JP6971103B2 - Information processing equipment, information processing methods, and programs - Google Patents

Description

The present invention relates to an information processing apparatus, an information processing method, and a program.

In the field of web search, a method of providing information on various products or services related to this query is performed according to a keyword input by a user (hereinafter, "query"). In addition, in order to provide information on products or services intended by users with high accuracy, a search method using machine learning is being developed.

Special Table 2002-530948 Gazette

However, in the conventional technique, the relevance to the query is low, and the information of the product or service not intended by the user may be provided. In addition, the conventional method using machine learning uses a simple linear model, and cannot consider the combination of words contained in the query. Also, in the field of general machine learning, research on methods using encoder-decoder models in addition to linear models is underway, but it was not supposed to be applied to the field of web search (for example, patents). See Document 1).

The present invention has been made in consideration of such circumstances, and one of the objects of the present invention is to provide an information processing apparatus, an information processing method, and a program capable of improving search accuracy.

One aspect of the present invention is an analysis unit that analyzes a query used for a search and divides the query into words, a conversion unit that converts the words divided by the analysis unit into feature vectors, and training data. A first estimation unit that estimates the category associated with the query used in the search by learning the relationship between the feature vector of the word included in the query and the category associated with the query in the training data. It is an information processing device provided.

According to one aspect of the present invention, the search accuracy can be improved.

It is a figure which shows the use environment of the search server 10 of 1st Embodiment. It is a figure which shows an example of the tree structure of the category of 1st Embodiment. It is a figure which shows an example of the flow of the learning process of the search server 10 of 1st Embodiment. It is a figure which shows an example of the flow of the category estimation process of the search server 10 of 1st Embodiment. It is a figure which shows typically the category estimation process in the category estimation part 36 of 1st Embodiment. It is a figure which shows an example of the search result page of 1st Embodiment. It is a figure explaining the model used in the evaluation experiment. It is a figure which shows the result of the evaluation experiment. It is a figure which shows the comparison result of the agreement rate by the number of clicks of an evaluation experiment. It is a figure which shows an example of the category estimated using the model 1 of the evaluation experiment. It is a functional block diagram of the search server 11 of the 2nd Embodiment. It is a figure which shows an example of the search log data of 2nd Embodiment. It is a figure which shows an example of the flow of the category estimation process of the search server 11 of 2nd Embodiment. It is a functional block diagram of the search server 12 of the 3rd Embodiment.

Hereinafter, an information processing apparatus, an information processing method, and an embodiment of a program will be described with reference to the drawings.

<1. First Embodiment>
Hereinafter, the first embodiment of the present invention will be described. The information processing apparatus of the present embodiment divides the query used for the search into words, converts the divided words into feature vectors, and estimates the category associated with the query. The "category" in the present embodiment is information indicating a field to which the content (product or service) of the content provided as a search result for the query used for the search belongs. In the present embodiment, an example in which the information processing device is a search server that provides search results for the contents of a product or service will be described.

<1-1. System configuration>
FIG. 1 is a diagram showing a usage environment of the search server 10 (information processing apparatus) of the present embodiment. The search server 10 is connected to the terminal device T or the like via the network NW. The network NW includes, for example, WAN (Wide Area Network), LAN (Local Area Network), the Internet, a dedicated line, a wireless base station, a provider, and the like.

The terminal device T is operated by a user who uses the search service provided by the search server 10. The terminal device T is, for example, a personal computer, a mobile phone such as a smartphone, a tablet terminal, or a computer device such as a PDA (Personal Digital Assistant). In the terminal device T, a browser or an application program that operates based on a user's operation sends, for example, a query requesting information provision to the search server 10, and receives information associated with the query from the search server 10.

The search server 10 transmits the page information of the search result associated with the query input from the terminal device T to the terminal device T. The search server 10 generates page information of the search result by using the search data in which the query is associated with the reference information for referencing the page of the site associated with the query. The reference information includes, for example, a URL (Uniform Resource Locator).

