Search system, search method, and program

Abstract

To provide a retrieval system which provides a retrieval result necessary for a user.SOLUTION: A retrieval system includes: a search query acquisition unit for acquiring a search word; a retrieval unit which performs document search on the search word acquired by the search query acquisition unit; and a storage unit which stores information on a document browsed out of documents retrieved by the retrieval unit and information on the search word in the document search in association with each other. The retrieval unit performs document search on the basis of the document information stored in the storage unit corresponding to the information on the search word acquired by the search query acquisition unit.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] The present disclosure relates to a search system, a search method, and a program.

Background Art

[0002] For companies, it is important to quickly reflect consumer needs and desires in products and services and improve competitiveness. Improving competitiveness requires accurate information, but currently, information is flooding, and it is difficult to collect accurate information. Patent Document 1 discloses a search system having a function of proposing recommended search keywords so that expected search results can be obtained when a user inputs a search term. By using such a search system, the possibility of collecting accurate information is increased. However, in the case of such a search system, the work of registering recommended keywords requires time and effort. Also, if the registered recommended keywords are insufficient, appropriate keywords may not be proposed, and it may not be possible to search for necessary information. For example, in a scenario where a user who does not have sufficient knowledge and experience regarding plants tries to find materials necessary for plant facility operation and maintenance work by search, it may not be possible to come up with appropriate search terms.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] There is a need for a search system that can provide desired search results regardless of the user's knowledge and skills.

[0005] The present disclosure provides a search system, a search method, and a program that can provide search results desired by a user.

Means for Solving the Problems

[0006] The search system of the present disclosure includes a search query acquisition unit that acquires a search term, a search unit that performs a document search for the search term acquired by the search query acquisition unit, and among the documents retrieved by the search unit, the viewed documents The probability information for each category indicating the probability of which category the document belongs to and the search term in the document search Dispersion expression and a storage unit that stores them in association with each other. The search unit performs the document search based on Dispersion expression the storage unit corresponding to the search term The probability information for each category acquired by the search query acquisition unit.

[0007] The search method of the present disclosure The computer includes a search query acquisition step of acquiring a search term, The computer a search step of performing a document search for the search term acquired in the search query acquisition step, The computer and a storage step of storing in association with each other the viewed documents among the documents retrieved in the search step The probability information for each category indicating the probability of which category the document belongs to and the search term in the document search. Dispersion expression In the search step, The computer the document search is performed based on Dispersion expression the storage corresponding to the search term The probability information for each category acquired in the search query acquisition step.

[0008] The program of the present disclosure causes a computer to execute a search query acquisition step of acquiring a search term, a search step of performing a document search for the search term acquired in the search query acquisition step, and a storage step of storing in association with each other the viewed documents among the documents retrieved in the search step The probability information for each category indicating the probability of which category the document belongs to and the search term in the document search. Dispersion expression In the search step, the document search is performed based on the search term Dispersion expression corresponding toThe probability information for each category Execute a process for performing the document search based on this.

Effect of the Invention

[0009] According to the search system, search method, and program of the present disclosure, a user can provide desired search results.

Brief Description of the Drawings

[0010]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Figure 8

Figure 9

Figure 10

Figure 11

Figure 12

Figure 13

Figure 14

Figure 15

Modes for Carrying Out the Invention

[0011] Hereinafter, the search system according to the embodiment will be described in detail with reference to FIGS. 1 to 15. <First Embodiment> (System Configuration) FIG. 1 is a block diagram showing an example of the search system according to the embodiment. The search system 10 is configured using one or more server devices and storage devices. The search system 10 searches for documents necessary for the user with high precision and presents the search results. The search system 10 includes a search query acquisition unit 11, a learning unit 12, a search unit 13, a search result output unit 14, a storage unit 15, and a search log management unit 16.

[0012] The search query acquisition unit 11 acquires the search query input by the user. The search query includes one or more search terms (keywords). The learning unit 12 creates a learned model and the like necessary for high-precision search. The learning unit 12 includes a document analysis unit 121, a document classification unit 122, a similar word list creation unit 123, and a learned model creation unit 124.

[0013] The document analysis unit 121 calculates a vector obtained by dispersing each word using a tool such as word2vec or doc2vec for each document to be searched and each word included in each document. For example, the document analysis unit 121 converts a document into a k-dimensional vector and a word into an l-dimensional vector for dispersed representation. The values and dimensions of the dispersed representation are set according to the amount of documents.

[0014] The document classification unit 122 classifies the documents to be searched into N categories. For example, the document classification unit 122 classifies each document represented as a k-dimensional vector by the document analysis unit 121 using a method such as the k-nearest neighbor method, and groups those with similar k-dimensional vector distances into the same group to create N groups. When classifying documents, attribute information of the documents may be used in combination. For example, as the attribute information of the document, attributes that can be recognized by people such as "manual", "report", "journal / record", etc. are given to each document, and the groups of those with similar vectors may be further classified by attribute information.

[0015] The similar word list creation unit 123 creates a similar word list in which words with similar l-dimensional vectors are grouped together using the word analysis results by the document analysis unit 121 and the like. In the similar word list, for example, "law" and "legal" and "legal affairs" are registered as the same group. As will be described later, in the search system 10, for example, when a user inputs "law" as a search term and searches, not only "law" but also "legal" and "legal affairs" can be searched. The similar word list is referred to in such a case and is created for the purpose of providing another search term ("legal" or "legal affairs") to the search unit 13.

[0016] The learned model creation unit 124 creates a learned model that learns the relationship between the search term and N document categories. For example, when the search term "fan" is input, information on the documents actually accessed by the user is collected from the search log, and it is tallied which of the N categories the document is classified into by the document classification unit 122, and the relationship between the distributed representation of "fan" and the document category is learned. For example, if as a result of the search, the distribution of the categories to which the actually viewed documents belong is category 1 = 50%, category 2 = 30%, and category 3 = 20%, the learned model creation unit 124 creates a learned model that, when input with the distributed representation (l-dimensional vector) of "fan", outputs category 1 = 0.5, category 2 = 0.3, and category 3 = 0.2. The pair of the word in distributed representation and the document with the category assigned can be obtained from the search log in which the user's search history is recorded. The search log records the search term and the documents actually viewed among the search results. Therefore, based on the records in the search log, a pair combining these is created. Then, using the word in distributed representation as the input and the category information represented by a one-hot vector or the like consisting of N elements as the correct label, this is used as learning data. Specifically, a correct label is assigned in which the element corresponding to the category to which the viewed document belongs is set to 1 and the remaining elements are set to 0. The learned model creation unit 124 performs learning using the learning data by means of a multi-layer neural network or the like, and generates a learned model that calculates classification probabilities for each of the N categories for an arbitrary l-dimensional vector.

[0017] The search unit 13 searches for documents using the search term included in the search query acquired by the search query acquisition unit 11, and outputs, as search results, the documents with high values of the relevance score described later to the search result output unit 14. The search unit 13 includes a search term setting unit 131, a search term analysis unit 132, a document score calculation unit 133, and a relevance score calculation unit 134. The search term setting unit 131 refers to the similar word list, reads out one or more other search terms similar to the search term input by the user, and adds them to the input search term. The search term analysis unit 132 calculates a distributed representation vector (an l-dimensional vector) of the search terms (including those added by the search term setting unit 131) using, for example, word2vec or the like.

[0018] The document score calculation unit 133 performs a search of documents using the search terms included in the search query and similar search terms added by the search term setting unit 131, and calculates a document score assigned to each document. The method for calculating the document score may generally be a method used in a full-text search engine. The document score is a value calculated for each document based on conditions such as, for example, how many times the search terms given in the search query appear in the document, or whether the search terms appear in the title or the body of the text.

