JP7251625B2 - Method and system for searching and displaying relevant documents - Google Patents

Description

The present invention relates to methods and systems for retrieving and displaying relevant documents.

Conventionally, a method of searching for and displaying documents related to a search query (related documents) is known. For example, in Japanese Patent Application Laid-Open No. 2018-10482 (Patent Document 1), a set of pairs (correct information) of a search query and a set of correct documents that are search target documents conceptually matching the search query are used. , a document concept search method for searching for documents to be searched that conceptually match a search query entered by a user. According to the document concept retrieval method, retrieval accuracy can be improved by using the correct information.

JP 2018-10482 A

Natural language processing in the field of artificial intelligence applies machine learning to convert the meaning of words into vector representations (distributed representations). Specifically, based on the distribution hypothesis that the meaning of a word is determined from the distribution of words that appear around that word, a neural network learns a large amount of sentence information and generates a vector representing the meaning of each word. Generate space.

The vector-to-vector distance between two linguistic units that have similar meanings to each other is relatively short. According to the document retrieval method based on vector representation, documents containing more linguistic units having meanings close to the retrieval query are ranked higher. Therefore, according to the document retrieval method, even documents that do not contain any character strings that match the retrieval query may be ranked high in the retrieval results. In such cases, the user may not be able to understand why the document was retrieved.

In the document concept search method disclosed in Patent Document 1, search target documents ranked in descending order of similarity between a search query concept vector and a search target document concept vector are displayed as search results. However, no consideration is given to presenting the basis of search results to the user.

SUMMARY OF THE INVENTION The present invention has been made to solve such problems, and an object thereof is a method of retrieving and displaying at least one related document related to a search query from a database in which a plurality of documents are stored. , to show the user the basis of the search results.

A method for retrieving and displaying related documents according to a first aspect of the present invention retrieves and displays at least one related document related to a search query from a database containing a plurality of documents. The method includes a searching step and a displaying step. The retrieving step calculates a distance between the vector representation of each of the plurality of documents and the vector representation of the search query using a vector space that converts any linguistic unit into a vector representation, and retrieves at least one related document according to the distance. Search for The displaying step displays each of the at least one related document. The displaying step includes displaying the linguistic units in a display mode according to the degree of relevance between the vector representation of each of the linguistic units included in the related document and the vector representation of the search query.

A system for retrieving and displaying related documents according to a second aspect of the present invention retrieves and displays at least one related document related to a search query from a plurality of documents. The system includes a database and a search processor. A plurality of documents are stored in the database. The search processing unit calculates the distance between the vector representation of each of the plurality of documents and the vector representation of the search query using a vector space that converts any linguistic unit into a vector representation, and determines at least one relation according to the distance. Search for documents. A search processor displays each of the at least one related document. The search processing unit displays the linguistic units in a display manner according to the degree of relevance between the vector representation of each of the linguistic units included in the related document and the vector representation of the search query.

According to the method and system for retrieving and displaying related documents according to the present invention, each of the plurality of linguistic units included in the related document and the vector representation of the search query are used to determine the relationship between the linguistic unit and the search query. By displaying the linguistic unit in a display mode according to the degree of relevance, it is possible to show the user the basis of the search result.

1 is an external view of an analysis case search system, which is an example of a system for searching and displaying related documents according to an embodiment; FIG. FIG. 2 is a functional block diagram showing the configuration of the analysis case search system of FIG. 1; It is a figure which shows an example of an analysis report. 3 is a flowchart for explaining the flow of learning processing performed by the learning processing unit of FIG. 2; 3 is a flowchart for explaining the flow of search processing performed by the search processing unit in FIG. 2; 3 is a diagram showing an example of a search result window displayed on the display by the display control unit shown in FIG. 2; FIG. FIG. 7 is a diagram showing the contents of a related document displayed in a search result window when the related document in FIG. 6 is selected; FIG. 10 is a diagram showing how each of a plurality of words included in a related document is highlighted using the degree of contribution of the word as the relevance between the word included in the related document and the search query; FIG. 10 is a diagram showing the contents of related documents displayed in the search result window when the highlighted linguistic unit is a sentence; FIG. 10 is a diagram showing the contents of related documents displayed in the search result window when the highlighted linguistic unit is a paragraph; FIG. 2 is a functional block diagram showing the configuration of an analysis case search system, which is an example of a system for searching and displaying related documents according to the embodiment, when the analysis case search system is connected to a plurality of client terminals via a network; be.