The search server 10 includes, for example, a communication unit 30, a query analysis unit 32 (analysis unit), a query conversion unit 34 (conversion unit), a category estimation unit 36 (first estimation unit, learning unit), and a search result generation. A unit 38 (generation unit) and a storage unit 40 are provided. The communication unit 30 communicates with the terminal device T or the like via the network NW. The communication unit 30 receives a query from the terminal device T via the network NW, and transmits the page information of the search result to the terminal device T. The communication unit 30 includes, for example, a communication interface such as a NIC.

The query analysis unit 32 analyzes the query input from the terminal device T and divides it into word levels. The query analysis unit 32 analyzes the query by, for example, morphological analysis, and divides the query into at least one word.

The query conversion unit 34 converts each of the words divided by the query analysis unit 32 into a feature vector. The query conversion unit 34 converts each of the divided words into a feature vector by using, for example, a tool (program) called word2vec. Word2vec is a tool that uses a neural network, and converts the text information contained in the input corpus into a feature vector that indicates the characteristics of the text information and outputs it. In this embodiment, it is assumed that the distributed representation of the query has been learned in advance in word2vec.

The category estimation unit 36 estimates the category associated with at least one feature vector input from the query conversion unit 34. The category estimation unit 36 estimates a category using, for example, an encoder / decoder model based on a recurrent neural network (RNN). In the following, an example in which the category estimation unit 36 uses the encoder / decoder model will be described.

The category estimation unit 36 includes, for example, an encoder 42 and a decoder 44. The encoder 42 and the decoder 44 use a set of a "word feature vector" included in the query used for the search and a "category" selected by the user on the search result page provided in response to the query as learning data. Perform learning processing.

The encoder 42 performs an encoding process on at least one feature vector input from the query conversion unit 34 by using a model trained using the above-mentioned training data (hereinafter, “encoding model””. When two or more feature vectors are input from the query conversion unit 34, encodes the first feature vector, then encodes the first feature vector and the second feature vector. The encoder 42 inputs the encoding processing result obtained by repeating the encoding processing to the decoder 44.

The decoder 44 sets a category to be associated with the query word based on the encoding processing result input from the encoder 42 using the model trained using the above-mentioned training data (hereinafter, “decode model””. The category has a tree structure, for example, and defines the fields to which the goods or services belong in a hierarchical manner. FIG. 2 is a diagram showing an example of the tree structure of the categories. In the example shown in the above, "men's fashion", "ladies' fashion" and the like are defined as the first layer H1, and "men's shoes" are defined as the second layer H2 below "men's fashion (first layer H1)". , "Men's back", etc. In addition, "sneakers", "business shoes", "sandals", "boots" are defined as the third layer H3 below "men's shoes (second layer H2)". Etc. are defined.

The decoder 44 learns the relationship between the encoding processing result (multidimensional vector) input from the encoder 42 and the tree structure of the category as shown in FIG. The decoder 44 estimates the category to be associated with the word of the query based on the encoding processing result, and outputs the category in order from the upper hierarchy in the tree structure of the category.

The search result generation unit 38 generates a search result page which is a search result based on the category estimated by the category estimation unit 36. The search result generation unit 38 transmits the generated search result page to the terminal device T via the communication unit 30.

Each functional unit of the search server 10 is realized, for example, by executing a program (software) by a hardware processor such as a CPU (Central Processing Unit) in a computer. In addition, some or all of these components are hardware (circuits) such as LSI (Large Scale Integration), ASIC (Application Specific Integrated Circuit), FPGA (Field-Programmable Gate Array), and GPU (Graphics Processing Unit). It may be realized by the part; including circuitry), or it may be realized by the cooperation of software and hardware.

The storage unit 40 stores search data D1 in which a query is associated with reference information for referencing a page of a site associated with the query, category data D2 indicating a category tree structure, and the like. The storage unit 40 is realized by, for example, a RAM (Random Access Memory), a ROM (Read Only Memory), an HDD (Hard Disk Drive), a flash memory, or a hybrid storage device in which a plurality of these are combined. Further, a part or all of the storage unit 40 may be an external device such as NAS or an external storage server that can be accessed by the search server 10.

<1-2. Learning process>
Hereinafter, the learning process of the search server 10 will be described. FIG. 3 is a diagram showing an example of the flow of learning processing of the search server 10. The category estimation unit 36 learns the relationship between the feature vector of the word included in the query and the category associated with the query.