[0019] The relevance score calculation unit 134 inputs the l-dimensional vector of the search terms converted by the search term analysis unit 132 into the learned model created by the learned model creation unit 124, and obtains a category inference result as the output from the learned model. When there are two or more search terms, the sum of the vectors may be used as the input. The category inference result output by the learned model consists of N elements whose sum is 1, and each element is a value between 0 and 1. The relevance score calculation unit 134 adds an arbitrary value (for example, "1") to each value of the N elements, and uses the value after addition as a weight for determining the relevance score for the search query. The relevance score calculation unit 134 multiplies the document score calculated by the document score calculation unit 133 for each document in the search results by the weight given for the category (any one of the N categories classified by the document classification unit 122) to which the document belongs, and calculates the relevance score. The search unit 13 arranges the search results in descending order of the relevance score value and outputs them to the search result output unit 14.

[0020] The search result output unit 14 outputs information on the documents with high relevance score values as search results to a display device, an electronic file, another system, etc. The memory unit 15 stores the document to be searched, the search log, the program for realizing the above-described search query acquisition unit 11, learning unit 12, search unit 13 (including the program of a search engine used in general full-text search), search result output unit 14, the indexer and index (inverted file) used in general full-text search, and various information necessary for the search process of the present embodiment (learned model, list of similar words, document category classification information, analysis result vectors of documents and words, etc.). The search log management unit 16 records the search log in the memory unit 15.

[0021] (Operation) (Learning Phase) Next, with reference to FIG. 2, the operation in the learning phase of the search system 10 will be described. FIG. 2 is a flowchart showing an example of the process in the learning phase according to the first embodiment. First, documents are aggregated (step S1). The user registers the document to be searched in the memory unit 15 as the target of subsequent processing. At this time, for paper documents, they are OCRed or the like to register the electronically filed documents. Next, the document analysis unit 121 executes morphological analysis processing on the title of the document, the words included in the document, and the attribute information attached to the document using a tool (step S2). The document analysis unit 121 records the analysis result in the memory unit 15.

[0022] Next, the document analysis unit 121 creates a distributed representation of the document (step S3). For example, the document analysis unit 121 extracts necessary nouns and the like from each morpheme after morphological analysis, and calculates a vector of the distributed representation for the extracted words using word2vec or dec2vec. The document analysis unit 121 adds up the distributed representations (title, words included in the document, analysis results of attribute information attached to the document) obtained for, for example, one document, and performs dimensionality reduction as necessary to calculate a k-dimensional vector that is the distributed representation of the document. For example, only those that appear a predetermined number of times or more among the words appearing in the document may be targeted for summation.

[0023] Next, the document classification unit 122 classifies the document (step S4). The document classification unit 122 classifies the k-dimensional vector of the entire document obtained in step S3 into N groups using the k-nearest neighbor method or the like. The document classification unit 122 records the classification result in the storage unit 15. For example, information such as document 1 belonging to category 1, document 2 belonging to category 2, document 3 belonging to category N, ··· is recorded.

[0024] Next, the document analysis unit 121 creates a distributed representation of words (step S5). For example, the document analysis unit 121 extracts necessary words (excluding auxiliary words etc.) from the analysis result of step S2, and uses word2vec or the like to obtain an l-dimensional vector which is the distributed representation of each word.

[0025] Next, the trained model creation unit 124 acquires pairs of search terms and documents from the search log (step S6). The search log is stored in the storage unit 15 in advance. For example, the search log records the time when the search was executed, the user ID or terminal ID of the user who executed the search, the search query (search term) used in the search, and the history of the document IDs of the documents actually accessed by the user from the search results. The trained model creation unit 124 refers to the search log and creates pairs of search terms and the document IDs of the documents actually accessed by the user one by one, and acquires pairs of search terms and documents (document IDs) from the search log.

[0026] Next, the trained model creation unit 124 generates training data (step S7). The trained model creation unit 124 attaches information (for example, one hot vector) indicating the category into which the document is classified to the documents in the pairs acquired in step S6 based on the result of step S4 (document category) (correct label). The trained model creation unit 124 performs this process for all pairs to create training data.

[0027] Next, the trained model creation unit 124 generates a trained model (step S8). The trained model creation unit 124 trains the training data created in step S7 using a multi-layer neural network or the like, and creates a trained model that outputs the probability for each category of the document selected for the search term when an l-dimensional vector representing the search term is input. The trained model creation unit 124 records the created trained model in the storage unit 15.

[0028] Next, the similar word list creation unit 123 generates a similar word list (step S9). The similar word list creation unit 123 calculates the similarity of the l-dimensional vectors between words using the processing result of step S5, and for example, selects a predetermined number of words in order from those with a similarity equal to or greater than a predetermined threshold value or those with a high similarity, and registers these as a set in the similar word list. When using word2vec or the like, it is possible to analyze the relationship between the target word and the words before and after in the sentence in which the target word is used, and quantify (vectorize) the meaning and properties of the target word. Therefore, by selecting vectors similar to the target word, words with similar meanings or strong relevance can be automatically selected. By registering such words in the similar word list, the administrator can perform searches using synonyms or co-occurring words without dictionary-registering words with similar meanings or strong relevance. The similar word list creation unit 123 records the created similar word list in the storage unit 15. Through the processing in the above learning phase, classification information of the document into N categories, a trained model that has learned the relationship between the search term and the category to which the search result belongs, and a similar word list are created.

[0029] (Search phase) Next, with reference to FIG. 3, the operation in the search phase of the search system 10 will be described. In the search phase, a search is performed using the document category classification information, the trained model, and the similar word list created in the learning phase. FIG. 3 is a flowchart showing an example of the processing in the search phase according to the first embodiment. First, the user inputs a search query (e.g., one or more search terms) on the search screen. The search query acquisition unit 11 receives the input of the search query (step S11). The search query acquisition unit 11 outputs the received search query to the search unit 13. In the search unit 13, first, the search term setting unit 131 adds similar search terms (step S12). Specifically, for each search term included in the search query, the search term setting unit 131 refers to the similar word list in the storage unit 15 to obtain words similar to the input search term. The search term setting unit 131 outputs the words similar to the input search term in pairs to the search term analysis unit 132. Note that the process of this step S12 can be omitted. For example, the user may be able to select whether to perform a search using only search terms or to perform a search using similar words in parallel. When the user selects to perform a search using only search terms, step S12 is omitted, and the search term setting unit 131 outputs only the input search term to the search term analysis unit 132.

[0030] Next, the search term analysis unit 132 converts the search terms obtained from the search term setting unit 131 into a distributed representation (step S13). For example, when the search term analysis unit 132 obtains the search terms "law", "legal", and "legal affairs" from the search term setting unit 131, it converts each word into a distributed representation of an l-dimensional vector using word2vec or the like. When the addition of similar search terms in step S12 is not performed, the search term analysis unit 132 converts the search term "law" into a distributed representation.

[0031] Next, the relevance score calculation unit 134 infers the category using the learned model (step S14). The relevance score calculation unit 134 inputs the search term expressed as an l-dimensional vector into the learned model in the storage unit 15, and obtains the probabilities for each of the N categories output by the learned model. When there are multiple search terms, the relevance score calculation unit 134 may input the sum of the multiple vectors corresponding to the search terms into the learned model. Next, the relevance score calculation unit 134 calculates weights based on the category inference result (step S15). For example, when the probability of category 1 is p1, the probability of category 2 is p2, ···, and the probability of category N is pN, the relevance score calculation unit 134 calculates p1 + 1, p2 + 1, ···, pN + 1, and uses them as the weights for categories 1 to N, respectively.