BEST MODE FOR CARRYING OUT THE INVENTION Below, embodiments of the present invention will be described in detail with reference to the drawings. In the following description, the same reference numerals are assigned to the same or corresponding parts in the drawings, and the description thereof will not be repeated in principle.

FIG. 1 is an external view of an analysis case search system 100, which is an example of a system for searching and displaying related documents according to an embodiment. As shown in FIG. 1, the analysis case retrieval system 100 includes a calculator 10, a display 60, a keyboard KB1, and a mouse MS1. A display 60, a keyboard KB1, and a mouse MS1 are connected to the computer 10. FIG.

The display 60 displays a search window Wn1 and a cursor Cr. The user operates the cursor by operating the mouse MS1. The user inputs a search query into the search window Wn1 by operating the keyboard KB1. In FIG. 1, a user enters a character string "bromic acid" as a search query into a search window Wn1 in order to search for documents describing analytical equipment suitable for analysis of bromic acid, analysis methods, analysis conditions, and the like. The case is shown.

(Configuration of Analysis Case Search System 100)
FIG. 2 is a functional block diagram showing the configuration of the analysis case search system 100 of FIG. As shown in FIG. 2 , the analysis case search system 100 includes a learning processing section 20 and a search processing section 40 .

The analysis case search system 100 searches a plurality of documents included in the database 30 based on the search query input by the user, and searches related documents related to the search query. The database 30 contains document data in the field of analysis cases. Document data in the field of analysis cases include, for example, analysis reports, analysis-related papers, and analysis-related patent documents. The analysis report is, for example, a document related to compound analysis as shown in FIG. 3, and includes information related to the analysis method of the compound to be analyzed, information related to the analysis equipment, information related to analysis conditions, and the like.

(Configuration of learning processing unit 20)
Referring again to FIG. 2 , learning processing unit 20 includes morphological analysis unit 21 , vector generation unit 23 , association degree learning unit 25 , corpus 27 , and database 30 . The corpus 27 is linguistic material in which a large amount of document data related to the field of analysis cases is systematized and accumulated for machine learning using natural language processing.

The morphological analysis unit 21 subdivides all the document data stored in the corpus 27 by dividing them into the smallest meaningful linguistic units (morphemes or words) by morphological analysis.

The vector generation unit 23 generates a vector space for converting a word into a vector representation by machine learning of the morphological analysis result of the morphological analysis unit 21 . The vector space is generated in the process of machine learning for models formed by neural networks. Examples of the model include a CBOW (Continuous Bag-of-Words) model that infers a central word (target) from surrounding words (context), or a skip-gram model that infers peripheral words from a central word. can be mentioned.

The vector generation unit 23 generates word vectors, which are semantic representations of words. The vector generation unit 23 generates a sentence vector representing the feature of the sentence and a document vector representing the feature of the document from the sum of feature amounts of words included in the document. Note that subdivision of the document data may be performed in arbitrary language units. Any type of linguistic unit may include letters, morphemes, words, sentences, or paragraphs. A sentence is the smallest unit of meaningful sentences, and a document consists of a plurality of sentences. Japanese sentences are separated by periods, and sentences written in languages such as English are separated by periods (full stops). By the above morphological analysis, the document is divided into sentences by punctuation marks or periods, and a sentence vector is generated based on the above machine learning for the result of the morphological analysis. A sentence vector may be generated for each paragraph by dividing the document into paragraphs.

