions for operating and managing the synonym estimation device 10. For example, the administrator registers learning data. In addition, the administrator gives an instruction to learn weighting conditions effective for specifying synonyms.

  The communication I / F unit 20 receives a query from the client terminal 12 via the network 11. Further, the communication I / F unit 20 receives various data and various instructions registered from the management terminal 13 via the network 11.

  The storage unit 21 is a storage device such as a hard disk or an optical disk. The storage unit 21 is not limited to the above-mentioned types of storage devices, and may be a semiconductor memory element such as a RAM (Random Access Memory) or a flash memory.

  The storage unit 21 stores an OS (Operating System) executed by the control unit 22 and various programs used for searching for synonyms. Furthermore, the storage unit 21 stores various data used in programs executed by the control unit 22. For example, the storage unit 21 stores learning data 30, feature data 31, and function data 32.

  The learning data 30 is data stored together with information indicating that the character string that can be regarded as a synonym is close in meaning. For example, in the learning data 30, each character string that can be regarded as a synonym is stored together with a score indicating the closeness of meaning. The score is determined by an administrator or the like. The learning data 30 is registered by an administrator or the like.

  The feature data 31 is data in which feature values are stored for each searched synonym candidate. The feature data 31 is generated by, for example, a derivation unit 44 described later. FIG. 8 is a diagram illustrating an example of a data configuration of feature data. The feature data 31 according to the present embodiment stores records for each synonym candidate, and the feature values of each feature are stored in tab-separated fields in order of predetermined features. As shown in FIG. 8, the feature data 31 includes fields 31A to 31D. The first field 31A is an area for storing a label indicating whether or not the synonym is registered in the learning data 30. The first field 31A stores “1” when the synonym is registered in the learning data 30, and stores “0” when the synonym is not registered in the learning data 30. The field 31B and subsequent fields are areas for storing feature values. The example of FIG. 8 shows a case where the number of feature values of each synonym candidate is three, and the feature values of each feature are stored in the fields 31B to 31D. The feature value is a real number. That is, as the feature value, a minus value or a numerical value less than the decimal point can be used. Thus, by setting the feature value as a real number, it is possible to hold not only the presence / absence of a characteristic but also the characteristic in detail.

  The function data 32 is data in which a function whose parameters have been changed by the learning unit 42 described later is stored. The function data 32 is registered by the learning unit 42.

  Returning to FIG. 7, the control unit 22 includes an internal memory for storing programs and control data that define various processing procedures, and executes various processes using these. The control unit 22 is realized by an integrated circuit such as an application specific integrated circuit (ASIC) or a field programmable gate array (FPGA). The control unit 22 is realized, for example, by executing a program stored in an internal storage device (not shown) using a RAM as a work area by a CPU (Central Processing Unit), an MPU (Micro Processing Unit), or the like. The

  The control unit 22 functions as various processing units by operating various programs. For example, the control unit 22 includes a reception unit 40, a generation unit 41, a learning unit 42, a search unit 43, a derivation unit 44, a specification unit 45, and a transmission unit 46.

  The accepting unit 40 accepts various data and various instructions from the management terminal 13. For example, the receiving unit 40 receives learning data and a learning instruction from the management terminal 13. When receiving the learning data from the management terminal 13, the receiving unit 40 registers the learning data 30 in the storage unit 21. In addition, when receiving a learning instruction from the management terminal 13, the receiving unit 40 instructs the generation unit 41 and the learning unit 42 to operate in the learning mode.

When the operation in the learning mode is instructed, the generation unit 41 performs a plurality of non-linear functions for calculating a new feature value by weighting a plurality of feature values by changing the number of feature values and combinations of feature values. Generate. For example, the generation unit 41 generates a plurality of functions for performing inner product calculation by weighting a plurality of feature values with independent parameters θ, by changing the number of feature values and combinations of feature values. FIG. 9 is a diagram for explaining generation of a function. The example of FIG. 9 indicates that feature values x_0 to x_i−1, x_i, and x_i + 1 to x_D and a plurality of feature values are obtained for a plurality of synonym candidates t−1 to t + 1, respectively. In this case, the generation unit 41 changes the number of feature values and the combination of the feature values, and weights the feature values x_0 to x_i−1, x_i, x_i + 1 to x_D with independent parameters θ, respectively, to function w_ {1} to w_ {g} to w_ {K} are generated. In the example of FIG. 9, θ ₁ to θ _g to θ _k are described for easy identification of the parameter θ of each function. This parameter θ is updated according to the degree to which the function contributes to the specification of synonyms.