[0032] Next, the document score calculation unit 133 performs a document search based on one or more search terms. For the following steps S16 to S18, a general full-text search processing method can be used. The document score calculation unit 133 refers to an index (such as an inverted file) (step S16). In the inverted file, the document ID of the document in which the word (search term) appears is registered in association with the word. The document score calculation unit 133 identifies the documents in which the search term appears based on the index, and counts the frequency of occurrence of the word in those documents, etc. (step S17). The values such as the frequency of occurrence are calculated in advance together with the index, and the document score calculation unit 133 obtains this value. Then, the document score calculation unit 133 calculates a document score for each document (step S18). The specific calculation method of the document score may vary according to the specifications of the search engine of the document score calculation unit 133 (for example, if the search term is included in the title, the document score is higher than when it is included in the text, if the update date and time of the document is newer, a higher document score is given, ···, etc.).

[0033] When the document score calculation unit 133 performs a search of documents and calculates a document score for each document, next, the relevance score calculation unit 134 calculates a relevance score for each document (step S19). For example, assume that the number of retrieved documents is three, namely Document 1 to Document 3, and the document scores of each document are Document Score 1 to Document Score 3, respectively. Also assume that the categories of Document 1 to Document 3 are Category 1 to Category 3, respectively. The relevance score calculation unit 134 calculates "(Document Score 1 × (p1 + 1))" for Document 1 to calculate the relevance score of Document 1. Similarly, the relevance score calculation unit 134 calculates "(Document Score 2 × (p2 + 1))" and "(Document Score 3 × (p3 + 1))", and sets them as the relevance score of Document 2 and the relevance score of Document 3, respectively.

[0034] Next, the search result output unit 14 displays the search results (step S20). The search unit 13 arranges the retrieved documents in descending order of the relevance score value, and outputs, for each document, search result information with the document name, link destination, content summary, etc. added to the search result output unit 14. The search result output unit 14 outputs the information formatted in a display format to a user terminal or the like.

[0035] When the user selects any one of the search results and views the corresponding document, information such as the selected document ID is notified to the search system 10. The search log management unit 16 acquires the notified document viewing information (step S21). For example, in addition to the document ID, the search log management unit 16 acquires information such as the search query acquired by the search query acquisition unit 11, the acquisition time thereof, the ID of the user who performed the search, the affiliation, the location, and the like. The search log management unit 16 stores the acquired search query and viewing information (document ID) and the like in the search log (step S22). The search log management unit 16 continuously accumulates the user's viewing history as a search log. After a certain period of time has passed, the learning unit 12 executes the process described with reference to FIG. 2 and performs learning again. The documents aggregated in step S1 are continuously added, and the creation of the distributed representation and the classification of the documents are also executed again at regular intervals. As a result, the search target of the documents is expanded, and by continuously advancing the learned model based on the accumulated search log, it becomes possible to provide search results that conform to the user's document viewing tendency.

[0036] Next, with reference to FIG. 4, the difference from document search by a general search engine will be described. FIG. 4 is a schematic diagram of the search algorithm according to the first embodiment. When the user inputs a search term, in the case of general document search, the search is performed only with that search term. However, in the search system 10 of this embodiment, the search term setting unit 131 refers to the similar word list 123a and creates a set 131a of search terms with added "similar search term 1", "similar search term 2", etc. The search is performed for each search term included in the set 131a of search terms. Therefore, for example, even when a user with little knowledge cannot input an accurate keyword for searching, searches using similar words, words with a deep relationship that are often used in the same context, etc. are automatically performed simultaneously, so the possibility of the user being able to search for the desired document is increased. In addition, since the similar word list 123a is automatically created based on the similarity of the vectors obtained from the morphological analysis of the words, for example, no maintenance cost such as registering similar words in a dictionary is required.

[0037] Once the search term is determined, general document search by a common search engine is then performed. Generally, the indexer 133c generates an index 15b from the document 15a in advance. The index 15b is registered in the storage unit 15. Then, the search engine refers to the index 15b, identifies the documents containing the search term, and calculates a document score for each document. The document score is calculated according to, for example, the frequency of occurrence of the search term in the document and whether the search term is included in the document title. In the case of general document search, the documents with higher values of this document score are extracted as top search results. However, the documents retrieved in such a way may indeed contain many search terms, but they may not be the content desired by the user. Therefore, in this embodiment, based on the record of past search logs, a pair of the search term and the document is created, and the document is further classified into one of N document categories. For the search results when a certain search term is input, the tendency of which category the documents actually browsed by the user belong to is learned (trained model 124a). Then, a relevance score is calculated by multiplying the weight according to the learned tendency by the document score (134a). In the relevance score, for the frequency of occurrence of the search term, etc. (document score), the record of the category to which the document belongs being actually browsed by the user is weighted. Therefore, simply because a large number of search terms appear, the relevance scores of documents in the same category as the documents not used by the user are suppressed to be low, and the relevance scores of documents in the same category as the documents frequently used by the user (documents with similar vectors) are highly evaluated. Also, as shown in the flow of FIG. 3, the search log is constantly updated, and if the learning phase of FIG. 2 is repeated, the documents that initially appeared at the top of the search results but are simply those with a large number of search terms and are content-wise outside the user's interest will gradually lower their search result rankings. On the contrary, even if a document was not initially ranked very high, if that document has been browsed a lot, that document, similar (belonging to the same category), and documents containing the search term will gradually appear at the top of the search results. Search results can be obtained such that documents classified into the same category as the documents of value to the user and the objects of interest are preferentially ranked at the top.According to the search system 10 of this embodiment as described above, search results (134b) corresponding to the user's browsing tendency can be obtained based on the relevance score.

[0038] (Effect) As described above, according to this embodiment, by using the distributed representation for each of the word and the document, it is possible to return search results according to the similarity of the search term and the document, so that it is possible to provide the information necessary for the user. For example, even for a user with little business experience and insufficient knowledge, when focusing on a certain word, not only the word itself but also the documents containing related words are presented as search results, so that the probability of accessing useful materials that can lead to problem-solving and prevention of troubles for the user is improved.

[0039] Note that the information necessary for a user may vary depending on the organization or role to which the user belongs. For example, the documents of interest to users in the quality control department may be different from those of users in the manufacturing department, and each document may belong to a different category. Therefore, when creating learning data based on the search log, it may be possible to create learning data separately for each user attribute and create a learned model for each user attribute. For example, based on the user ID of the quality control department, the searches and document viewing results by the users in the quality control department are extracted from the search log, and a learned model 1 suitable for the quality control department is created. Similarly, the searches and document viewing results of the users in the manufacturing department are extracted from the search log, and a learned model 2 suitable for the manufacturing department is created. Further, a learned model 3 is created from the entire search log regardless of the department. Then, the user can select from three options: "obtain search results suitable for the quality control department", "obtain search results suitable for the manufacturing department", and "regardless of department" during the search. For example, when "obtain search results suitable for the quality control department" is selected, the weights are calculated based on the learned model 1. As a result, valuable search results can be obtained for the users in the quality control department. Also, even for users in the quality control department, when they want to search for documents beyond their normal business scope, they can select "obtain search results suitable for the manufacturing department" or "regardless of department" to perform the search. In this way, it is possible to find documents according to the viewing records of many other users.

[0040] <Second Embodiment> Hereinafter, the search system 10A of the second embodiment will be described with reference to FIGS. 5 to 10. In the first embodiment, it was assumed that based on the search log, the relationship between the search terms specified by the user and the documents actually viewed could be learned. However, in a situation where there is no accumulation of search logs, such as at the start of the system, effective learning cannot be performed. In the second embodiment, mainly for the period until the search log is accumulated and learning becomes possible, a function of proposing search terms that enable effective searches for users is provided.