First, the learning data is input to the category estimation unit 36 of the search server 10 (S101). As the learning data, for example, the log data when the search service is used is used. As the learning data, for example, a set of a "word feature vector" included in the query used for the search and a "category" selected by the user on the search result page provided in response to the query is used.

Next, when the encoder 42 of the category estimation unit 36 encodes the "feature vector of the word included in the query" (S103) and the decoder 44 decodes the encoding processing result by the encoder 42 (S105), this query and The parameters in this encoder model and decoder model are adjusted so that a set of "categories" is output (S107).

For example, as a result of this decoding process, the decoder 44 calculates a score indicating the degree of association with the word of the query for each category defined as the first layer. The decoder 44 adjusts the parameters in the decoder model so that the score of the category paired with the query is high in the training data. The category estimation unit 36 may adjust the parameters by, for example, the back propagation method. The same processing is performed for the second and lower layers. This completes the processing of this flowchart.

In addition, in the log data, among the data of multiple categories existing for a certain query, the data of the category with a small number of clicks (data in which the number of clicks (selected number of times) is equal to or less than a predetermined threshold value) is excluded and excluded. The later log data may be used as training data. This makes it possible to improve the accuracy of the encoder model and the decoder model.

<1-3. Category estimation process>
Hereinafter, the category estimation process of the search server 10 will be described. FIG. 4 is a diagram showing an example of the flow of the category estimation process of the search server 10.

First, the query analysis unit 32 acquires the query input from the terminal device T via the communication unit 30 (S201). Next, the query analysis unit 32 analyzes the query input from the terminal device T and divides it into word levels (S203). The query analysis unit 32 analyzes the query by, for example, morphological analysis, and divides the query into word levels.

Next, the query conversion unit 34 converts each of the words divided by the query analysis unit 32 into a feature vector (S205). The query conversion unit 34 converts each of the divided words into a feature vector by using, for example, word2vec.

Next, the encoder 42 performs encoding processing on the feature vector of the first word among the feature vectors converted by the query conversion unit 34 (S207). Next, the encoder 42 determines whether or not the encoding process for the feature vectors of all the words is completed (S209).

When the encoder 42 determines that the encoding processing for the feature vectors of all the words has not been completed, the feature vector of the second word and the feature vector of the first word among the feature vectors converted by the query conversion unit 34 Encoding processing is performed again using the encoding processing result for the feature vector (S207). Next, the encoder 42 determines whether or not the encoding process for the feature vectors of all the words is completed (S209). Hereinafter, the same processing is repeated until the encoding processing for the feature vectors of all words is completed. When the encoder 42 determines that the encoding processing for the feature vectors of all words is completed, the encoder 42 inputs the encoding processing result obtained as a result of repeating the above-mentioned encoding processing to the decoder 44.

Next, the decoder 44 performs decoding processing on the encoding processing result input from the encoder 42 (S211). As a result of this decoding process, the decoder 44 calculates a score indicating the degree of association with the word of the query for each category defined as the first layer. For example, the category with the highest score is estimated as the category associated with the query.

Next, the decoder 44 determines whether or not the estimation of the lowest layer category is completed (S213). When the decoder 44 determines that the estimation of the lowermost category has not been completed, the decoder 44 performs the decoding process again using the estimated category information and the decoding process result (S211). Next, the encoder 42 determines whether or not the estimation of the category of the lowest layer is completed (S213). Hereinafter, the same process is repeated until the estimation of the lowest category is completed. When the decoder 44 determines that the estimation of the lowest layer category is completed, the decoder 44 inputs the information of the estimated category to the search result generation unit 38.

FIG. 5 is a diagram schematically showing the category estimation process in the category estimation unit 36. In the example shown in FIG. 5, the query analysis unit 32 divides the query "men's sneaker company A" input from the terminal device T into three words "men's", "sneaker", and "company A", and queries. An example is shown in which the conversion unit 34 converts each of these three words into a feature vector and inputs it to the category estimation unit 36.