The word vectors, sentence vectors, and document vectors generated by the vector generation unit 23 are sent to the association degree learning unit 25 . The association level learning unit 25 includes a word vector learning unit 25a, a word-sentence learning unit 25b, and a word-document learning unit 25c.

Since the distance between vectors in the vector space of words represents the semantic similarity between words, the word vector learning unit 25a calculates the semantic relevance between words in the vector space and the vector distance between words. do. Similarly, the word-sentence learning unit 25b calculates the degree of relevance between the word and the sentence in the vector space and the vector distance, and the word-document learning unit 25c calculates the degree of relevance between the word and the document in the vector space, and Compute vector distances. Note that the distance between two vector representations can be, for example, cosine distance (cosine similarity). The smaller the cosine distance between the two vector representations, the closer the meaning of the two linguistic units respectively represented by the two vectors.

Calculation results in the word vector learning unit 25a, the word-sentence learning unit 25b, and the word-document learning unit 25c are stored in the database 30 having a multidimensional vector space with each coordinate axis being a word, a sentence, and a document. Stored with the data as word vectors, word-sentence vectors, and word-document vectors.

By accumulating data centered on a specific field, such as analysis cases, in the corpus 27, a vector representation more suitable for that field can be obtained. As a result, it is possible to improve the search accuracy for sentences in a specific field. The corpus 27 is composed of an internal corpus that stores in-house reports, technical reports, application news, etc., and an external corpus that collects data published on the web such as Wikipedia (registered trademark). may Since the external corpus is intended to improve the learning of vector representations, it is possible to prevent a decrease in search speed by excluding the external corpus from the search target.

FIG. 4 is a flowchart for explaining the flow of learning processing performed by the learning processing unit 20 of FIG. A step is simply denoted as S below. As shown in FIG. 4, in S11, the morphological analysis unit 21 divides the document data (learning data) stored in the corpus 27 into a plurality of words by morphological analysis using an existing dictionary.

In S13 following S11, the vector generation unit 23 generates word vectors, which are semantic representations of words, by machine learning based on the results of the morphological analysis in S11, and also generates sentence vectors representing the characteristics of sentences and text vectors representing the characteristics of the sentences. A document vector, which is a feature vector, is generated from the sum of feature amounts of words included in the document.

In S15 following S13, the word-sentence learning unit 25b calculates the degree of relevance and the vector distance between the word and the sentence in the vector space. In S17 following S15, the word-document learning unit 25c calculates the degree of relevance and vector distance between the word and the document in the vector space.

In S19 following S17, the relevance learning unit 25 converts the document data of the corpus 27 used as learning data and the calculation results of S13, S15, and S17 into word vectors, word-sentence vectors, and word-document vectors. Store in database 30 .

(Configuration of search processing unit 40)
Referring to FIG. 2 again, search processing unit 40 includes input unit 1 , analysis unit 11 , feature extraction unit 13 , search unit 15 , display control unit 17 and output unit 5 . A user inputs a search query to the input unit 1 of the search processing unit 40 . Search queries include, for example, analysis-related search keywords, analysis-related compound names, and analysis-related analyte names. The input unit 1 includes a keyboard KB1 and a mouse MS1. Output unit 5 includes display 60 .

The analysis unit 11 performs morphological analysis on the search query input to the input unit 1 based on a predefined search dictionary, and divides the search query into words. The feature extraction unit 13 uses the vector space generated by the learning processing unit 20 to calculate the vector representation of the search query.

The search unit 15 uses the vector representation of the search query acquired from the feature extraction unit 13 to search the database 30 for related documents related to the search query. The search unit 15 searches the database 30 for related documents whose distance from the vector representation of the search query is smaller than a threshold. The search unit 15 outputs to the display control unit 17 a search result in which the shorter the distance between the vector representation of the search query and the vector representation of each of the retrieved related documents, the related documents are ranked higher. . The display control unit 17 controls the output unit 5 so that related documents are displayed in the order of ranking by the search unit 15 . The output unit 5 displays information on the display 60 according to the control result of the display control unit 17 .