[0041] (Configuration) FIG. 5 is a block diagram showing an example of a search system according to the second embodiment. Among the configurations according to the second embodiment of the present disclosure, those that are the same as the configurations constituting the search system 10 according to the first embodiment are denoted by the same reference numerals, and their descriptions are omitted. The search system 10A according to the second embodiment includes a search term proposal unit 17 in addition to the configuration of the search system 10 of the first embodiment. The search term proposal unit 17 analyzes the relationship between words in the document to be searched, and creates a score table that can index words highly relevant to a certain word (= search term) based on that word. Then, based on the score table from the search term, it assigns highly relevant words and proposes the assigned words as search keywords to be added to the search terms specified by the user.

[0042] The search term proposal unit 17 includes a classification setting unit 171, a reading analysis unit 172, a reduction unit 173, an attribute setting unit 174, and a score table creation unit 175. The classification setting unit 171 sets the classification of the document to be searched. The classification of each document is specified by the user based on the source of the document. The reading analysis unit 172 reads the text data included in the document to be searched by OCR (Optical Character Recognition) or the like, and morphologically analyzes the read text data using a tool to decompose it into morphemes. The reduction unit 173 performs a process of deleting misrecognized words, particles, low-frequency words, overly high-frequency words, etc. among the words appearing in the document, and leaving the words representing the characteristics of the document. The attribute setting unit 174 sets an attribute for each word. For example, the attribute setting unit sets any one of the three attributes of "equipment name", "company name, etc.", and "event" for each word. The attributes are specified by the user. The score table creation unit 175 creates a score table associating highly relevant words with each other. The search term proposal unit 17 determines the search keywords to be proposed to the user using this score table.

[0043] (Operation) FIG. 6 is a flowchart showing an example of search term proposal processing according to the second embodiment. First, the user registers the electronic data of a document file including text in the search system 10A. The search system 10A acquires the electronic data of the document file (step S31) and registers the document file in the storage unit 15. Next, the user classifies the registered documents based on their sources (step S32). For example, the user classifies each document into "business standard", "trouble / accident", "inspection / construction", "daily business", etc. according to its source (department that issued the document, etc.) or document type. The user associates classification information indicating the document type with the electronic file of the document with an identification ID and registers it in the search system 10A. For example, the user registers a document with a description of a non-conformance case as being classified as "trouble / accident" and a document related to the internal business standard as being classified as "business standard". In the search system 10A, the classification setting unit 171 receives this classification information and registers the classification information in the storage unit 15 in association with the document.

[0044] Next, the reading and analysis unit 172 reads out the documents and classification information registered in the storage unit 15, performs character recognition on the text data in the documents by OCR or the like as necessary, and morphologically analyzes the recognized text data (step S33). The reading and analysis unit 172 recognizes text data such as the following data 0, for example.

[0045] · Data 0 <category> Trouble / Accident ...Around 20:30, the extraction of equipment A was completed On September 8 (Tuesday), we received a notice from the customer that the water supply volume had increased and there was a suspicion of leakage. The unit stopped at 11:51...

[0046] Here <category>The tag describes the classification information (document type) set in step S32. The tag describes the text data recognized by OCR. When the reading analysis unit 172 performs morphological analysis on this, it is converted as follows, for example, into data 1 below.

[0047] · Data 1 <category> Trouble / Accident …Around 20:30, the extraction of Equipment A was completed. On September 8 (Tuesday), we received a notice from a customer that the water supply had increased and there was a suspicion of leakage. The unit was stopped at 11:51. …

[0048] Next, the reduction unit 173 deletes misrecognized words, auxiliary words, and words with low / high appearance frequencies (step S34). The reduction unit 173 creates the following data by reducing unnecessary morphemes from the above data. For example, the reduction unit 173 deletes misrecognized words that occur at a certain frequency from the text data with part-of-speech extraction. Also, the reduction unit 173 deletes words, auxiliary words, and words with very few appearance times that appear in any document.

[0049] ·Data 2 <category> Trouble / Accident … / Equipment A / Withdrawal / Completed / Customer / Water supply increased / Suspicion of leakage / Contact / Unit stopped / …

[0050] The reduction unit 173 associates the remaining words with part-of-speech tags, the document ID, and the location where they appear (the nth word counted from the beginning), and records them in the storage unit 15.

[0051] Next, the user sets attributes for the part-of-speech tags of the above data 2 created by the reduction unit 173. The user picks up the words classified as nouns among the part-of-speech tags, extracts the words related to the names of plant facilities and their parts, components, etc. from them, and gives "facility name" as the attribute of these words. The user picks up the words classified as proper nouns, extracts the words related to company names, etc. from them, and gives "company name, etc." as the attribute of these words. The user picks up the words classified as nouns among the part-of-speech tags, extracts the words related to the events of trouble / accident from them, and gives "event" as the attribute of these words. The user sets the words and the attributes given to them in the search system 10A. In the search system 10A, the attribute setting unit 174 acquires this setting and records the combination of words and attributes in the storage unit 15.

[0052] Next, the score table creation unit 175 calculates the co-occurrence frequency between words (step S36). For a word having any of the three attributes of "facility name", "company name, etc.", and "event", when an arbitrary word is given, the score table creation unit 175 calculates the number of occurrences of the other word in the range of the previous m words and the next n words. Suppose, as illustrated in FIG. 7, a word "water supply volume" is extracted as an "event". Also, let m = 4 and n = 6. The score table creation unit 175 extracts, for example, the previous 4 words and the next 6 words of "water supply volume" as illustrated in the frame 71 of FIG. 7. Then, the score table creation unit 175 counts the co-occurrence frequency with "water supply volume" within the extracted range. For example, in the extracted range, "equipment A" appears only once. Therefore, the co-occurrence frequency of "equipment A" with respect to "water supply volume" is 1. If "equipment A" appears twice within the range of the previous 4 words and the next 6 words, the numerical value of "equipment A" in the table within the frame 71 becomes 2.

[0053] In parallel with step S36, the user sets an initial value of the weight for each classification of the document. For example, for "business standard" among "business standard", "trouble / accident", "inspection / construction", and "daily business", an initial weight of "1.0" is set, for "trouble / accident", "0.9", for "inspection / construction", "0.7", and for "daily business", "0.3", etc. When the use of the search system 10A is started and the search log is sufficiently accumulated, this weight is updated. For example, if there are many views of the documents classified as "daily business" among the search results, the weight may be updated to a value larger than "0.3". The score table creation unit 175 acquires the set weight and multiplies the weight by the number of occurrences (co-occurrence frequency) of each word described in the table of the frame 71. The score table creation unit 175 multiplies the number of occurrences by different weights according to each document of "business standard", "trouble / accident", "inspection / construction", and "daily business". In the case of this example, <category>Since "Trouble / Accident" is set, the score table creation unit 175 multiplies the appearance frequency of each word by "0.9". The score table creation unit 175 uses the product of the appearance count and the weight as the "score", and records the information of the table T1 shown in the frame 72 of FIG. 7 in the storage unit 15 (step S38). The score table creation unit 175 performs the above processing on all documents, and accumulates the scores while multiplying by the weights according to the documents. For example, as shown in the frame 81 of FIG. 8, for "water supply amount", the total value of the scores of each word is calculated, and the number of columns increases as the number of documents increases. Also, for words having any of the three attributes of "equipment name", "company name, etc.", and "event", and other words other than "water supply amount", the same processing is performed to calculate the total score. For example, as shown in the frame 82 of FIG. 8, for words other than "water supply amount", such as other words indicating "equipment name", "company name, etc.", and "event" (for example, "equipment A", etc.), the same processing is performed to obtain the total score value.

[0054] The score table creation unit 175 calculates the scores between words in this way. When calculating the scores for all documents and words, the score table creation unit 175 summarizes the relationship between the scores between words in three types of pairs: "equipment name" and "event", "company name, etc." and "event", and "company name, etc." and "equipment name". An example of the reference table summarized for "equipment name" and "event" is shown in FIG. 9. The search term proposal unit 17 proposes additional search terms according to the search terms input by the user (step S39). For example, when the user inputs "boiler" as the search term, the search term proposal unit 17 refers to the reference table illustrated in FIG. 9 and obtains "leakage", "fatigue", "crack"... with high scores for "boiler". The search term proposal unit 17 proposes these words to the user as candidates for the second search term. Refer to FIG. 10 here.

