JP2020071678A - Information processing device, control method, and program - Google Patents

Abstract

To provide a technique capable of enhancing the accuracy improving effect by effectively using learning data and also shortening the learning time, in order learning in information search.SOLUTION: An information processing device includes: search means for searching for a search object document with a search query text; and storage means for storing the search object document and a learning search query text associated with the search object document. The device includes: creation means for creating additional text information for the search object document using the learning search query text associated with the search object document; and learning means for performing order learning using a set of the learning search query text and document data including additional text information for the search object document.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a document search technique for presenting a document considered to be appropriate for a specified search condition from a document group to be searched.

  Conventionally, in order to present an appropriate search result to a user, the relationship between search conditions and terms (character strings cut out by a certain standard such as morphological analysis and N-Gram) included in each document included in a document group is related. There is a technique for calculating the statistical value. These techniques are referred to as similarity search and the like (hereinafter, this technique will be collectively referred to as similarity search in the description of the present invention, and will be distinguished from the search by order learning described later in the present invention).

  In addition, when the learning data and the documents to be searched are similar, a feature quantity is modeled by machine learning, and when a new search condition is specified, ranking adjustment is performed based on the learning model to perform similar search. There is a rank learning technique that improves the accuracy of.

  A large amount of learning data is required for rank learning, but it is difficult to collect learning data. It is possible to collect learning data from the user's search log after the similar search is started as a system, but evaluation of the search results may impose a load on the user, and it is not possible to collect a sufficient amount of logs. could not say it all. Before starting operation, it will be limited to learning data created by the developer for testing.

  Patent Document 1 is directed to a technique for finding a question (question sentence of learning data) that is most similar to an inquiry from a user in response to a prepared answer (so-called FAQ document group) and returning a corresponding answer. Thus, even if the number of question sentences is small, a technology for improving topic estimation accuracy is provided.

  Specifically, by replacing a word appearing in the question sentence of the learning data with a word in the corresponding answer, the question sentence of the learning data is expanded, that is, the number of learning data items is increased. In order to exclude unnatural question sentences from the expanded question sentences, the probabilistic language model is used to calculate the existence probability of the question sentence, and only when the existence probability exceeds a certain threshold, it is used as learning data.

JP, 2017-37588, A

  However, in the technique of Patent Document 1, it is determined using a probabilistic language model whether or not the expanded question sentence is appropriate, but the replaced word is included in the answer prepared in advance. It is possible that technical terms or term specific to an organization are used. In that case, there are cases where the probabilistic language model is insufficient and the question sentences are not appropriately expanded.

Further, even if the expansion of the question text is properly processed, the technical user or the organization-specific term may not be input by the actual user. Further, the technique itself has the purpose of amplifying the question sentence, and as a result, the time required for the learning process may increase exponentially.
It is an object of the present invention to provide a technique capable of enhancing the accuracy improvement effect and effectively shortening the learning time by effectively using the learning data in order learning in information retrieval.

  The present invention is an information processing apparatus comprising: a search unit that searches a search target document by a search query text; and a storage unit that stores the search target document and a learning search query text associated with the search target document. Then, using the learning search query text associated with the search target document, creating means for creating additional text information for the search target document, the learning search query text, and the addition to the search target document And a learning means for performing order learning using a set with document data including text information.

  According to the present invention, an object of the present invention is to provide a technology capable of enhancing the effect of improving accuracy by effectively using learning data in order learning in information retrieval and shortening the learning time. Become.