FIG. 5 is a flowchart for explaining the flow of search processing performed by the search processing unit 40 of FIG. As shown in FIG. 5, the input unit 1 receives a search query input by the user in S21. In S23 following S21, the analysis unit 11 performs morphological analysis on the search query to divide the search query into minimum unit morphemes (words). In S25 following S23, the feature extraction unit 13 calculates a vector representation of the search query using the result of the morphological analysis of the search query and the vector space generated by the learning processing unit.

In S27 following S25, the search unit 15 searches the database 30 storing learning data vectorized by learning the corpus 27 for related documents related to the search query. In S27, documents related or highly relevant to the search query are retrieved as the search results. A document highly relevant to a search query is a document having a high degree of association between a word and a document in the vector space and a short vector distance, which is obtained by calculating the association between the word and the document in advance. The search unit 15 ranks the searched related documents in ascending order of vector distance.

In S29 following S27, the display control unit 17 displays the plurality of retrieved related documents on the output unit 5 based on the ranking by the retrieval unit 15. FIG. A user can determine a related document to browse according to the ranking of relevance between each of the at least one retrieved related document and the search query.

FIG. 6 is a diagram showing an example of a search result window Wn2 displayed on the display 60 by the display control section 17 of FIG. As shown in FIG. 6, the related documents D1 to D4 are displayed in order with the ranks 1 to 4 in the search result window Wn2. A hyperlink is set for each of the related documents D1 to D4. In FIG. 6, among the plurality of related documents retrieved from the database 30, the related document D1 is most relevant to the search query "bromic acid".

Again referring to FIG. 5, when the user operates the cursor Cr to select the hyperlink of the related document D1 in FIG. View the content of related documents. FIG. 7 shows the contents of the related document D1 displayed in the search result window Wn2 when the related document D1 of FIG. 6 is selected.

As shown in FIG. 7, the display control unit 17 controls the vector representation of the search query and for each of a plurality of words included in the related document D1, according to the distance between the vector representation of the word and the vector representation of the search query. The word is highlighted by changing the color of the surrounding area of the word in the search result window Wn2. The distance between the vector representation of the words included in the related document D1 and the vector representation of the search query is a range R1 greater than the distance Ds3, a range R2 greater than the distance Ds2 (<Ds3) and less than or equal to the distance Ds3, and a distance Ds1 (<Ds2). It is roughly divided into four stages: a range R3 with a distance Ds2 or less and a range R4 with a distance Ds1 or less. In the color map CM1, different colors CL1-CL4 are assigned to the ranges R1-R4, respectively. Note that the relevance between each word and the search query may be displayed as a continuous color change (gradation) in the color map CM1.

The relevance between the word WD5 highlighted with the color CL2 and the search query is greater than the relevance between the word (not shown) highlighted with the color CL1 and the search query. The relevance of words WD2, WD4, and WD7 highlighted with color CL3 to the search query is greater than the relevance of word WD5 to the search query. The relevance of words WD1, WD3, WD6, WD8 highlighted with color CL4 to the search query is greater than the relevance of words WD2, WD4, WD7 to the search query.

Again referring to FIG. 5, in S33 subsequent to S31, the display control unit 17 determines whether or not a selection operation (for example, a double-click by a mouse operation) has been performed on the language unit highlighted in the search result window. judge. If a selection operation has been performed on the highlighted linguistic unit (YES in S33), the display control unit 17 sets the selected linguistic unit in the search query in S35, and returns the process to S23. For example, when the user double-clicks the mouse MS1 while the cursor Cr is superimposed on the area surrounding the word WD2 in FIG. 7, the word WD2 is set as the search query, and the search process starts from S23 in FIG. be done. The user can determine a new search query by focusing on the relevance to the current search query.

If the highlighted linguistic unit has not been selected (NO in S33), the display control unit 17 determines in S37 whether or not the search result window has been closed. If the operation to close the search result window (for example, pressing the button Bn3 in FIG. 7) has not been performed (NO in S37), the display control unit 17 returns the process to S33. If the search result window has been closed (YES in S37), the display control unit 17 terminates the process.