First, the encoder 42 performs encoding processing on the feature vector of the first word "men's", and obtains the first encoding processing result R1. Next, the encoder 42 performs an encoding process using the feature vector of the second word "sneaker" and the first encoding processing result R1 to obtain the second encoding processing result R2. Next, the encoder 42 performs an encoding process using the feature vector of the third word “Company A” and the second encoding processing result R2, and obtains the third encoding processing result R3. The encoder 42 inputs the third encoding processing result R3 to the decoder 44.

Next, the decoder 44 performs decoding processing on the third encoding processing result R3 input from the encoder 42. Here, the decoder 44 calculates a score indicating the degree of association with the three words "men's", "sneakers", and "company A" for each category of the highest layer (first layer H1). The decoder 44 associates the category with the highest score among the scores calculated for each category with the three words "men's", "sneakers", and "company A" (in this example, "" Men's fashion ").

Next, the decoder 44 performs decoding processing using the decoding processing result (first decoding processing result R4) for the third encoding processing result R3 and the data indicating the estimated category “men's fashion”. Here, the decoder 44 has three words, "men's", "sneakers", and "company A", for each category of the second layer H2 located below the estimated category "men's fashion (first layer)". Calculate a score that indicates the degree of association with. The decoder 44 associates the category with the highest score among the scores calculated for each category with the three words "men's", "sneakers", and "company A" (in this example, "" Men's shoes ").

Next, the decoder 44 performs decoding processing using the decoding processing result (second decoding processing result R5) for the first decoding processing result R4 and the data indicating the estimated category “men's shoes”. Here, the decoder 44 has the three words "men's", "sneakers", and "company A" for each category of the third layer H3 located below the estimated category "men's shoes (second layer)". Calculate a score that indicates the degree of association with. The decoder 44 associates the category with the highest score among the scores calculated for each category with the three words "men's", "sneakers", and "company A" (in this example, "" Sneakers ").

By the above processing shown in FIG. 5, the categories associated with the three words "men's", "sneakers", and "company A" are "men's fashion (first layer)" and "men's shoes (second layer)". , And the estimation results of the category "sneakers (third layer)" are obtained.

Next, the search result generation unit 38 generates a search result page, which is a search result, based on the estimation result of the category input from the decoder 44, and transmits it to the terminal device T (S215). FIG. 6 is a diagram showing an example of the search result page P1. In the example shown in FIG. 6, the search result generation unit 38 generates the search result page P1 in which the estimated third layer category “sneakers” are arranged at the most visible position in the page.

In the example shown in FIG. 6, in addition to the estimated category "sneakers", the other category "boots" under the estimated second layer category "men's shoes (second layer)". , "Sandals", "business shoes", etc. are displayed in descending order of score, for example. The search result generation unit 38 transmits the generated search result page P1 to the terminal device T via the communication unit 30. The user of the terminal device T can acquire the search result according to the user's intention by operating the terminal device T and selecting one category included in the search result page P1. This completes the processing of this flowchart.

In the example shown in FIG. 6, an example of transmitting the search result page P1 that causes the user to select the estimated category to the terminal device T has been described. However, the search result generation unit 38 may transmit the page of the product or service related to the estimated category to the terminal device T instead of the search result page P1 that causes the user to select the estimated category. ..

<1-3. Evaluation experiment>
In order to evaluate the category estimation accuracy of the search server 10 in this embodiment, an evaluation experiment using three models as shown in FIG. 7 was performed. Model 1 is an encoder / decoder model used in the search server 10 of the present embodiment. Model 2 is a linear model using a conventional linear classifier. Model 3 is a (results-based) model based on the log of the prior art. In this model 3, the category with the most clicks is selected in the search log. In this evaluation experiment, among the categories clicked by the user on the search result page provided according to the actual use in the search, the category with the largest number of clicks was defined as the "correct answer category". In this evaluation experiment, the query actually used for the search is input to the search server 10 that performs the category estimation process using each model, and the match rate between the category estimated by the search server 10 and the correct answer category. Was calculated.

FIG. 8 is a diagram showing the results of the above-mentioned evaluation experiment. As shown in FIG. 8, in all of the first layer category, the second layer category, and the third layer category, the matching rate of the estimation result using the model 3 (log model) is the highest, followed by. , The matching rate of the estimation result using the model 1 (encoder decoder model) was high, and the matching rate of the estimation result using the model 2 (linear model) was the lowest.