  Here, the function loses the independence of individual feature values as the number of feature values to be weighted increases, and shows the same tendency. In addition, as the number of feature values to be weighted increases, the number of functions generated increases, and the processing load for calculating new feature values also increases. Therefore, the generation unit 41 generates a non-linear function by changing the number of feature values within a predetermined value or less. In the present embodiment, this predetermined value is set to 3, for example. This predetermined value may be designated by the administrator from the management terminal 13. That is, the generation unit 41 generates a non-linear function that calculates a new feature value by changing the combination of feature values and weighting the feature values within a range of three or less feature values. The synonym estimation device 10 according to the present embodiment identifies a synonym from a calculation result obtained by weighting a new feature value calculated from a plurality of feature values and a plurality of nonlinear functions. For example, a synonym is specified from a calculation result obtained by weighting a new feature value calculated from a plurality of feature values and a plurality of nonlinear functions with independent parameters w.

  When the learning unit 42 is instructed to perform an operation in the learning mode, the learning unit 42 can newly determine a character string based on the learning data 30 in order of significance, and is calculated from a plurality of feature values and a plurality of nonlinear functions. A weight for a feature value and a weight for a feature value obtained by combining each nonlinear function are learned. Specifically, the learning unit 42 first derives a feature value for each character string stored in the learning data 30. For example, the learning unit 42 derives a feature value of each feature shown in FIG. 2 as a real number for each character string that can be regarded as a synonym. In addition, when the feature value of each character string is stored in the learning data 30, it is not necessary to derive the feature value. Then, the learning unit 42 separates the feature values of the real numbers of the respective features in a predetermined feature order by tab delimiters in a separate record for each character string, and learns feature data for each character string of the learning data 30. 31 is generated. Then, the learning unit 42 uses the parameter θ and the parameter w as predetermined initial values, respectively, and calculates the feature values of the character strings of the feature data 31 for learning and the feature values assigned to the functions. The listwise learning is performed so that the result of weighting the feature value becomes closer to the score of each character string that can be regarded as a synonym of the learning data 30, and the weight value (individual parameter θ) for the feature value combined with each nonlinear function ) And the weight value (individual parameter w) for each feature value and the new feature value are updated. Then, the learning unit 42 sets, as the function data 32, each function for calculating a new feature value in which each parameter θ specified by learning is set, and each feature value indicating each feature value and a weight for the new feature value. Register in the storage unit 21. The function data 32 may store various parameters such as the parameter θ and the parameter w.

  Here, an example of function parameter update will be described. References (Zhe Cao, Tao Qin, Tie-Yan Liu, Ming-Feng Tsai, and Hang Li. Learning to rank: from pairwise approach to listwise approach. In Proceeding of the 24th international conference on Machine learning, ICML 07, P129-136, New York, USA, 2007. ACM) will explain Top k ListNet.

In Top k ListNet, the probability of obtaining the order of the top k documents from the n ⁽ⁱ⁾ synonym candidates obtained when the query q ⁽ⁱ⁾ is given is expressed by the following equation (1). Define as follows.

Here, f _w represents a ranking function for the document feature vector x _jt , and z ⁽ⁱ⁾ (f _w ) represents a list of scores of f _w for document permutations.

  In model learning, synonym learning is performed using the feature value and score of each character string that can be regarded as a synonym of the learning data 30, and the cross-entropy for the learning data 30 is minimized, so that the parameters of the ranking function Estimate w. The cross entropy is expressed as the following formula (2).

g _k represents a permutation composed of k documents, and g represents a subset thereof. The following formula (3) shows the parameter update formula in this case. Py ⁽ⁱ⁾ represents the probability of occurrence of a correct subset g of the document for the query q ⁽ⁱ⁾ , and Pz ⁽ⁱ⁾ (f _w ) represents the probability of the subset g obtained from the parameters of the learning model.

  When k is large, the above update formula is difficult to calculate, and therefore normally only the case of k = 1 is handled. The update formula when k = 1 is shown in the following formula (4).

In the reference literature, a simple linear neural network model in which the feature vector x _jt is weighted by the parameter w as shown in the following equation (5) is adopted for the ranking function f _w .