It is a figure showing an example of functional composition concerning an embodiment of the present invention. It is a block diagram which shows an example of the hardware constitutions applicable to the information processing apparatus 100 which concerns on embodiment of this invention. 3 is an example of a document to be a similarity search target according to the embodiment of the present invention. 3 is an example of learning data according to the embodiment of the present invention. It is an example of a generated feature vector according to the embodiment of the present invention. It is an example of the characteristic word extracted from the learning data concerning the embodiment of the present invention. It is an example of a generated feature vector according to the embodiment of the present invention. It is an example of a graph showing a distribution of the number of learning data according to the embodiment of the present invention. It is an example of a flowchart illustrating a process at the time of learning according to the embodiment of the present invention. 6 is an example of a flowchart illustrating a process of similarity search / reranking based on a learning result according to the embodiment of the present invention. 3 is an example of a setting item according to the embodiment of the present invention. It is a figure which shows an example of the storage method of the learning language information which concerns on embodiment of this invention.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

  In the present invention, the search result of a conventional document is re-designated by machine learning using machine learning. This is called order learning. In particular, in the present invention, for convenience of explanation, the process of determining the learning model in advance is “generation of the learning model”, and the search results based on the search conditions of the user or the like are actually designated again by using the generated learning model. The process will be referred to as "reranking".

  FIG. 1 is a diagram showing an example of a functional configuration according to an embodiment of the present invention. This functional configuration can be mainly divided into the following three types. First, the conventional similarity search is used, and 131 is the portion related to the conventional similarity search. Further, the portions 101 to 104 generate a learning model based on the learning data. The portions 111 to 114 perform similarity search and re-ranking based on the generated learning model.

  The learning data preprocessing unit 101 selects learning data to be actually learned from the learning data (user logs such as search conditions and selection of correct answers) stored in the learning data storage unit 121, and linguistically from the learning data group. It is a functional unit that extracts features. Criteria for performing the selection and extraction are stored in the setting storage unit 122. The result of the preprocessing is stored in the learning language information storage unit 123.

  The information search unit 131 for the query described in the learning data is called, and the search target document storage unit 124 that stores the document group is searched. Information retrieval is assumed to be not a highly accurate search engine using “rank learning” described in the present invention, but other rank learning, or rank learning itself of the present invention, or any method It may be information retrieval. Anyway, it is only necessary that the search target document storage unit 124 can appropriately search for a document desired by the user.

  In the learning feature vector generation unit 102, focusing on one of the learning data, the documents in the search results of the query (search condition) in the learning data are compared, and the linguistic feature is determined as “feature vector” for each document. ". Further, the learning feature vector mapping unit 103 maps the feature vector to the coordinate space of the corresponding dimension. Based on this mapping, the learning model generation unit 104 expresses the result of re-ranking (rank learning) as a learning model, and stores the learning model in the learning model storage unit 125.

  Next, in actual operation, a configuration of processing in which a user or another application inputs a search condition and returns an appropriate result to a calling side using a learning model of a result learned in advance will be described.

  The user condition receiving unit 111 receives a search condition (query) of a user (or another application). The information search unit 131 searches the search target document storage unit 124 based on the search condition (query) and returns the search result. The re-ranking feature vector generation unit 112 compares the search condition with each search result to generate a feature vector. The process of generating a feature vector from one focused query and one document is the same as that of the learning feature vector generation unit 102, but in consideration of the fact that there are multiple queries themselves and the possibility that they differ depending on the learning theory during learning. , For convenience, they are divided into two functional units. In practice, the same or different ones are included in the present invention.

  The re-ranking feature vector mapping unit 113 maps the feature vector to the coordinate space generated by the learning model generation unit 104 and stored in the learning model storage unit 125. Based on this mapping, the re-ranking unit 114 performs a re-ranking process of each document with respect to the search result based on the above-mentioned user's search condition.

  FIG. 2 is a block diagram showing an example of a hardware configuration applicable to the information processing apparatus 100 according to the embodiment of the present invention.

  As illustrated in FIG. 2, the information processing apparatus 100 includes a CPU (Central Processing Unit) 201, a RAM (Random Access Memory) 202, a ROM (Read Only Memory) 203, an input controller 205, and a video controller 206 via a system bus 204. , A memory controller 207, a communication I / F controller 208, etc. are connected.

The CPU 201 centrally controls each device and controller connected to the system bus 204.