In the search result window Wn2 shown in FIG. 7, the user can confirm the relevance of each of the plurality of words included in the related document D1 to the search query as a difference in highlight color. Even if the related document D1 does not contain a character string that matches the search query, the user can visually grasp the reason why the related document D1 was retrieved. Further, by changing the highlight color of the word according to the distance between the vector representation of each of the multiple words and the vector representation of the search query, each of the multiple words contained in the retrieved related documents and the search query can be shown to the user as a direct relationship that does not depend on other words other than the word.

In FIG. 7, for each of a plurality of words included in related documents, the case where the relevance between the word and the search query is expressed as the distance between the vector representation of the word and the vector representation of the search query has been described. The relevance between the word and the search query may be expressed as the contribution of the word to the distance between the vector representation of the relevant document and the vector representation of the search query.

A document W _i stored in the database 30 is expressed as a set of words wd _i,k as shown in Equation (1) below. The natural number i is any natural number from 1 to D (>1). Each of natural numbers k and t is any natural number from 1 to natural number N (>1).

A plurality of documents stored in the database 30 are expressed as a set of sentences W as shown in Equation (2) below.

The distance _Dsi between the document _Wi and the search query Q is expressed as in Equation (3) below.

The function f in equation (3) is a function that returns a vector representation of its arguments. Functions f include, for example, Doc2Vec, K-hot vectors, linear combinations of One-hot vectors, vector representations by counting words, and topic models.

In order to obtain the degree of contribution of words wd _i,t included in document W _i in equation (1) to distance Dsi, document W _i ,/t obtained by deleting words wd _{i ,t} from document W _i is expressed as follows: It is defined as in formula (4).

The distance Ds _i,/t between the document Wi, _/t and the search query Q is expressed as in Equation (5) below.

The greater the contribution of word wd _i,t to the relevance between search query Q and document _Wi , the smaller the relevance between search query Q and document _Wi,/t . That is, the greater the contribution of the word wd _i,t to the relevance between the search query Q and the document _Wi , the greater the distance Dsi _,/t . As a result, the difference between the distance Ds _i,/t and the distance Ds _i increases. The contribution Cn _i,n of the word wd _i, t is represented by the following equation (6).

FIG. 8 is a diagram showing how each of a plurality of words included in the related document D1 is highlighted using the degree of contribution of the word as the relationship between the word included in the related document D1 and the search query. The contents of the search result window Wn2 in FIG. 8 are obtained by replacing the color map CM1 of the search result window Wn2 in FIG. 7 with the color map CM2. Other than this, they are the same, so the description will not be repeated.

As shown in FIG. 8, the display control unit 17 changes the color of the peripheral area of each of the words included in the related document D1 in the search result window Wn2 according to the degree of contribution of the word. to highlight the word. The degree of contribution of the words included in the related document D1 is the range R11 smaller than the contribution Cn1, the range R12 greater than or equal to the contribution Cn1 and smaller than the contribution Cn2 (>Cn1), and the range R12 greater than or equal to the contribution Cn2 and Cn3 ( >Cn2) It is divided into four stages: a range R13 smaller than Cn2 and a range R14 greater than or equal to Cn3. In the color map CM2, different colors CL1 to CL4 are assigned to the ranges R11 to R14, respectively. Note that the degree of contribution of words may be displayed as a continuous color change (gradation) in the color map CM2.

Relevance between each of the plurality of linguistic units included in the related document and the search query by changing the highlight color of the word according to the degree of contribution of each of the words included in the related document can be presented to the user as a comprehensive relevance that reflects the relevance between other linguistic units and the search query.

7 and 8, the case where the linguistic unit highlighted in the search result window is a word has been described. The linguistic units highlighted may be other than words. By changing the type of linguistic unit, it is possible to show the user the grounds on which related documents have been retrieved from a plurality of viewpoints.