Here, since the reference literature employs a linear neural network model for the ranking function, only a linear relationship between the output z and the input x can be learned. Only a linear relationship between output z and input x can be learned. In the present embodiment, the function φ _t generated by the generation unit 41 has a nonlinear relationship between the output and the input. Therefore, it is extended so that a nonlinear relationship between the output and the input can be learned. As a result, the ranking function is changed as shown in Equation (6).

Φ (x _j ⁽ⁱ⁾ ) represents a vector composed of D functions φ _t . Each function φ _t is represented by a sigmoid function shown in the following formula (7), and each function has a parameter θ _t indicating an independent weight.

Parameter w so as to minimize the loss function of Top 1 ListNet making changes to the ranking function to update the theta _t. The update formulas are shown in the following formulas (8) to (10).

In the function φ _t , a combination of features included in the input x _j ⁽ⁱ⁾ is generated and given a nonlinear weight. Thereby, the input x _j ⁽ⁱ⁾ is mapped to an arbitrary D dimension, and nonlinear learning between the input and the output is performed.

  The learning unit 42 individually updates the independent parameter θ and parameter w using equations (8) and (9). As a result, the weight for the feature combination that contributes to predicting the correct answer is changed to be high, and the weight for the feature combination that does not contribute to the prediction of the correct answer is changed to be low. The parameter changing method is not limited to this. Various other methods can be applied. For example, the learning unit 42 obtains a calculation result obtained by weighting a new feature value calculated from a plurality of feature values and a plurality of nonlinear functions for each character string that can be regarded as a synonym stored in the learning data 30. If the closest synonym string can be identified at the highest level, the value of the operation result is increased, and if it cannot be specified, the feature value of the nonlinear function is weighted so that the value of the operation result is decreased. The weighting for the value and the new feature value may be changed. That is, the learning unit 42 may update the parameter θ and the parameter w depending on whether the synonym character string having the closest meaning can be specified.

  When a query is input from the client terminal 12, the search unit 43 searches a plurality of synonym candidates for the input query. This search method may be any as long as it can search for synonym candidates. In the present embodiment, for example, a plurality of synonym candidates are searched using label propagation.

  The deriving unit 44 derives the feature value as a real number for each of the searched synonym candidates. For example, the deriving unit 44 derives a real feature value of each feature shown in FIG. 2 for each of a plurality of searched synonym candidates. Then, the derivation unit 44 separates the feature values of the real numbers derived for each of the synonym candidates into separate records for each of the synonym candidates, and separates the feature values of the features in the predetermined feature order by tab separation. Feature data 31 is generated.

  When a query is input from the client terminal 12, the specifying unit 45 uses each function stored as the function data 32 in the storage unit 21, and uses a plurality of real number feature values in the feature data 31 to generate new real number feature values. Is calculated. Then, the specifying unit 45 performs a weighting operation on each feature value of the real number of the feature data 31 and the calculated feature value of the new real number with each independent parameter w stored as the function data 32 in the storage unit 21. Based on the calculation result, a synonym having a meaning close to that of the query input from the synonym candidates is identified. For example, the synonym candidates are ranked in descending order of the value of the operation result obtained by the weighting operation, and the largest synonym candidate is identified as a synonym having a close meaning. This identification method may be any as long as it can identify synonyms having similar meanings. For example, the identifying unit 45 may rank the synonym candidates in descending order of the feature values without distinguishing the feature values of the features.

  Here, each function generated by the generation unit 41 has a parameter θ so that the value of the calculation result is large when there is a high probability that the corresponding synonym character string can be specified at the highest level from a plurality of synonym candidates. When the parameter w is changed and the probability is low, the parameter θ and the parameter w are changed so that the value of the calculation result becomes small. Therefore, a function with a low probability that the synonym character string can be specified at the highest level has a smaller feature value value, and therefore has less influence when specifying a synonym that has a meaning close to the query.

  The transmission unit 46 transmits the synonym ranked highest to the client terminal 12 as a suggestion query. Thereby, on the client terminal 12, a synonym close to the query instructing the search is displayed as a suggestion query.

[3. Action (Operation of Synonym Estimation Device)]
Next, the effect | action of the synonym estimation apparatus 10 which concerns on a present Example is demonstrated. First, a flow of learning processing in which the synonym estimation device 10 according to the present embodiment learns a function effective for specifying synonyms will be described. FIG. 10 is a flowchart showing the procedure of the learning process. This learning process is executed at a predetermined timing, for example, when a learning instruction is received from the management terminal 13.