  Further, the ROM 203 or the external memory 211 will be described later, which is necessary for realizing the functions executed by the control programs of the CPU 201, such as BIOS (Basic Input / Output System), OS (Operating System), and each server or each PC. Various programs are stored. In addition, information necessary for carrying out the present invention is stored. The external memory may be a database.

  The RAM 202 functions as a main memory, a work area, etc. of the CPU 201. The CPU 201 implements various operations by loading a program or the like required for executing the processing from the ROM 203 or the external memory 211 into the RAM 202 and executing the loaded program.

  The input controller 205 also controls inputs from a keyboard (KB) 209 and a pointing device such as a mouse (not shown).

  The video controller 206 controls display on a display device such as the display 210. The display may be a display such as a liquid crystal display. These are used by the administrator as needed.

  The memory controller 207 is an external storage device (hard disk (HD)) that stores a boot program, various applications, font data, user files, edit files, various data, etc., a flexible disk (FD), or a PCMCIA (Personal Computer). Controls access to an external memory 211 such as a CompactFlash (registered trademark) memory connected to a Memory Card International Association (Card) card slot via an adapter.

  The communication I / F controller 208 connects / communicates with an external device via a network and executes communication control processing on the network. For example, communication using TCP / IP (Transmission Control Protocol / Internet Protocol) is possible.

  The CPU 201 can display on the display 210, for example, by executing an outline font rasterizing process in the display information area in the RAM 202. Further, the CPU 201 enables a user instruction with a mouse cursor (not shown) or the like on the display 210.

  Various programs to be described later for implementing the present invention are recorded in the external memory 211, and are executed by the CPU 201 by being loaded into the RAM 202 as necessary.

  FIG. 3 is an example of a document to be searched according to the embodiment of the present invention. As an example, the FAQs (Frequently Asked Questions) used to support software products are described, but the documents to be selected in the present invention include texts such as papers, newspaper articles, company regulations, and posts on SNS. Anything is acceptable, and it is not limited to FAQ.

  In the examples described in 300a and 300b, one FAQ includes a “FAQ ID” that uniquely identifies the FAQ, a “category” that classifies the contents of the entire FAQ, and support contents such as an inquiry of the user of the software product. On the other hand, it is composed of fields such as “question” which is a guide to which FAQ should be referred to and “answer” in which the response is described, and includes description of the contents corresponding to each field. Here, the FAQ ID is for associating a character string of an inquiry (query) with a specific FAQ in learning data described later, and is not limited to a field name. The names and the number of other fields are not limited.

  FIG. 4 is an example of learning data according to the embodiment of the present invention. As an example of the learning data in FIG. 4, learning data (10 cases) 400a corresponding to the FAQ 300a and learning data (10 cases) 400b corresponding to the FAQ 300b in FIG. 3 are illustrated.

  Each learning data has a query 401 (or sometimes referred to as an inquiry or a search condition) and an FAQ correspondence 402. A FAQ ID is stored in the FAQ association 402 and identifies which of the FAQs shown in FIG. 3 the learning data corresponds to.

These learning data can be obtained from a log collected by, for example, selecting an FAQ that matches the intention of the inquiry content (query) input by the user when the user actually uses the similar search engine of the present invention. it can. Further, the FAQ corresponding to the inquiry content (query) may be specified and collected from an email or the like exchanged between the user and the person in charge of support, not as a part of the function of the present invention.
FIG. 5 is an example of the generated feature vector according to the embodiment of the present invention. The similarity search in the present invention includes “at the time of learning” and “at the time of inquiry” in which the user actually browses the FAQ. In either case, a feature vector is generated.

  The feature vector is subjected to language processing on one of the documents (FAQ in this example) obtained as a result of searching the search target document storage unit 124 by the information search unit 131 using the query 401 included in the learning data as the search condition. It is a sequence of numerical values obtained by (this is regarded as a vector).

  Each of the three fields (similarity calculation field 501a) included in one FAQ and the query 401 is calculated by the three similarity calculation methods described in the similarity index 502. For example, the value 1.2 of the item indicated by 503 is a value obtained by calculating the text in the query 401 and the question field by the similarity calculation method called BM25.