Further, FIG. 9 is a diagram showing a comparison result of the agreement rate for each number of clicks in the evaluation experiment. As shown in FIG. 9, for any number of clicks, the matching rate of the estimation result using the model 3 is the highest, then the matching rate of the estimation result using the model 1 is the highest, and the estimation using the model 2 is performed. The concordance rate of the results was the lowest. It was found that even when the number of clicks was small (even in the so-called tail query), the matching rate of the estimation results using Model 3 was the highest.

In model 1 used in this evaluation experiment, since the click frequency is not taken into consideration, the category with a large number of clicks and the category with a small number of clicks are processed in the same way, so that they are vulnerable to noise contained in the training data. is assumed.

Further, although the categories estimated using the model 1 were different from the correct answer categories in the search log, there were some categories that could be judged to have been correctly estimated. FIG. 10 is a diagram showing an example of a category estimated using the model 1. As shown in FIG. 10, for example, for the query "newlywed gift", the category estimated using model 1 was the category related to kitchen utensils, while the correct answer category was related to the game. It was a category. This is due to the ambiguity of the query, and it can be considered that the estimation result of such model 1 may be judged to be correct.

Also, for example, in response to the query "Enter a large wallet passbook", the category estimated using Model 1 was a category related to women's fashion, while the correct answer category was a category related to men's fashion. there were. This is due to the fact that the categories related to the products are duplicated, and it can be considered that the estimation result of the model 1 may be judged to be correct.

Also, for example, for the query "mobile phone cute strap", the category estimated using model 1 was the category related to smartphones, while the correct answer category was the category related to ladies fashion. rice field. This is due to the difficulty of category selection for the query, and it can be considered that such an estimation result of Model 1 may be judged to be correct.

In addition, as an example in which the estimation process of model 1 is not determined to be the correct answer, the morphological analysis for the query by the query analysis unit 32 is not appropriate, or a specific word included in the query has a great influence on the category estimation. In some cases, the relationship with other words was not properly evaluated.

Regarding the category estimated using model 1 as described above, although it is different from the "correct answer category" in the search log, considering that it can be judged that the correct estimation process is performed, model 1 The matching rate of the categories in the first layer was 0.9238 (92.38 [%]). Considering this point, it can be determined that the accuracy of category estimation using the encoder / decoder model in the search server 10 of the present embodiment is high.

In addition, the coverage of the search log for the query in the estimation process using the model 3 was 0.3671 as a whole. Therefore, when the model 3 is used, it may not be possible to estimate an appropriate category unless the query exists in the search log. On the other hand, in the estimation using the model 1, even if the query does not exist in the search log, the category can be estimated appropriately. For example, even if the query "Men's, sneakers, company B" is not learned, the query "Men's, sneakers, company A" is learned, and "Company B" and "Company B" are learned. If it is learned that "Company A" is a close word in the vector expression, it is possible to appropriately estimate the category for "Men's, sneakers, Company B".

According to the search server 10 of the present embodiment described above, the category associated with the query can be appropriately estimated, and the search accuracy can be improved. Further, according to the search server 10 of the present embodiment, it is possible to appropriately estimate the category even when an unknown query that has not been used for the search in the past is input.

It should be noted that a determination unit for determining the break of a set of sessions of access from the user who operates the terminal device T to the search server 10 based on the encoding processing result of the encoder 42 may be further provided. For example, when the position of the vector indicated by the encoding processing result on the vector coordinate space changes significantly, it may be determined that the session of access from the user has been switched.

<2. 2nd Embodiment>
Hereinafter, a second embodiment of the present invention will be described. The information processing apparatus (search server) of the present embodiment is different from the first embodiment in that it performs both the category estimation process based on the encoder decoder model and the category estimation process based on the log data. Therefore, for the configuration and the like, the drawings and related descriptions described in the first embodiment will be referred to, and detailed description will be omitted.

<2-1. System configuration>
FIG. 11 is a functional block diagram of the search server 11 of the present embodiment. The search server 11 further includes, for example, a query determination unit 46 and a second category estimation unit 48 (second estimation unit) in addition to the components of the search server 10 of the first embodiment.