FIG. 9 shows the contents of the related document D1 displayed in the search result window Wn2 when the highlighted linguistic unit is a sentence. The color map CM2 shown in FIG. 9 shows the distribution of the degree of contribution of sentences.

As shown in FIG. 9, the relevance between the sentence ST4 highlighted with the color CL2 and the search query is greater than the relevance between the sentence (not shown) highlighted with the color CL1 and the search query. The relevance between the sentences ST3 and ST6 highlighted with the color CL3 and the search query is greater than the relevance between the sentence ST4 and the search query. The relevance between the sentences ST1, ST2, ST5 and ST7 highlighted with the color CL4 and the search query is greater than the relevance between the sentences ST3 and ST6 and the search query.

FIG. 10 shows the contents of the related document D1 displayed in the search result window Wn2 when the highlighted linguistic unit is a paragraph. The color map CM2 shown in FIG. 10 indicates the distribution of paragraph contributions.

As shown in FIG. 10, the relevance between the paragraph PR3 highlighted with the color CL2 and the search query is greater than the relevance between the paragraph (not shown) highlighted with the color CL1 and the search query. The relevance of paragraph PR1 highlighted with color CL3 to the search query is greater than the relevance of paragraph PR3 to the search query. The relevance between paragraph PR2 highlighted with color CL4 and the search query is greater than the relevance between paragraph PR1 and the search query.

Note that the linguistic unit highlighted in the search result window Wn2 does not have to be of one type, and may be selected from the group consisting of characters, morphemes, words, sentences, paragraphs, and any combination thereof. For example, letters and morphemes may be the linguistic units highlighted, or words, sentences, and paragraphs may be the linguistic units highlighted.

FIG. 11 shows an analysis case search when an analysis case search system 100A, which is an example of a system for searching and displaying related documents according to the embodiment, is connected to a plurality of client terminals 80a to 80n via a network 70. It is a functional block diagram which shows the structure of system 100A. As shown in FIG. 11, the analysis case search system 100A enables search of analysis cases in response to a request from a user or the like, for example, in an Internet environment, and provides the search results to the user or the like.

An analysis case retrieval system 100A and a plurality of client terminals 80a to 80n are communicably connected via a network 70 for information communication such as the Internet. The analysis case search system 100A and the plurality of client terminals 80a to 80n constitute a client server system 1000. FIG. A communication unit 61 in the analysis case search system 100A is an interface with the network 70 . The control unit 65 includes a CPU (Central Processing Unit) and controls the entire analysis case search system 100A including the learning processing unit 20 and the search processing unit 40 .

The memory 67 stores a learning processing program for the learning processing unit 20 described above, a search processing program for the search processing unit 40, and the like. The control unit 65 reads out these programs from the memory 67 and executes predetermined processes and the like shown in FIGS. 4 and 5. FIG. Note that the analysis case search system 100A can also be positioned as a server device connected to the network 70. FIG. That is, in the analysis case search system 100A, the display processing (display process) by the search processing unit 40 is performed on the server side. By connecting an existing client terminal to the server device, it is possible to show the user the reason why the relevant document was retrieved via the client terminal.

The display processing by the search processing unit 40 may be performed in a plurality of client terminals 80a to 80n (client side). By connecting the client terminal to the existing server device, it is possible to show the user the reason why the related document was retrieved via the client terminal.

As described above, according to the method and system for searching and displaying related documents according to the embodiment, it is possible to show the user the basis of the search result.

[Aspect]
It will be appreciated by those skilled in the art that the multiple exemplary embodiments described above are specific examples of the following aspects.

(Section 1) A method for retrieving and displaying related documents according to one aspect retrieves and displays at least one related document related to a search query from a database including a plurality of documents. The method includes a searching step and a displaying step. The retrieving step calculates a distance between the vector representation of each of the plurality of documents and the vector representation of the search query using a vector space that converts any linguistic unit into a vector representation, and retrieves at least one related document according to the distance. Search for The displaying step displays each of the at least one related document. In the display step, the language is displayed in a display mode according to the degree of relevance between the language unit and the search query based on the vector expression of each of the plurality of language units included in the related document and the vector expression of the search query. Including the step of displaying units.