  In the table 500a, nine numerical values are described in a tabular form, and the ones arranged in a row are referred to as feature vectors (nine dimensions in this example), but the description in the table below is also equated with the feature vector. .. That is, it is called a feature vector 500a.

  The fields used for calculation in this example are merely examples. It is presumed that the FAQID is usually not used because it has no linguistic meaning, but it is not necessary to use all the other three fields. In particular, it is possible to use a language that has a strong linguistic feature.

  Furthermore, the three methods described in the similarity index 502 are merely examples, and there are various other calculation methods. Details of these are omitted because they are well-known techniques in natural language processing.

  FIG. 6 is an example of characteristic words extracted from the learning data according to the embodiment of the present invention. In the embodiment of the present invention, the “feature word” is extracted from the learning data 400. To extract the characteristic word, for example, in a certain text group, a word extracted from the text A is statistically compared with a word extracted from a text other than the text A in the text group to characterize the text A (other A word that is determined to be distinguished from a document) is extracted as a characteristic word of the text A. There are various types of feature word extraction methods, but these details are omitted because they are well known in the natural language processing technology.

  The characteristic word 600a is obtained by extracting the top 10 characteristic words from all the learning data (learning data 400a in FIG. 4) whose correct answer is FAQ ID = 3198. Here, the "top 10" is given "importance" to each word in the above-described feature word extraction process. For example, "10" is written in the "feature word count" of the setting storage unit 122 in FIG. Therefore, the top 10 most important ones are selected. Alternatively, by describing "0.5" in "feature word importance", words having an importance of 0.5 or more may be selected regardless of the number.

  The characteristic word 600b describes an example in which the top 10 characteristic words are extracted from all the learning data items (learning data 400b in FIG. 4) in which the correct answer is FAQID = 0064.

  FIG. 7 is an example of the generated feature vector according to the embodiment of the present invention.

  In addition to the feature vector 500a described in FIG. 5, the feature vector 500b has an item called a learning language information field 701 added. As the learning language information, for example, the word selected by the process of extracting the characteristic word from the learning data 400 of FIG. 4, that is, the characteristic word 600 of FIG. 6 is used as an example. The information stored in the characteristic word 600 is not actually a FAQ field, but as shown in FIG. 12, for example, for the FAQ of FAQ ID = 3198, it corresponds to the item described as 3198 of the index 1201. It is stored in association with the associated learning language information 1202, which is acquired and treated as a logical field of the FAQ of FAQ ID = 3198.

  That is, in FIG. 5, only three fields are used for generating a feature vector, but in FIG. 7, a logical learning language information field 701 including a feature word as information is used.

  As described above, the information contained in the learning language information field 701 can be used for learning, for example, by collecting the information actually input by the user as a log.

  The fields of the FAQ, such as “question” and “answer”, are sentences written in “correct words” created by a person who is familiar with the related business and technology, and technical terms are used.

  However, the actual user does not know the “technical term” and uses his / her own word as the search condition. Therefore, important technical terms do not always have an effect on rank learning.

  In the method using the learning language information field 701 of the present invention, the user's search log, that is, the user's word itself is the learning target, and thus an effect of more effective learning result can be obtained.

  Further, in the feature vector 500c, the feature vector is generated only from the learning language information field 701 (501c). As a result, the effect of shortening the learning time can be obtained.

  Explaining in detail, there are several factors that determine the learning time, and among them, there is the dimension of the "feature vector" generated based on the learning data and the learning data. For example, if three feature calculation methods are applied to the three fields of “question”, “answer”, and “category” in the FAQ, the feature vector becomes 3 × 3 = 9 dimensions. On the other hand, as shown in 500c, by limiting the similarity calculation field to one learning language information field, it becomes a three-dimensional feature vector, and the effect of shortening the time required for order learning can be obtained.