  As illustrated in FIG. 10, the generation unit 41 generates a plurality of nonlinear functions that calculate a new feature value by weighting a plurality of feature values by changing the number of feature values and the combination of the feature values (Step S <b> 10). ). The learning unit 42 derives a feature value for each character string stored in the learning data 30, and generates learning feature data 31 in which the feature values of the derived features are stored (step S11).

  Then, the learning unit 42 learns weights for a plurality of feature values and new feature values, and weights for feature values obtained by combining non-linear functions, which can determine the character strings of the learning data 30 in the order of meaning. (Step S12). The learning unit 42 stores, as function data 32, each function that calculates a new feature value in which the calculated parameter θ is set, and each feature value and a weight w for the new feature value as function data 32. It memorize | stores in the memory | storage part 21 (step S13), and complete | finishes a process.

  Next, the flow of the synonym specification process in which the synonym estimation device 10 according to the present embodiment specifies a synonym will be described. FIG. 11 is a flowchart showing the procedure of the synonym specifying process. This synonym specifying process is executed at a predetermined timing, for example, when a query is input from the client terminal 12.

  As shown in FIG. 11, the search unit 43 searches for a plurality of synonym candidates for the input query (step S20). The deriving unit 44 derives a feature value for each of the searched synonym candidates, and stores the feature data 31 storing the derived feature values of each feature in the storage unit 21 (step S21).

  The specifying unit 45 reads the function data 32 (step S22). Using the functions stored in the function data 32, the specifying unit 45 calculates new feature values from the feature values of the features in the feature data 31 (step S23). The specifying unit 45 performs a weighting operation on each feature value of the feature data 31 and the calculated new feature value using the parameter w stored as the function data 32, and ranks the synonym candidates in descending order of the value of the operation result. Then, a synonym having a meaning similar to the query input from the synonym candidates is specified (step S24).

  The transmission unit 46 transmits the synonym ranked in the highest rank as a suggestion query to the client terminal 12 (step S25), and ends the process.

[4. effect]
As described above, the synonym estimation device 10 stores the learning data 30 stored together with information indicating that the character string that can be regarded as a synonym is close in meaning in the storage unit 21 (corresponding to an example of a storage unit). Further, the synonym estimation device 10 generates a non-linear function for calculating a new feature value by weighting a plurality of feature values by using the generation unit 41 (corresponding to an example of a generation unit), the number of feature values, and the feature value. Generate multiple by changing the combination of. Then, the synonym estimation device 10 has a plurality of feature values and a plurality of non-linear functions that allow the learning unit 42 (corresponding to an example of learning means) to determine a character string in the order of meaning based on the learning data 30. The weighting for the new feature value respectively calculated from the above and the weighting for the feature value obtained by combining each nonlinear function is learned. Then, the synonym estimation device 10 uses the plurality of feature values of the synonym candidates for the plurality of synonym candidates searched from the query requested to be searched by the specifying unit 45 (corresponding to an example of the specifying unit). Then, a new feature value is calculated using each non-linear function that weights the feature value by weighting the combined feature value learned by the learning means, and the plurality of feature values and the calculated new feature value are Based on the calculation result calculated by the weighting learned by the learning means, a synonym having a meaning similar to the query is specified from a plurality of synonym candidates. As described above, the synonym estimation device 10 generates a plurality of nonlinear functions for calculating new feature values, weights the plurality of feature values and the new feature values, and weights the feature values obtained by combining the nonlinear functions. By learning, it is possible to obtain effective new feature values without bothering the user and identify synonyms taking into account new feature values, so synonyms close to the query can be more accurately Can be identified.

  In the synonym estimation device 10, the learning unit 42 changes the weight value of a nonlinear function having a high probability of being identified at the highest level, and changes the weight value of a nonlinear function having a low probability to be low. Thereby, the synonym estimation apparatus 10 can improve the specific precision of the character string of a synonym.

  Further, in the synonym estimation device 10, the generation unit 41 generates a non-linear function by changing the number of feature values within a predetermined value or less. As a result, the synonym estimation device 10 can prevent an unnecessarily long function for weighting feature values and suppress the generation of an arithmetic expression with a small influence of each feature value. It is possible to generate an arithmetic expression effective for specifying

[5. Others]
As described above, some of the embodiments of the present application have been described in detail with reference to the drawings. However, these are merely examples, and various modifications, including the aspects described in the disclosure section of the invention, based on the knowledge of those skilled in the art, It is possible to implement the present invention in other forms with improvements.