According to the method described in item 1, according to the degree of relevance between the linguistic unit and the search query based on the vector representation of each of the plurality of linguistic units included in the related document and the vector representation of the search query By displaying the linguistic unit in this display mode, it is possible to show the user the basis of the search result.

(Section 2) In the method described in Section 1, the display mode of the linguistic unit includes the color of the surrounding area of the linguistic unit.

According to the method described in item 2, the relevance between each of the plurality of linguistic units included in the retrieved related documents and the search query is visually grasped as the difference in color of the surrounding area of the linguistic unit. be able to.

(Section 3) In the method described in Section 1 or Section 2, the degree of relevance between the linguistic unit and the search query is determined by the relationship between the vector representation of each of the plurality of linguistic units and the vector representation of the search query. Distance.

According to the method described in paragraph 3, the relevance between each of the plurality of linguistic units included in the retrieved related documents and the search query is directly related without depending on other linguistic units other than the relevant linguistic unit. It can be shown to the user as gender.

(Item 4) In the method described in item 1 or 2, the degree of relevance between the language unit and the search query is the degree of contribution of each of the plurality of language units. The degree of contribution is a value obtained by subtracting the distance between the vector representation of the related document and the vector representation of the search query from the distance between the vector representation of the search query and the vector representation of at least one related document excluding the relevant linguistic unit. is.

According to the method described in paragraph 4, the relevance between each of the plurality of language units included in the related documents retrieved and the search query is determined by determining the relevance between the search query and other language units other than the relevant language unit. can be shown to the user as a comprehensive relevance reflecting

(Section 5) In the method according to Section 1, the search step ranks the document higher as the distance between the vector representation of the search query and the vector representation for each of the at least one related document is shorter. including. The displaying step includes displaying the at least one relevant document as ranked by the searching step.

According to the method of item 5, the user can decide which related document to browse according to the ranking of relevance between each of the retrieved at least one related document and the search query.

(Item 6) In the method of item 1, the types of the plurality of linguistic units are selected from the group consisting of characters, morphemes, words, sentences, paragraphs, and any combination thereof.

According to the method described in item 6, by changing the types of a plurality of linguistic units, it is possible to show the user the reason why the related document was retrieved from a plurality of viewpoints.

(Section 7) The method described in Section 1 is performed in a client-server system. The display process is performed on the server side.

According to the method described in Item 7, by connecting an existing client terminal to the server device, it is possible to show the user the reason why the related document was retrieved via the client terminal.

(Section 8) The method described in Section 1 is performed in a client-server system. The display process takes place on the client side.

According to the method described in Item 8, by connecting the client terminal to the existing server device, it is possible to show the user the reason why the related document was retrieved via the client terminal.

(Item 9) In the method according to item 1, in the searching step, when a certain linguistic unit included in the displayed plurality of linguistic units is selected by the user, at least one linguistic unit is used as a search query. retrieving two related documents.

According to the method described in item 9, the user can determine a new search query by focusing on the relevance to the current search query.

(Item 10) In the method described in Item 1, the vector space is generated by machine learning that performs natural language processing on a corpus including a database.

According to the method described in item 10, in the process of machine learning for a model using a corpus, a vector space in which the features of the corpus are highly accurately reflected can be generated.

(Item 11) The system for retrieving and displaying related documents according to Item 11 retrieves and displays at least one related document related to a search query from a plurality of documents. The system includes a database and a search processor. The database contains multiple documents. The search processing unit calculates the distance between the vector representation of each of the plurality of documents and the vector representation of the search query using a vector space that converts any linguistic unit into a vector representation, and determines at least one relation according to the distance. Search for documents. A search processor displays each of the at least one related document. The search processing unit displays the relevant Display linguistic units.

According to the system of item 11, based on the vector representation of each of the plurality of linguistic units contained in the related document and the vector representation of the search query, according to the degree of relevance between the linguistic unit and the search query By displaying the relevant linguistic units in the display mode, it is possible to show the user the grounds for the search results.