  Further, the information in the learning language information field 701 is stored in the learning language information storage unit 123, but the information stored in the learning language information storage unit 123 in the present invention is limited to feature words. is not. In particular, all the learning data queries may be stored in association with each FAQ ID. In this case, the entire sentence of the query of the learning data is stored in the learning language information field 701, but it can be similarly executed because it is analyzed by natural language processing at the time of order learning and linguistic features such as words are extracted. is there. Further, the examples are not limited to the characteristic words or the whole sentence, but may be linguistically indicating the characteristics of the learning data.

  However, the use of feature words shortens the time required for order learning, as described above. Feature words have less linguistic information than the full-text query of learning data. Therefore, by limiting the extracted feature words, it is possible to obtain an effect of shortening the time required for rank learning as compared with storing the entire query text.

  Furthermore, learning data does not exist for all FAQs. That is, the search result by the information search unit 131 at the time of learning includes an inappropriate answer in addition to the FAQ designated by the user as an appropriate answer. This is because there is at least one learning data item in a proper FAQ (correct answer document), but other documents do not always have corresponding learning data items. In fact, among the FAQs, the FAQs that the user browses and browses in response to an inquiry are generally biased, and a relatively large proportion of FAQs are not even browsed. For that reason, there are many FAQs that do not have any corresponding learning data.

  If the learning data does not exist, the feature vector cannot be generated if the learning target is limited to only the learning language information field 701. In order to cope with such a case, for example, by adding the score when the information search unit 131 searches the query of the learning data as one of the features, it is possible to generate a feature vector in which at least one feature is not 0. The effect can be obtained.

  Regarding this score, as an example, only Score 500c is described as Score 702, but it goes without saying that it may be added to Score 500a and 500b.

  FIG. 8 is an example of a graph showing the distribution of the number of learning data items according to the embodiment of the present invention. In the above paragraphs, the processing for extracting the characteristic word in the learning data preprocessing unit 101 has been described. 11 is a graph for explaining that, as another method, learning data used for learning is narrowed down by the learning data preprocessing unit 101.

  In the first place, the order learning includes learning data used by the functional units 102 to 104 related to learning and user conditions used by the functional units 112 to 114 that are accepted by the user condition acceptance unit 111 during actual operation. Among them, the similarity of linguistic features is utilized to generate a learning model and improve the accuracy of retrieval.

  However, the learning data is obtained by the user's search log, that is, by the user actually inputting the search conditions and then selecting an appropriate answer from the obtained search results. However, the user does not always select an appropriate answer.

  For example, you may mistakenly view the answer as an appropriate answer, or you may specify that the answer is unrelated to the user's original intention and that it is an appropriate answer because it happened to be interesting. There is also. An optimal learning model cannot be obtained by learning even by using such inappropriate learning data. Therefore, it is necessary to classify appropriate learning data and inappropriate learning data.

<First embodiment for classifying learning data>
First, a method of selecting appropriate learning data according to the number of learning data having each FAQ as a correct answer will be presented. Based on the description that the number of learning executions in the setting parameter 1100 (FIG. 11) is 10, only when there are 10 or more learning data corresponding to one FAQ, those learning data are regarded as appropriate learning data and used for rank learning. Show the method to use.

  The learning data number graph 800 is the number of learning data having each FAQ as a correct answer. The horizontal axis represents the FAQ ID, and the vertical axis represents the number of corresponding learning data items. The learning data are arranged from left to right in descending order.

  As described above, it is general that the learning data is biased toward a specific FAQ. The FAQ IDs in the range of 804, that is, the FAQ IDs on the left of the dotted line 802, indicate that there are 10 or more corresponding learning data. In the example, the FAQ with the most learning data is 40, the FAQ 300a (the FAQ ID is 3198) illustrated in FIG. There is data.

  On the other hand, the FAQ ID in the range of 805, that is, the portion sandwiched between the dotted lines 802 and 803 has less than 10 learning data, and the range of 806 (right of the dotted line 803) has even one learning data. It corresponds to no FAQ.