[0055] FIG. 10 is a diagram showing the flow of the search process according to the second embodiment. Using FIG. 10, the process of step S39 will be described in detail. For example, when a user uses the search system 10A, the user logs in to the system using account information. The search system 10A obtains account information such as a user ID and performs authentication and the like. The search system 10A records the user ID in the storage unit 15 (step S391). In the storage unit 15, user attribute information (affiliation, title / status, job type, etc.) associated with the user ID is registered in advance. Next, the user starts a search. For example, the user enters "boiler" as a search term. Then, the search query acquisition unit 11 acquires the search term entered by the user (step S392). The search query acquisition unit 11 outputs "boiler" to the search term proposal unit 17 and instructs the proposal of the second search term. When the search term proposal unit 17 acquires "boiler", it refers to the data in the score table stored in the storage unit 15 where the "equipment name" is "boiler" and extracts candidates for the second search term. At this time, the search term proposal unit 17 extracts candidates for the second search term according to the user attributes indicated by the user ID. Specifically, in the score table, words having attributes of "event" or "company name, etc." are registered for "boiler" (attribute = "equipment name"). For example, if the department to which the user indicated by the user ID belongs is a department that performs maintenance of boilers, etc., the search term proposal unit 17 extracts candidates for the second search term in descending order of score value from the combination of "boiler" and "event" (for example, the score table in FIG. 9). Also, for example, if the department to which the user belongs is a department that conducts boiler sales, the search term proposal unit 17 extracts candidates for the second search term in descending order of score value from the combination of "boiler" and "company name, etc.". Thus, it is possible to propose search terms along with the user's interests. Note that the number of search terms to be proposed may be two or more (for example, the second search term, the third search term). Also, the attributes and their combinations are just examples, and other attributes may be set, or a score table may be created with other combinations (for example, "event" and "event", etc.). The search term proposal unit 17 transmits the extracted search term candidates to the user terminal (step S393). On the search screen of the user terminal, candidates for the second search term are displayed as shown in frame 100 of FIG. 10. The user selects "boiler, leakage" and executes the search.The search query acquisition unit 11 acquires "boiler, leakage" (step S394). The search query acquisition unit 11 instructs the search unit 13 to perform a search based on "boiler, leakage". The search unit 13 performs a document search using, for example, a general full-text search algorithm (document score calculation unit 133) (step S395). The search result output unit 14 transmits the search results to the user terminal.

[0056] (Effect) As described above, according to the present embodiment, even for a user with little business experience and insufficient knowledge, when focusing on a single search term, it is possible to present two related words that are relevant to it and scored according to the source. As a result, even users with insufficient skills have an improved possibility of reaching the desired documents that lead to problem-solving and prevention of troubles. Also, even when no search logs are accumulated and a learned model cannot be created, effective search terms can be proposed. Although the second embodiment is a technology targeted until search logs are accumulated, it can also be used after a learned model is created.

[0057] <Third Embodiment> Hereinafter, the search system 10B of the third embodiment will be described with reference to FIGS. 11 to 15. In the first embodiment, the search terms input by the user are converted into vectors in a distributed representation, and other related search terms and documents are searched based on the vector similarity. In the third embodiment, sensor information and alert information output from the plant facilities to be monitored and its monitoring system are acquired instead of the search terms, vectors in a distributed representation of the sensor information and the alert information are calculated, related data is searched, and presented to the user to support efficient monitoring.

[0058] (Configuration) FIG. 11 is a block diagram showing an example of a search system according to the third embodiment. The search system according to the third embodiment can be combined with either the search system 10 of the first embodiment or the search system 10A of the second embodiment. However, FIG. 11 shows a configuration example in which the monitoring information search unit 18, which is the configuration of the third embodiment, is combined with the search system 10 of the first embodiment. Among the components according to the third embodiment, those that are the same as the functional components constituting the search system 10 according to the first embodiment are denoted by the same reference numerals, and their descriptions are omitted. The search system 10B according to the third embodiment includes a monitoring information search unit 18 in addition to the configuration of the search system 10 of the first embodiment. Further, documents related to the monitoring of plant facilities (such as monitoring manuals and accident cases) are registered in the storage unit 15.

[0059] The monitoring information search unit 18 includes a monitoring information acquisition unit 181, a monitoring information analysis unit 182, a sensor classification unit 183, a learning data creation unit 184, a support information output unit 185, and a log management unit 186. The monitoring information acquisition unit 181 acquires monitoring information from a monitoring system 20 that monitors plant facilities and the like. The monitoring information includes sensor information and alert information. The sensor information includes the name and ID of the sensor, the measured value measured by the sensor, the measurement date and time, and the like. The alert information includes the ID, title, and content of the alert. For example, the alert information is information such as "001, Temperature anomaly detected in the boiler, XX day XX hour XX minute XX second... At boiler No. X, temperature sensor X has detected a temperature higher than normal." Here, "001" is the alert ID, "Temperature anomaly detected in the boiler" is the alert title, and the subsequent text data is the alert content.

[0060] The monitoring information analysis unit 182 analyzes the monitoring information acquired by the monitoring information acquisition unit 181 to obtain the sensor name and alert name (title) included in the monitoring information. The monitoring information analysis unit 182 calculates vectors obtained by dispersedly representing the acquired sensor name and alert name using a tool such as word2vec. The monitoring information analysis unit 182 may use, as the dispersed representation, not only the sensor name but also text including the behavior information of the sensor (such as "high" or "low" for a temperature sensor). For example, when acquiring monitoring information indicating that "the temperature measured by the temperature sensor is high", the monitoring information analysis unit 182 converts each of "temperature sensor" and "temperature is high" into a vector, and calculates the vector of their sum as the dispersed representation regarding "the temperature measured by the temperature sensor is high".

[0061] The sensor classification unit 183 calculates vectors obtained by dispersedly representing the sensor names included in the documents related to the monitoring of the plant facilities registered in the storage unit 15 using a tool such as word2vec. The sensor classification unit 183 classifies the sensor names into any of a predetermined number of categories based on the similarity of the vectors of the sensor names, for example, by the k-nearest neighbor method or the like. For example, the plant facilities have a plurality of systems, and a plurality of sensors are provided in each system. The documents related to the monitoring of the plant facilities include, for example, descriptions for checking the measured values by a plurality of sensors in the same system together, and descriptions for taking measures based on the measured values of the sensors in a certain system when an abnormality occurs in that system. When grouping those with similar vectors obtained by dispersed representation, sensors in the same system are more likely to be classified into the same group as those with a higher similarity than sensors in different systems. If sensors in the same system belong to the same group, it is possible to provide information for checking other sensors in the same system as related sensor information described later.

[0062] The learning data creation unit 184 creates learning data for creating a learned model based on the log records of the documents actually browsed by the user for the related documents presented in the support information described later. This learning data is provided to the learning unit 12 described in the first embodiment.

[0063] When an abnormality in the measured value measured by the sensor or alert information occurs, the support information output unit 185 presents to the user an HTML-formatted email or dashboard including support information regarding the abnormality or alert information. For example, when an abnormality is detected by a temperature sensor, the support information presents information on sensors closely related to the temperature sensor and information on documents closely related to the abnormality of the temperature sensor.

[0064] The log management unit 186 records in the log the sensor name included in the support information output by the support information output unit 185, or information on sensors closely related to the behavior of the measured values measured by the sensor name and the sensor, information on documents closely related, information on the sensors actually referred to by the user, and information on the documents actually browsed by the user. This log is recorded in the storage unit 15.