It should be noted that, regarding the above-described embodiments and modifications, it is possible to appropriately combine the configurations described in the embodiments within a range that does not cause any inconvenience or contradiction, including combinations not mentioned in the specification at the time of filing. is scheduled from

It should be considered that the embodiments disclosed this time are illustrative in all respects and not restrictive. The scope of the present invention is indicated by the scope of the claims rather than the above description, and is intended to include all changes within the scope and meaning equivalent to the scope of the claims.

1 input unit, 5 output unit, 10 calculator, 11 analysis unit, 13 feature extraction unit, 15 search unit, 17 display control unit, 20 learning processing unit, 21 morphological analysis unit, 23 vector generation unit, 25 relevance learning unit, 25a word vector learning unit, 25b word-sentence learning unit, 25c word-sentence learning unit, 27 corpus, 30 database, 40 search processing unit, 60 display, 61 communication unit, 65 control unit, 67 memory, 70 network, 80a-80n client terminal, 100, 100A analysis case search system, 1000 client server system, KB1 keyboard, MS1 mouse.

Claims (10)

1. A computer-implemented method of retrieving and displaying at least one relevant document associated with a search query from a database containing a plurality of documents, comprising:
calculating a distance between the vector representation of each of the plurality of documents and the vector representation of the search query using a vector space that converts any linguistic unit into a vector representation, and determining the at least one related document according to the distance; a search step of searching for
displaying each of the at least one related document;
The display step includes
The linguistic unit is displayed in a display mode according to the degree of relevance between the linguistic unit and the search query based on the vector representation of each of the multiple linguistic units contained in the related document and the vector representation of the search query. including the step of displaying,
The magnitude of the relevance is the contribution of each of the plurality of linguistic units,
The degree of contribution is calculated from the distance between the vector representation of the search query and the vector representation of each of the at least one related document excluding the linguistic unit, and the distance between the vector representation of the related document and the vector representation of the search query. The method, which is the value after subtracting the 2. The method of claim 1, wherein the display aspect includes the color of the surrounding area of the linguistic unit. 3. The method according to claim 1 or 2, wherein said measure of relevance is a distance between a vector representation of each of said plurality of linguistic units and a vector representation of said search query. the searching step includes ranking the related documents higher as the distance between the vector representation of the search query and the vector representation for each of the at least one related documents is shorter;
2. The method of claim 1, wherein the displaying step comprises displaying the at least one related document according to ranking by the searching step. 2. The method of claim 1, wherein the plurality of linguistic unit types are selected from the group consisting of letters, morphemes, words, sentences, paragraphs, and any combination thereof. The method is performed in a client-server system,
2. The method of claim 1, wherein the displaying step is performed at the server side. The method is performed in a client-server system,
2. The method of claim 1, wherein the displaying step is performed on the client side. The searching step includes, when a certain linguistic unit included in the plurality of displayed linguistic units is selected by the user, retrieving the at least one related document using the linguistic unit as the search query. The method of claim 1. 2. The method of claim 1, wherein the vector space is generated by machine learning performing natural language processing on a corpus containing the database. A system for retrieving and displaying at least one related document related to a search query from a plurality of documents, comprising:
a database containing the plurality of documents;
calculating a distance between the vector representation of each of the plurality of documents and the vector representation of the search query using a vector space that converts any linguistic unit into a vector representation, and determining the at least one related document according to the distance; and displaying each of the at least one related document,
The search processing unit is
The linguistic unit is displayed in a display mode according to the degree of relevance between the linguistic unit and the search query based on the vector representation of each of the multiple linguistic units contained in the related document and the vector representation of the search query. display and
The magnitude of the relevance is the contribution of each of the plurality of linguistic units,
The degree of contribution is calculated from the distance between the vector representation of the search query and the vector representation of each of the at least one related document excluding the linguistic unit, and the distance between the vector representation of the related document and the vector representation of the search query. , the system.

Images