  As mentioned above, some training data is inappropriate, and it seems that it occurs with a certain probability and rarely. Therefore, when the number of learning data corresponding to a specific FAQ is one or several, it is unlikely that there is inappropriate learning data, but on the other hand, even if there is one, there is inappropriate learning data. , The bad effect as learning data has a big influence. When the characteristic words are extracted as described above, many inappropriate characteristic words are selected.

  On the other hand, when there are dozens of corresponding learning data, even if one inappropriate learning data is included, statistically linguistic features are extracted from dozens of learning data. Among them, there will be almost no adverse effects.

  That is, even if there is inappropriate learning data with a certain low probability, the more learning data corresponding to a particular FAQ, the higher the possibility of ignoring, and the less the learning data, the more ignoring. It will not be possible. From that point of view, for example, when there are less than 10 pieces of learning data corresponding to the FAQ, it is decided not to use those learning data, so that the cause of adversely affecting learning is excluded.

  Further, in the example of the graph 800, since the FAQs that are actually frequently inquired are concentrated in 804, this portion is subjected to rank learning with a large amount of learning data and the accuracy is improved, so that an appropriate result for the user can be obtained. Therefore, even if the learning in the range of 805 and 806, which is rarely inquired, is not performed at all, the probability that the user feels a problem is low.

  With the method described above, it is possible to obtain the effect that the optimum learning can be performed according to the actual usage frequency of the user.

<Second embodiment for classifying learning data>
Describe other methods. If the query included in the learning data is appropriate, even if the learning is not performed (that is, even if the search is performed by the information search unit 131), the correct answer FAQ is relatively high. Therefore, for example, when the correct answer FAQ that is associated is within the 50th place as a result of searching with the learning data query using the 50th place as a threshold, the learning data is regarded as good learning data.

<Third embodiment for classifying learning data>
As a method similar to the second embodiment, not the rank but the similarity (search score) between the query and the FAQ search results is used as the threshold. That is, if the degree of similarity is a certain value or more, it is considered as good learning data.

<Fourth embodiment for classifying learning data>
As a fourth embodiment, there is also a method of combining both the third and fourth embodiments and using both the rank and the similarity as threshold values. Other than the correct answer in the learning data and the search results, numerical information related to the degree of similarity, linguistic information such as words contained inside, and other learning data and correct answer FAQ instead of a single learning data It goes without saying that any method may be used as long as it can be classified using the statistical values obtained from the above.

  FIG. 9 is an example of a flowchart for explaining the processing at the time of learning according to the embodiment of the present invention, that is, the generation of the learning model.

  In step S901, the learning data stored in the learning data storage unit 121 is read.

  In step S902, information used for rank learning is extracted from the learning data and stored in the learning language information storage unit 123. For example, the characteristic word shown in FIG. 6 is extracted and stored in association with the FAQ corresponding to the learning data as shown in FIG. At that time, the learning data to be processed may be selected in advance as described in FIG. As another example, as described above, the query of the learning data may be stored as it is in association with each FAQ ID. That is, it is possible to store only without substantially performing processes such as extraction and selection.

  In steps S903 to S911, if all the learning data read in step S901 or some of the learning data is selected in step S902, the iterative process is performed on all the selected learning data.

  In step S904, attention is paid to one of the learning data, and in step S905, the information search unit 131 searches the search target document storage unit 124 by the query of the learning data. Here, one or more documents are hit for the query. In some cases, there is no hit document, but in that case, the following processing is interrupted and the next learning data is focused on by repeating the processing.

  Steps S906 to S910 are an iterative process for the document hit by the query of interest.

  In step S907, attention is paid to one of the documents hit by the query in question.

  In step S908, learning language information corresponding to the document of interest, for example, if the document of interest has FAQ ID = 3198, the characteristic word corresponding to (the FAQ ID of) the document shown in FIG. 12 is acquired.