[0065] FIG. 12 is a diagram showing an example of support information according to the third embodiment. As shown in the figure, the support information presents the content of the abnormality, related sensors (XX sensor 1, XX sensor 2) related to the temperature sensor A that detected the abnormality, and documents (document 1.doc, document 2.pdf) related to the temperature sensor A. The support information is sent as an email or displayed on the dashboard of an application or the like used by the user. The user is guided to consider countermeasures with reference to related sensors and related documents for the detected abnormality. When the user presses the relevant location of the related sensor or related document of the support information, the behavior of the selected sensor is displayed or the selected document is displayed. When the user performs such an operation, this operation is notified to the search system 10B, and the log management unit 186 records the notified operation in the log. For example, when the user selects and presses XX sensor 1, the time of the press, temperature sensor A, the temperature measured by temperature sensor A, information on XX sensor 1 actually checked by the user, etc. are recorded in the log. The same applies when the user selects document 1.doc.

[0066] (Operation) (Learning phase) FIG. 13 is a flowchart showing an example of the processing in the learning phase according to the third embodiment. First, the sensor classification unit 183 acquires sensor names and the like (step S41). For example, the sensor classification unit 183 acquires all sensor names and all alert names included in documents related to the monitoring of all plant facilities prepared in advance. Next, the sensor classification unit 183 converts all sensor names and the like into a distributed representation (step S42). For example, the sensor classification unit 183 calculates vectors in a distributed representation for all sensor names and all alert names using word2vec or the like. The sensor classification unit 183 may calculate vectors in a distributed representation including behavior information such as "high" or "low" for the measured values of each sensor. Next, the sensor classification unit 183 classifies the sensor names and the like (step S43). The sensor classification unit 183 classifies the distributed sensor names and the like into L groups using a clustering method such as the k-nearest neighbor method. This classification information is referred to when the support information output unit 185 presents sensors related to the sensor when an abnormality occurs in the sensor measurement value.

[0067] Next, the learning data creation unit 184 acquires pairs of sensor and other information and browsing information from the logs recorded by the log management unit 186 (step S44). The logs include information on the sensors that directly caused the abnormality detection presented in the support information, sensors related to the sensors, and documents related to the sensors. The learning data creation unit 184 creates pairs of sensor names and related documents actually accessed by the user one by one based on the log records, and acquires pairs of sensor and other information and browsing information from the logs.

[0068] Next, the learning data creation unit 184 generates learning data (step S45). The learning data creation unit 184 generates learning data for document search. Based on the result of step S4 (document category), the learning data creation unit 184 attaches information indicating the category into which the document is classified to the document in the pair of sensor names and documents acquired in step S44 (correct label). The learning data creation unit 184 performs this process for all pairs of sensor names and documents to create learning data. The learning data creation unit 184 outputs this learning data to the learning unit 12 described in the first embodiment.

[0069] Next, the learning unit 12 creates a learned model in the same manner as in the first embodiment (step S46). The learning unit 12 replaces the relationship between the sensor name and the actually accessed related document with the relationship between the search term and the actually viewed document among the search results, and creates (updates) a learned model using the learning data created in step S45. As a result, a learned model is constructed such that when a sensor name is input as a search term, documents classified into the same category as the related documents actually viewed by the user preferentially appear at the top of the search results.

[0070] (Output process of support information) FIG. 14 is a flowchart showing an example of the output process of support information according to the third embodiment. First, the monitoring information acquisition unit 181 acquires monitoring information (sensor information, alert information) (step S51). For example, the monitoring information includes information other than the related sensors and related documents among the information exemplified in FIG. 12. Next, the monitoring information analysis unit 182 converts the monitoring information into a distributed representation (step S52). For example, the monitoring information analysis unit 182 acquires the sensor name from the monitoring information and calculates a vector in which the sensor name (for example, temperature sensor A) is represented in a distributed manner. Alternatively, the monitoring information analysis unit 182 acquires the sensor name included in the monitoring information and the words indicating its behavior (temperature X1 ° C), and calculates a vector in which they are represented in a distributed manner. The monitoring information analysis unit 182 outputs the calculated vector to the sensor classification unit 183. Further, the monitoring information analysis unit 182 outputs the sensor name or the words indicating the sensor name and its behavior before conversion to the search query acquisition unit 11 to instruct the search for documents.

[0071] Next, the sensor classification unit 183 selects related sensors (step S53). For example, the sensor classification unit 183 determines to which group the distributed representation vector converted in step S52 belongs among the groups classified in step S43, and selects the sensor names belonging to the same group. At this time, if there are a large number of sensor names to be selected, the sensor classification unit 183 may select a predetermined number in order from those similar to the vector obtained in step S52. Further, the sensor classification unit 183 may select sensor names with a past reference count of a predetermined number or more from among the sensor names belonging to the same group based on the past reference count record logged in the log. The sensor classification unit 183 outputs the selected sensor names to the support information output unit 185.

[0072] On the other hand, the search query acquisition unit 11 that has acquired the sensor name or the like outputs the sensor name or the like to the search unit 13. The search unit 13 performs a search based on the sensor name or the like by the same process as in the first embodiment (FIG. 3), and selects information (= related documents) for the top n of the search results (step S54). The search unit 13 outputs the information of the selected search results to the support information output unit 185.

[0073] Next, the support information output unit 185 creates and outputs support information (step S55). Using the monitoring information acquired in step S51, the information of the related sensors selected in step S53, and the information of the related documents (search results) selected in step S54, the support information output unit 185 creates the support information illustrated in FIG. 12 and outputs it via e-mail or the like.

[0074] Next, the log management unit 186 records the operations on the support information in a log (step S56). In the example of FIG. 12, when XX sensor 1 is selected, the log management unit 186 records XX sensor 1 as the actually referenced sensor in the log in association with temperature sensor A or temperature sensor A and temperature X1°C. Similarly, when document 1 is selected, the log management unit 186 records document 1 as the actually viewed document in the log in association with temperature sensor A or temperature sensor A and temperature X1°C.

[0075] Next, the learning data creation unit 184 reads out the pair of the sensor name etc. recorded in the log and the viewed document and outputs it to the search log management unit 16. The search log management unit 16 records the pair of the sensor name etc. and the viewed document in the search log (step S57). Thereby, when the learning process in FIG. 2 is performed next time, the record of the actually viewed document can be reflected when notifying the alert information.

[0076] Next, the log management unit 186 reads out the pair of the sensor name etc. recorded in the log and the related sensors that have been referenced and counts the number of references to the referenced sensors (step S58). For example, if the number of times the information of XX sensor 1 has been actually referenced for temperature sensor A so far is 1 time, the log management unit 186 updates this value to 2 times and records this value in the storage unit 15 in association with temperature sensor A and XX sensor 1. This value is referenced when the related sensor is selected next time (step S53).

[0077] (Effect) When certain monitoring information is obtained and it is necessary to present information such as that of sensors related to the monitoring items, a method of manually performing the settings of the relevant sensors can be considered. However, with such a method, the work becomes enormous and it takes an enormous amount of time until the setting work is completed. On the other hand, according to the search system 10B, by using a distributed representation for sensors and alert information described in natural language, relevant sensors can be presented to the user without prior structuring. Also, by performing document search using the sensor name or alert name as a search term in the search method according to the first embodiment, useful information for dealing with alerts and the like can be provided to the user.

[0078] FIG. 15 is a diagram showing an example of the hardware configuration of the search system according to the embodiment. The computer 900 includes a CPU 901, a main memory device 902, an auxiliary storage device 903, an input / output interface 904, and a communication interface 905. The above-described search systems 10, 10A, and 10B are implemented in the computer 900. And each of the above-described functions is stored in the auxiliary storage device 903 in the form of a program. The CPU 901 reads the program from the auxiliary storage device 903 and expands it in the main memory device 902, and executes the above processing according to the program. Also, the CPU 901 secures a storage area in the main memory device 902 according to the program. Further, the CPU 901 secures a storage area in the auxiliary storage device 903 for storing data being processed according to the program.