  For example, in the above embodiment, a case has been described in which a synonym candidate ranked highest is transmitted as a suggestion query, but the present invention is not limited to this. For example, an OR search may be performed on the synonyms ranked within a predetermined upper order together with a query.

  Further, the synonym estimation device 10 described above may be realized by a plurality of server computers, and depending on the function, an external platform or the like may be realized by calling an API (Application Programming Interface) or network computing. The configuration can be changed flexibly.

  For example, in the above-described embodiment, the synonym estimation apparatus 10 searches for synonym candidates for the input query and derives feature values for the synonym candidates, but is not limited thereto. is not. The synonym candidate search and the derivation of the feature value for the synonym candidate are performed by another server device such as a Web server, and the synonym estimation device 10 receives the feature data 31 from the other server device, The synonyms ranked higher may be returned to the other server device.

  Further, the “means” described in the claims can be read as “section (module, unit)” or “circuit”. For example, the search means can be read as a search unit or a search circuit.

DESCRIPTION OF SYMBOLS 10 Synonym estimation apparatus 21 Memory | storage part 22 Control part 30 Data for learning 31 Feature data 32 Function data 40 Reception part 41 Generation part 42 Learning part 43 Search part 44 Derivation part 45 Identification part 46 Transmission part

Claims (5)

Storage means for storing learning data stored together with information indicating the closeness of meaning of a character string that can be regarded as a synonym;
Generating means for generating a plurality of non-linear functions for calculating a new feature value by weighting a plurality of feature values by changing the number of feature values and combinations of feature values;
Weights for new feature values calculated from the plurality of feature values and a plurality of non-linear functions, respectively, based on the learning data, and a combination of the non-linear functions. Learning means for learning weights for values;
For each of a plurality of synonym candidates searched from a query for which a search is requested, each nonlinearity obtained by weighting the feature values by weighting the combined feature values learned by the learning means from the plurality of feature values of the synonym candidates A new feature value is calculated using a function, and the plurality of feature values and the calculated new feature value are calculated based on a calculation result obtained by weighting learned by the learning unit. A specifying means for specifying a synonym having a meaning close to the query;
A synonym estimation device comprising: 2. The synonym according to claim 1, wherein the learning unit changes a weight value of a nonlinear function having a high probability of being identified at the highest level and changes a weight value of a nonlinear function having a low probability to be low. Estimating device. The synonym estimation device according to claim 1, wherein the generation unit generates the nonlinear function by changing the number of feature values within a range of a predetermined value or less. A synonym estimation method executed by a computer,
A storage step of storing learning data stored together with information indicating the closeness of meaning of a character string that can be regarded as a synonym;
Generating a plurality of non-linear functions for calculating a new feature value by weighting a plurality of feature values by changing the number of feature values and the combination of feature values;
Based on the learning data, weights for new feature values calculated from the plurality of feature values and a plurality of nonlinear functions capable of determining character strings in the order of meaning, and feature values obtained by combining the nonlinear functions A learning process for learning weights for,
For each of a plurality of synonym candidates searched from a query for which a search is requested, each nonlinearity obtained by weighting the feature values by weighting the combined feature values learned in the learning step from the plurality of feature values of the synonym candidates A new feature value is calculated using a function, and the plurality of feature values and the calculated new feature value are calculated based on a calculation result obtained by weighting learned in the learning step. A specific step of identifying synonyms that are close in meaning to the query;
The synonym estimation method characterized by including. A storage procedure for storing learning data stored together with information indicating the closeness of meaning of a character string that can be regarded as a synonym;
A generation procedure for generating a plurality of non-linear functions for calculating a new feature value by weighting a plurality of feature values by changing the number of feature values and the combination of feature values,
Based on the learning data, weights for new feature values calculated from the plurality of feature values and a plurality of nonlinear functions capable of determining character strings in the order of meaning, and feature values obtained by combining the nonlinear functions A learning procedure for learning weights for,
For each of a plurality of synonym candidates searched from a query for which a search is requested, each nonlinearity obtained by weighting the feature values by weighting the combined feature values learned in the learning procedure from the plurality of feature values of the synonym candidates A new feature value is calculated using a function, and the plurality of feature values and the calculated new feature value are calculated based on a calculation result obtained by weighting learned in the learning procedure. A specific procedure for identifying synonyms that are close in meaning to the query;
Is executed by a computer.

Images