The query determination unit 46 refers to the search log stored in the storage unit 40, and determines whether or not the query input from the terminal device T exists in the past search log. In the search server 11, when the query determination unit 46 determines that the query input from the terminal device T does not exist in the past search log, the search server 11 performs category estimation processing based on the encoder decoder model similar to the first embodiment. .. On the other hand, when the query determination unit 46 determines that the query input from the terminal device T exists in the past search log, the search server 11 performs category estimation processing based on the log data.

The second category estimation unit 48, for example, refers to the search log data D3 stored in the storage unit 40 and estimates the category. The second category estimation unit 48 selects, for example, the category with the largest number of clicks in the search log data D3 stored in the storage unit 40. FIG. 12 is a diagram showing an example of the search log data D3. In FIG. 12, when the query is "Ladies B company", category 1 of the first layer is "ladies fashion", category 2 of the second layer is "wallet / accessory", and category 3 of the third layer is. The number of clicks (256999) in the category of "wallet" is the highest. Therefore, when the query is "Ladies'B company", the category 1 of the first layer is "Ladies' fashion" and the category 2 of the second layer is the category associated with this query in the second category estimation unit 48. Is "wallet / accessory", and category 3 of the third layer estimates "wallet".

<2-2. Category estimation process>
Hereinafter, the category estimation process of the search server 11 will be described. FIG. 13 is a diagram showing an example of the flow of the category estimation process of the search server 11.

First, the query analysis unit 32 acquires the query input from the terminal device T via the communication unit 30 (S301). Next, the query determination unit 46 refers to the search log stored in the storage unit 40, and determines whether or not the query input from the terminal device T exists in the past search log (S303).

When the query determination unit 46 determines that the query input from the terminal device T exists in the past search log, the second category estimation unit 48 refers to the search log data D3 stored in the storage unit 40. , Estimate the category (S305). The second category estimation unit 48 inputs the estimated category information into the search result generation unit 38.

On the other hand, when the query determination unit 46 determines that the query input from the terminal device T does not exist in the past search log, the query analysis unit 32 analyzes the query input from the terminal device T to the word level. Divide (S307). Next, the query conversion unit 34 converts each of the words divided by the query analysis unit 32 into a feature vector (S309).

Next, the encoder 42 performs encoding processing on the feature vector of the first word among the feature vectors converted by the query conversion unit 34 (S311). Next, the encoder 42 determines whether or not the encoding process for the feature vectors of all the words is completed (S313).

When the encoder 42 determines that the encoding processing for the feature vectors of all the words has not been completed, the feature vector of the second word and the feature vector of the first word among the feature vectors converted by the query conversion unit 34 Encoding processing is performed again using the encoding processing result for the feature vector (S311). Next, the encoder 42 determines whether or not the encoding process for the feature vectors of all the words is completed (S313). Hereinafter, the same processing is repeated until the encoding processing for the feature vectors of all words is completed. When the encoder 42 determines that the encoding processing for the feature vectors of all words is completed, the encoder 42 inputs the encoding processing result obtained as a result of repeating the above-mentioned encoding processing to the decoder 44.

Next, the decoder 44 performs decoding processing on the encoding processing result input from the encoder 42 (S315). Next, the decoder 44 determines whether or not the estimation of the lowest layer category is completed (S317). When the decoder 44 determines that the estimation of the lowermost category has not been completed, the decoder 44 performs the decoding process again using the estimated category information and the decoding process result (S315). Next, the encoder 42 determines whether or not the estimation of the category of the lowest layer is completed (S317). Hereinafter, the same process is repeated until the estimation of the lowest category is completed. When the decoder 44 determines that the estimation of the lowest layer category is completed, the decoder 44 inputs the information of the estimated category to the search result generation unit 38.

Next, the search result generation unit 38 generates a search result page which is a search result based on the estimation result of the category input from the decoder 44 or the second category estimation unit 48, and transmits it to the terminal device T (S319). ). This completes the processing of this flowchart.