  In step S909, a feature is calculated from the query of interest and the field to be learned of the document of interest. Here, the characteristic word acquired in step S908 is also set as a learning target in the (logical) learning language information field 701. The feature vectors exemplified by 500b and 500c in FIG. 7 are generated by performing calculations for a plurality of calculation methods and fields. Furthermore, the generated feature vector is mapped to the coordinate space. At this time, since there is a description in the learning data as to whether or not the document under consideration is the correct answer for the query under consideration, a label indicating that the session is being held is stored in association with the coordinates of the mapping destination. When the mapping of the feature vector to the coordinate axes has been completed for all the learning data and the document hit by the search, the process proceeds to step S912.

  In step S912, an appropriate learning model is generated based on all the mapped feature vectors, and the learning model is stored in the learning model storage unit 125. Regarding the generation of the learning model, various methods have been presented and are well-known techniques, so a detailed description thereof will be omitted.

  This completes the process of generating the learning model using the learning data and the search target document. This completes the description of the flowchart of FIG.

  FIG. 10 is an example of a flowchart illustrating the information search / reranking process based on the learning result according to the embodiment of the present invention.

  In step S1001, the user condition receiving unit 111 receives the search condition (query) of the user (or another application).

  In step S1002, the information search unit 131 searches the search target document storage unit 124 with the received query and returns the hit document (FAQ or the like).

  Steps S1003 to S1007 are repetitive processing for all the hit documents.

  In step S1004, attention is paid to one of the documents as the search result.

  In step S1005, learning language information corresponding to the document of interest is acquired. For example, if the document of interest is the FAQ 300a in FIG. 3, the learning language information 1202 corresponding to 3198 in the index 1201 of the learning language information storage unit 123 in FIG. 12 is acquired. As an example, 10 characteristic words 1203 are described.

  In step S1006, a feature vector obtained from the query accepted by the user condition accepting unit 111 and the FAQ 300a including the logical learning language information field 701 is generated. For example, feature vectors such as 500b and 500c in FIG. 7 are obtained.

  Further, the feature vector is mapped from the learning model storage unit 125 to the coordinate system including the learning model described in FIG. 9 (the learning model may be read once at the start of the flowchart in FIG. 10).

  As a result of the iterative process, the number of documents hit by the search condition (query) of the user is mapped. As a result, the repeating process of steps S1003 to S1007 is completed, and the process proceeds to step S1008. Note that unlike the case described in the flowchart of FIG. 9, in the above-mentioned query, since the correct FAQ is not known, a label indicating whether or not it is the correct answer is not shown.

  In step S1008, the mapping of the feature vectors corresponding to all the documents hit with respect to the user's search condition (query) is compared with the learning model generated in the process of FIG. 9 to rank the maps. Since this ranking is a well-known technique in correspondence with the method of generating the learning model, detailed description thereof will be omitted.

  The documents ranked in the above process (the documents that hit the query) are presented to the caller of the search process. This is the end of the description of the flowchart using FIG.

  It is needless to say that the structure and contents of the various data described above are not limited to this, and may be composed of various structures and contents according to the use and purpose.

  Although some embodiments have been described above, the present invention can be embodied as, for example, a system, an apparatus, a method, a computer program, a recording medium, or the like, and more specifically, a plurality of devices. The present invention may be applied to a system composed of, or may be applied to a device composed of one device.

  Further, the computer program in the present invention is a computer program that allows the computer to execute the processing method of the flowcharts shown in FIGS. 9 to 10, and the storage medium of the present invention can execute the processing method in FIGS. 9 to 10 by the computer. Various computer programs are stored. The computer program according to the present invention may be a computer program for each processing method of each device in FIGS. 9 to 10.

  As described above, a computer in which a recording medium recording a computer program that realizes the functions of the above-described embodiments is supplied to a system or apparatus, and the computer (or CPU or MPU) of the system or apparatus is stored in the recording medium. It goes without saying that the object of the present invention can also be achieved by reading and executing the program.

  In this case, the computer program itself read from the recording medium realizes the novel function of the present invention, and the recording medium storing the computer program constitutes the present invention.