[0079] A program for realizing all or part of the functions of the search systems 10, 10A, and 10B may be recorded on a computer-readable recording medium, and the program recorded on this recording medium may be read into a computer system and executed to perform processing by each functional unit. Here, the "computer system" shall include hardware such as an OS and peripheral devices. Also, the "computer system" shall include a homepage providing environment (or display environment) if the WWW system is being used. Also, the "computer-readable recording medium" refers to portable media such as CDs, DVDs, USBs, and storage devices such as hard disks built into a computer system. Also, when this program is distributed to the computer 900 via a communication line, the computer 900 that has received the distribution may expand the program in the main storage device 902 and execute the above processing. Also, the above program may be for realizing a part of the functions described above, or may further be capable of being realized in combination with a program already recorded in the computer system for the functions described above. Also, the above-described search systems 10, 10A, and 10B may be configured as a system consisting of a plurality of computers 900 such as client terminals and server terminals. In this case, a configuration may be adopted in which some functions (for example, the search query acquisition unit 11 and the search result output unit 14) are provided to the client terminal, and other functions are provided to a server terminal capable of communicating with the client terminal. Also, in the case of such a system, it may be configured to execute each function in the server terminal in accordance with a search request from the client terminal.

[0080] As described above, several embodiments according to the present disclosure have been described, but all of these embodiments are presented as examples and are not intended to limit the scope of the invention. These embodiments can be implemented in various other forms, and various omissions, replacements, and changes can be made without departing from the gist of the invention. These embodiments and their modifications are included in the scope and gist of the invention, and are also included in the invention described in the claims and its equivalent scope.

[0081] <Supplementary Note> The search systems, search methods, and programs described in the embodiments can be understood as follows, for example.

[0082] (1) The search systems 10, 10A, 10B according to the first aspect include a search query acquisition unit 11 that acquires a search term, a search unit (document score calculation unit 133) that performs document search for the search term acquired by the search query acquisition unit, and a storage unit 15 that stores in association information on the viewed documents among the documents searched by the search unit and information on the search term in the document search. The search unit performs the document search based on the information on the documents in the storage unit corresponding to the information on the search term acquired by the search query acquisition unit. When performing a document search by specifying a search term, since the actually viewed documents are provided as search results, search results with high utility value for the user can be provided.

[0083] (2) The search systems 10, 10A, 10B according to the second aspect are the search systems 10, 10A, 10B in (1), and further include an output unit that preferentially outputs the documents to be searched by the search unit (document score calculation unit 133) based on the information on the documents in the storage unit corresponding to the information on the search term acquired by the search query acquisition unit. Since the actually viewed documents are preferentially output, documents with high utility value for the user can be provided as search results.

[0084] (3) The search systems 10, 10A, and 10B according to the third aspect are the search systems 10, 10A, and 10B in (2), and include a search term analysis unit 132 that calculates a search term vector obtained by distributing the search term, a vector of the distributed word, and a learned model that has learned the relationship between the vector of the word and the category to which the document actually viewed among the search results obtained when the word is used as a search term, and a weight calculation unit (relevance score calculation unit 134) that calculates a weighting for each of a plurality of the categories based on the search term vector and the learned model, and an evaluation value calculation unit (relevance score calculation unit 134) that calculates a second evaluation value by multiplying the first evaluation value of the document in the search unit by the weighting of the category to which the document belongs for each of the documents searched by the search unit. The search unit calculates a first evaluation value for the searched document, and the output unit outputs a search result in which the searched documents are sorted based on the second evaluation value. Since the score of the document is calculated by weighting more heavily the documents belonging to the same category as the document actually viewed by the user among the search results, and the documents with high scores are provided as search results, it is possible to provide the search results with documents having high utility value for the user.

[0085] (4) The search systems 10, 10A, and 10B according to the fourth aspect are the search systems 10, 10A, and 10B in (3), and further include a search term setting unit 131 that adds, as search keywords, words whose similarity to the search term vector is equal to or greater than a predetermined threshold. Thereby, it is possible to perform a search including words similar to the search term. Therefore, even when the knowledge of the user performing the search is uncertain, the possibility of finding the desired document is improved.

[0086] (5) The search systems 10, 10A, and 10B according to the fifth aspect are the search systems 10, 10A, and 10B in (3) to (4), and further include a similar word list creation unit 123 that creates a similar word list in which a second word whose similarity to the vector obtained by distributing the first word is equal to or greater than a predetermined threshold is associated with the first word. Accordingly, when a user specifies a search term, other keywords similar thereto can be automatically set. Also, since the creation of the similar word list can be automatically performed, there is no need to maintain the similar word list.

[0087] (6) The search systems 10, 10A, 10B according to the sixth aspect are the search systems 10, 10A, 10B of (3) to (5), and include a document analysis unit 121 that calculates vectors obtained by dispersedly representing each of the documents to be searched and the words included in the documents, a document classification unit 122 that classifies the documents such that those with similar document vectors belong to the same category, and a learned model creation unit 124 that learns the relationship between a search term, the category based on the classification result of the document classification unit of the documents actually viewed from among the search results of the search term, and creates the learned model. Accordingly, for each document searched by a general search engine, a weight can be assigned according to the category of the document.

[0088] (7) The search system 10A according to the seventh aspect is the search system 10A of (1) to (6), and further includes a morphological analysis unit (reading analysis unit 172) that performs morphological analysis on the text data included in the document, a score table creation unit 175 that counts the co-occurrence frequency of one of the words segmented by the morphological analysis unit with other words arranged before and after the word, multiplies the counted co-occurrence frequency by a weight according to the type of the document, and creates a score table in which an evaluation value between words is set, and a search term proposal unit 17 that refers to the score table for the search term, acquires words with a high evaluation value for the search term, and outputs them as candidates for a second search term. Accordingly, even when a learned model cannot be created, other keywords highly relevant to the search term specified by the user can be proposed.

[0089] (8) The search system 10A according to the eighth aspect is the search system 10A of (7), wherein the search term proposal unit 17 acquires, from the score table, words that match the attribute of the user who input the search term. Thereby, candidates for search terms that match the user's needs can be proposed.

[0090] (9) The search system 10A according to the ninth aspect is the search system 10A of (7) to (8), wherein the search term proposal unit outputs, as candidates for the second search term, according to the storage status of information on the viewed documents and information on the search terms in the storage unit. For example, even when the storage status of information on the viewed documents and information on the search terms in the storage unit is small (when there is little data), the user convenience can be improved by the search term proposal unit outputting candidates for the second search term.

[0091] (10) The search system 10B according to the tenth aspect is the search system 10B of (1) to (9), further comprising a monitoring information acquisition unit 181 that acquires monitoring information including measured values and alerts of sensors output by the monitoring system 20, a monitoring information analysis unit 182 that acquires a sensor name or an alert name from the monitoring information and outputs the sensor name or the alert name as the search term to the search query acquisition unit, and a support information output unit 185 that outputs the search result when the sensor name or the alert name is used as the search term, together with the monitoring information. Thereby, documents related to the measured values and alerts of the sensors output by the monitoring system can be searched and presented, improving the convenience for the monitor.

[0092] (11) The search system 10B according to the 11th aspect is the search system 10B of (1) to (9), and includes a monitoring information acquisition unit 181 that acquires monitoring information including measured values and alerts of sensors output by the monitoring system 20, a monitoring information analysis unit 182 that acquires a sensor name or an alert name from the monitoring information and calculates a monitoring information vector in which the sensor name or the alert name is expressed in a distributed manner, and a support information output unit 185 that outputs another sensor name having a vector in a distributed expression similar to the monitoring information vector together with the monitoring information. Thereby, it is possible to present sensors related to the measured values and alerts of the sensors output by the monitoring system, and the convenience of the monitor can be improved.