According to the search server 11 of the present embodiment described above, the category associated with the query can be appropriately estimated, and the search accuracy can be improved. Further, according to the search server 11 of the present embodiment, when the query input from the terminal device T does not exist in the past search log, the category estimation process based on the encoder decoder model is performed and the query is input from the terminal device T. If the query exists in the past search log, the category estimation accuracy can be further improved by performing the category estimation process based on the log data.

<3. Third Embodiment>
Hereinafter, a third embodiment of the present invention will be described. The information processing apparatus (search server) of the present embodiment is different from the first embodiment in that it includes one encoder and a plurality of estimation units (decoders). Therefore, for the configuration and the like, the drawings and related descriptions described in the first embodiment will be referred to, and detailed description will be omitted.

<3-1. System configuration>
FIG. 14 is a functional block diagram of the search server 12 of the present embodiment. The search server 12 replaces the category estimation unit 36 of the search server 10 of the first embodiment with, for example, one encoder 50, three estimation units (category estimation unit 52, price range estimation unit 54, brand estimation unit 56). ) And.

The encoder 50 uses the learning data and uses an encoding model to perform encoding processing on at least one feature vector input from the query conversion unit 34. When two or more feature vectors are input from the query conversion unit 34, the encoder 50 performs encoding processing on the first feature vector, and then encodes the first feature vector and 2 Encoding processing is performed using the second feature vector. The encoder 42 inputs the encoding processing result obtained by repeating the encoding processing to the category estimation unit 52, the price range estimation unit 54, and the brand estimation unit 56.

The category estimation unit 52 performs decoding processing on the encoding processing result input from the encoder 50 by using the decoding model trained using the training data, and estimates the category to be associated with the query word. And input to the search result generation unit 38.

The price range estimation unit 54 performs decoding processing on the encoding processing result input from the encoder 50 using the decoding model trained using the training data, and the product or service to be associated with the word of the query. The price range of is estimated and input to the search result generation unit 38. The training data of the price range estimation unit 54 includes, for example, a "word feature vector" included in the query used for the search and a "price of the product or service" selected by the user on the search result page provided in response to the query. Use the pair with.

The brand estimation unit 56 performs decoding processing on the encoding processing result input from the encoder 50 using the decoding model trained using the training data, and performs decoding processing on the product or service to be associated with the word of the query. The brand is estimated and input to the search result generation unit 38. The learning data of the brand estimation unit 56 includes, for example, a "word feature vector" included in the query used in the search and a "brand of goods or services" selected by the user on the search result page provided in response to the query. Use the pair with.

The search result generation unit 38 is based on the category estimation result input from the category estimation unit 52, the price range estimation result input from the price range estimation unit 54, and the brand estimation result input from the brand estimation unit 56. Then, a search result page, which is a search result, is generated and transmitted to the terminal device T.

According to the search server 12 of the present embodiment described above, the category associated with the query can be appropriately estimated, and the search accuracy can be improved. Further, according to the search server 12 of the present embodiment, the category estimation result input from the category estimation unit 52, the price range estimation result input from the price range estimation unit 54, and the brand estimation unit 56 are input. Since the search result page which is the search result is generated based on the estimation result of the brand, it is possible to further provide the search result page according to the user's intention.

In addition to or in place of the category estimation unit 52, the price range estimation unit 54, and the brand estimation unit 56, an estimation unit for estimating other features may be provided.

According to the embodiment described above, an analysis unit that analyzes the query used for the search and divides the query into words, a conversion unit that converts the words divided by the analysis unit into feature vectors, and a conversion unit. First estimation to estimate the category associated with the query used in the search by learning the relationship between the feature vector of the word included in the query in the training data and the category associated with the query in the training data. By providing a unit, the search accuracy can be improved.

Although the embodiments for carrying out the present invention have been described above using the embodiments, the present invention is not limited to these embodiments, and various modifications and substitutions are made without departing from the gist of the present invention. Can be added.