  As a recording medium for supplying the computer program, for example, a flexible disk, a hard disk, an optical disk, a magneto-optical disk, a CD-ROM, a CD-R, a DVD-ROM, a magnetic tape, a non-volatile memory card, a ROM, an EEPROM, A silicon disk, a solid state drive, etc. can be used.

  Further, by executing the computer program read by the computer, not only the functions of the above-described embodiments are realized, but also an OS (operating system) or the like running on the computer is executed based on the instruction of the computer program. It goes without saying that a case where some or all of the actual processing is performed and the functions of the above-described embodiments are realized by the processing is also included.

  Furthermore, after the computer program read from the recording medium is written in the memory provided in the function extension board inserted in the computer or the function extension unit connected to the computer, the function is executed based on the instruction of the computer program code. It goes without saying that a case where the CPU or the like provided in the expansion board or the function expansion unit performs a part or all of the actual processing and the processing realizes the functions of the above-described embodiments is also included.

  Further, the present invention may be applied to a system including a plurality of devices or an apparatus including one device. Further, it goes without saying that the present invention can be applied to the case where it is achieved by supplying a computer program to a system or an apparatus. In this case, by reading the recording medium storing the computer program for achieving the present invention into the system or apparatus, the system or apparatus can enjoy the effects of the present invention.

  Further, by downloading and reading a computer program for achieving the present invention from a server, a database or the like on a network using a communication program, the system or apparatus can enjoy the effects of the present invention.

  It should be noted that the present invention also includes all configurations that combine the above-described embodiments and modifications thereof.

100 information processing device 101 learning data pre-processing unit 102 learning time searching unit 103 information searching unit 104 learning feature vector generation unit 105 learning feature vector mapping unit 106 learning model generation unit 111 user condition acceptance unit 112 user condition search unit 113 re Ranking feature vector generation unit 114 Reranking feature vector mapping unit 115 Reranking unit 121 Learning data storage unit 122 Setting storage unit 123 Learning language information storage unit 124 Search target document storage unit 125 Learning model storage unit

Claims (9)

An information processing apparatus comprising: a search unit that searches a search target document with a search query text; and a storage unit that stores the search target document and a learning search query text associated with the search target document,
Creating means for creating additional text information for the search target document using the learning search query text associated with the search target document;
An information processing apparatus comprising: a learning unit that performs order learning using a set of the learning search query text and document data including additional text information for the search target document.   The information processing apparatus according to claim 1, wherein the learning unit performs order learning using a set of the search query text for learning and document data including only additional text information for the search target document. ..   The information processing apparatus according to claim 1, wherein the learning unit performs rank learning using feature data calculated from the learning search query text and the document data.   The information processing apparatus according to claim 3, wherein the feature data includes a search score when the search target document corresponding to the document data is searched by the learning search query text.   The information processing apparatus according to claim 1, wherein the additional text information includes a characteristic word extracted from the learning search query text.   The information processing apparatus according to claim 5, wherein the number of characteristic words included in the additional text information is limited according to a predetermined value.   The learning unit may include additional text information for the search target document in the document data when the number of learning search query texts associated with each of the search target documents exceeds a predetermined value. The information processing device according to claim 1. A control method for an information processing apparatus, comprising: a search unit that searches a search target document with a search query text; and a storage unit that stores the search target document and a learning search query text associated with the search target document. hand,
A creating step in which creating means creates additional text information for the search target document by using the learning search query text associated with the search target document;
A control method for an information processing apparatus, comprising: a learning step in which a learning step performs learning in order using a set of the learning search query text and document data including additional text information for the search target document. A program executable in an information processing apparatus including a search unit that searches a search target document with a search query text, and a storage unit that stores the search target document and a learning search query text associated with the search target document And
The information processing device,
Creating means for creating additional text information for the search target document using the learning search query text associated with the search target document;
A program that functions as a learning unit that performs order learning using a set of the learning search query text and document data including additional text information for the search target document.

Images