[0093] (12) The search systems 10, 10A, and 10B according to the 12th aspect are search systems including a server terminal capable of communicating with a client terminal. The client terminal includes a search query acquisition unit that acquires a search term. The server terminal includes a search unit that performs a document search for the search term acquired by the search query acquisition unit, and a storage unit that stores information related to the viewed documents among the documents retrieved by the search unit in association with information related to the search term in the document search. The search unit performs the document search based on the information related to the document in the storage unit corresponding to the information related to the search term acquired by the search query acquisition unit.

[0094] (13) The search method according to the 13th aspect includes a search query acquisition step of acquiring a search term, a search step of performing a document search for the search term acquired in the search query acquisition step, and a storage step of storing information related to the viewed documents among the documents retrieved in the search step in association with information related to the search term in the document search. In the search step, the document search is performed based on the information related to the document corresponding to the information related to the search term acquired in the search query acquisition step.

[0095] (14) The program according to the 14th aspect causes a computer 900 to execute a search query acquisition step of acquiring a search term, a search step of performing a document search for the search term acquired in the search query acquisition step, and a storage step of storing, in association with each other, information regarding a viewed document among the documents searched in the search step and information regarding the search term in the document search. In the search step, a process of performing the document search is executed based on information regarding the document corresponding to the information regarding the search term acquired in the search query acquisition step.

Explanation of Signs

[0096] 10, 10A... Search system 11... Search query acquisition unit 12... Learning unit 121... Document analysis unit 122... Document classification unit 123... Similar word list creation unit 124... Trained model creation unit 13... Search unit 131... Search term setting unit 132... Search term analysis unit 133... Document score calculation unit 134... Relevance score calculation unit 14... Search result output unit 15... Storage unit 16... Search log management unit 17... Search term proposal unit 171... Classification setting unit 172... Reading analysis unit 173... Reduction unit 18... Monitoring information search unit 181... Monitoring information acquisition unit 182... Monitoring information analysis unit 183... Sensor classification unit 184... Learning data creation unit 185... Support information output unit 186... Log management unit 174... Attribute setting unit 175 ··· Score table creation unit 900 ··· Computer 901 ··· CPU 902 ··· Main memory device 903 ··· Auxiliary storage device 904 ··· Input / output interface 905 ··· Communication interface< / category> < / category> < / category> < / category> < / category>

Claims

1. A search query acquisition unit that acquires a search term; A search unit that performs a document search for the search term acquired by the search query acquisition unit; A storage unit that stores, in association with each other, probability information for each category indicating the probability of which category a viewed document among the documents searched by the search unit belongs to, and a distributed representation of the search term in the document search; Comprising: The search unit performs the document search based on the probability information for each category in the storage unit corresponding to the distributed representation of the search term acquired by the search query acquisition unit. A search system.

2. An output unit that preferentially outputs documents in the category with a high probability based on the probability information for each category in the storage unit corresponding to the distributed representation of the search term acquired by the search query acquisition unit, among the documents to be searched by the search unit; The search system according to claim 1, further comprising:

3. A search term analysis unit that calculates a search term vector obtained by distributing the search term; A learned model that has learned the relationship between the vector of a distributed word and the category to which the actually viewed document belongs among the search results obtained when that word is used as a search term, and the search term vector; based on this, a weight calculation unit that calculates weights for each of a plurality of the categories; An evaluation value calculation unit that calculates a second evaluation value by multiplying the first evaluation value of each of the documents searched by the search unit by the weight of the category to which the document belongs; Comprising: The search unit calculates a first evaluation value for the searched document; The output unit outputs a search result in which the searched documents are sorted based on the second evaluation value. The search system according to claim 2.

4. A search term setting unit that adds a word whose similarity to the search term vector is equal to or greater than a predetermined threshold to the search term; The search system according to claim 3, further comprising:

5. A similar word list creation unit that creates a similar word list in which a second word whose similarity to the vector obtained by distributing a first word is equal to or greater than a predetermined threshold is recorded in association with the first word; The search system according to claim 3 or claim 4, further comprising:

6. A document analysis unit that calculates vectors obtained by distributing each of the documents to be searched and the words included in the documents; A document classification unit that classifies the documents such that those with similar document vectors belong to the same category; A trained model creation unit that learns the relationship between a search term and the category based on the classification result of the document classification unit of the document actually viewed from among the search results of the search term, and creates the trained model. The search system according to any one of claims 3 to 5, further comprising **Claim 7** A morphological analysis unit that performs morphological analysis on the text data included in the document; A score table creation unit that counts the co-occurrence frequency of one of the words segmented by the morphological analysis unit with other words arranged before and after the word, multiplies the counted co-occurrence frequency by a weight corresponding to the type of the document, and creates a score table in which an evaluation value between words is set; A search term proposal unit that refers to the score table for the search term, acquires a word having a high evaluation value for the search term, and outputs it as a candidate for a second search term; The search system according to any one of claims 1 to 6, further comprising **Claim 8** The search term proposal unit acquires, from the score table, a word that matches the attribute of the user who input the search term. The search system according to claim 7. **Claim 9** The search term proposal unit outputs, as a candidate for the second search term, when there is no record of the search term and the document actually viewed among the search results by the search term stored in the storage unit. The search system according to claim 7 or 8, characterized in that **Claim 10** A monitoring information acquisition unit that acquires monitoring information including measurement values and alerts of sensors output by a monitoring system; A monitoring information analysis unit that acquires a sensor name or an alert name from the monitoring information, and outputs the sensor name or the alert name as the search term to the search query acquisition unit; An assistance information output unit that outputs the result of the document search when the sensor name or the alert name is used as the search term, together with the monitoring information; The search system according to any one of claims 1 to 9, further comprising **Claim 11** A monitoring information acquisition unit that acquires monitoring information including measurement values and alerts of sensors output by a monitoring system; A monitoring information analysis unit that acquires a sensor name or an alert name from the monitoring information, and calculates a monitoring information vector in which the sensor name or the alert name is expressed in a distributed manner; An assistance information output unit that outputs, together with the monitoring information, another sensor name having a vector of a distributed expression similar to the monitoring information vector; The search system according to any one of claims 3 to 9, further comprising

12. A search system comprising a server terminal capable of communicating with a client terminal, wherein the client terminal comprises a search query acquisition unit that acquires a search term, and the server terminal comprises a search unit that performs a document search for the search term acquired by the search query acquisition unit, and a storage unit that stores, in association with each other, probability information for each category indicating the probability of which category a viewed document among the documents retrieved by the search unit belongs to, and a distributed representation of the search term in the document search, and the search unit performs the document search based on the probability information for each category in the storage unit corresponding to the distributed representation of the search term acquired by the search query acquisition unit. The search system is characterized by this.

13. A computer has a search query acquisition step of acquiring a search term, a search step in which the computer performs a document search for the search term acquired in the search query acquisition step, and a storage step in which the computer stores, in association with each other, probability information for each category indicating the probability of which category a viewed document among the documents retrieved in the search step belongs to, and a distributed representation of the search term in the document search, and in the search step, the computer performs the document search based on the probability information for each category corresponding to the distributed representation of the search term acquired in the search query acquisition step. Search method.

14. A program for causing a computer to execute a search query acquisition step of acquiring a search term, a search step of performing a document search for the search term acquired in the search query acquisition step, and a storage step of storing, in association with each other, probability information for each category indicating the probability of which category a viewed document among the documents retrieved in the search step belongs to, and a distributed representation of the search term in the document search, and in the search step, causing the computer to execute a process of performing the document search based on the probability information for each category corresponding to the distributed representation of the search term acquired in the search query acquisition step. Program.

Images