10, 11, 12 ... Search server (information processing device)
30 ... Communication unit 32 ... Query analysis unit (analysis unit)
34 ... Query conversion unit (conversion unit)
36 ... Category estimation unit (1st estimation unit, learning unit)
38 ... Search result generation unit (generation unit)
40 ... Storage unit 42 ... Encoder 44 ... Decoder 46 ... Query judgment unit (judgment unit)
48 ... Second category estimation unit (second estimation unit)
50 ... Encoder 52 ... Category estimation unit 54 ... Price range estimation unit 56 ... Brand estimation unit

Claims (12)

An analysis unit that analyzes the query used for the search and divides the query into words.
A conversion unit that converts words divided by the analysis unit into feature vectors, and a conversion unit.
The first estimation that estimates the category associated with the query used in the search by learning the relationship between the feature vector of the word included in the query in the training data and the category associated with the query in the training data. With a department ,
The first estimation unit is
An encoder that performs encoding processing on the feature vector of the word,
A decoder that performs decoding processing on the processing result of the encoder and outputs a category associated with the query used in the search.
Equipped with
The decoder performs a first decoding process on the processing result of the encoder to output a first category, and performs a second decoding process on the result of the first decoding process to perform a second decoding process. Output category,
Information processing device. The encoder performs a first encoding process on the first feature vector among the feature vectors of the plurality of words, and then uses the result of the first encoding process and the second feature vector. Encoder processing,
The information processing apparatus according to claim 1. The category has a tree structure and has a tree structure.
The first category is the category defined at the highest level in the tree structure.
The second category is a category defined below the first category.
The information processing apparatus according to claim 1 or 2. It further comprises a generator that generates a search result page for the query used in the search based on the category estimated by the first estimate.
The information processing apparatus according to any one of claims 1 to 3. The analysis unit analyzes the query used in the search by morphological analysis and divides the query into words.
The information processing apparatus according to any one of claims 1 to 4. Further provided with a learning unit that learns the relationship between the feature vector of the query word used in the search and the category selected by the user on the search result page for the query used in the search.
The learning unit excludes categories from the learning data in which the number of times selected by the user is equal to or less than a predetermined threshold value.
The information processing apparatus according to any one of claims 1 to 5. Further comprising a second estimation unit that estimates the category associated with the query used in the search based on the log of the category selected by the user on the search result page for the query used in the search.
The information processing apparatus according to any one of claims 1 to 6. Further, a determination unit for determining whether or not the query used for the search exists in the log is provided.
When the determination unit determines that the query used in the search exists in the log, the second estimation unit estimates a category related to the query used in the search.
When the determination unit determines that the query used for the search does not exist in the log, the first estimation unit estimates a category related to the query used for the search.
The information processing apparatus according to claim 7. The first estimation unit is
An encoder that performs encoding processing on the feature vector of the word,
A plurality of decoders that perform decoding processing on the processing result of the encoder are provided.
The information processing apparatus according to claim 1. An encoder that performs encoding processing on the feature vector of the word,
When the position of the vector indicated by the processing result of the encoder on the vector coordinate space changes significantly, it is determined that the session of access from the user has been switched, so that a session of a set of access from the user to the information processing apparatus is performed. Judgment unit that determines the break of
Further prepare,
The information processing apparatus according to claim 1. The computer
The query used in the search is parsed, the query is split into words, and
The divided words are converted into feature vectors and
By learning the relationship between the feature vector of the word included in the query in the training data and the category associated with the query in the training data, the category associated with the query used in the search is estimated .
In the estimation of the above category,
Encoding processing is performed on the feature vector of the word.
Decoding processing is performed on the result of the encoding processing, and the category associated with the query used in the search is output.
In the decoding process, the first decoding process is performed on the result of the encoding process to output the first category, and the second decoding process is performed on the result of the first decoding process to perform the second decoding process. Output the category of
Information processing method. On the computer
The query used for the search is analyzed, and the query is divided into words.
The divided words are converted into feature vectors and
By learning the relationship between the feature vector of the word included in the query in the training data and the category associated with the query in the training data, the category associated with the query used in the search is estimated .
In the estimation of the above category,
Encoding processing is executed for the feature vector of the word,
The decoding process is executed for the result of the encoding process, and the category associated with the query used in the search is output.
In the decoding process, the first decoding process is executed on the result of the encoding process to output the first category, and the second decoding process is executed on the result of the first decoding process. Output the second category,
program.

Images