JPWO2020079751A1 - Literature search method and literature search system - Google Patents

Abstract

High-speed literature search is performed without the need to replace or combine words in the literature data to solve similar words contained in the literature data, and even if a search keyword that does not exist in the search target is entered, the search is performed. As a result, a document search method and a document search system that can provide useful information are provided. Therefore, the database 20 interprets similar words as those in which the distance between the vectors is short by converting the words contained in the literature data into vectors expressing the meanings automatically acquired by machine learning, and combines, replaces, and index words. No need to set. When the search target included in the search query does not exist in the database 20, the search unit 15 refers to the conversion table 30 in which similar search targets similar to the search target are stored, and searches for similar search targets similar to the search target. get. Then, the conversion table 30 is searched by the search query after replacing the search target included in the search query with the acquired similar search target.

Description

The present invention relates to a document search method and a document search system for searching a search target from a large amount of document data based on an input search query.

With the development of computer technology and communication networks in recent years, there is an increasing demand for searching desired document data, information, etc. from a database in which a large amount of digitized documents, information, etc. are stored. In a general information retrieval system, a search process (matching process with an index term) is executed for a database by inputting a search term (keyword) that best represents the content that the searcher wants to search, and a specific condition is specified. Document data and the like that match are output as search results.

The literature contains words that are disagreeable (for example, differences in translation of interfaces, interfaces, etc., differences in okurigana, etc.) and similar words (for example, computers, personal computers, computers, etc.). Then, it was necessary to devise to treat these (hereinafter referred to as similar words) as the same thing. For example, search omissions were reduced by replacing similar words to be searched with all combinations or by associating multiple index terms with documents. However, as the number of combinations increases, there are problems that the search time becomes long, the optimum index term cannot be set, and the necessary search cannot always be performed. Similarly, it was necessary to resolve similar terms in search queries.

For example, in Patent Document 1, when the document data is registered, a fixed keyword is added to the document data to be registered and registered in the information retrieval system, and when the search is performed, the fixed keyword is specified to search the document database in a short time. The technology is disclosed.

By the way, in natural language processing in the field of artificial intelligence, a technique of applying machine learning to convert a word into a vector expressing the meaning of the word is used instead of handling the word itself. For example, a large amount of sentence information is learned by a neural network, a vector space representing the meaning of a word is generated from the distribution of words appearing before and after a certain word, and each word is converted into a vector.

By converting a word into a vector, similar words can be interpreted as having close distances between the vectors. Therefore, if the words included in the search target are converted into vectors, it is not necessary to combine, replace, or set index terms for similar words. When acquiring a word vector space by machine learning, the literature data to be searched is used as training data, or a general document available on the Web is added to the literature data to be used as training data to obtain a word vector. Expressions can be acquired automatically.

Japanese Unexamined Patent Publication No. 2007-200352

With the above method, vector representations of all the words contained in the literature data to be searched can be obtained, but since the words included in the search query are diverse, the vector space in which the words included in the search query are generated in advance. It cannot be handled if it is not included in. This is an unavoidable problem when there is a shortage of words in the learning data or when it is a new word. Here, a word that cannot be converted into a vector is called an unknown word.

Patent Document 1 uses similar words including various search terms and expressions as common registered keywords, and searches a literature database in which common registered keywords are embedded in the documents. At this time, since the similar words are converted into the common registered keywords, the information on the difference between the similar words is lost, and the accuracy of the reranking of the search results is lowered. In addition, since it is necessary to update all the documents in the document database when updating the common registration keyword, there is a problem that the calculation cost is high.

The present invention has been made in view of the above-mentioned problems, that is, the problem of similar words in a search target or a search query, and is a document search method and a document that can shorten the search time of documents containing similar words and improve the search accuracy. The purpose is to provide a search system.

Another object of the present invention is to provide a document search method and a document search system that can provide useful document information as a search result even when the word included in the search query is an unknown word.

As one means for solving the above-mentioned problems, the present invention is a document search method for searching a search target document stored in a document database based on an input search query, and automatically performs the document database by machine learning. The process of converting to a vector expressing the acquired meaning, the process of determining whether the search keyword included in the search query can be converted into a vector, and the process of determining whether the search keyword (word) included in the search query cannot be converted into a vector. If so, the step of converting to the word closest to the word contained in the vector space using the conversion table, the step of vectorizing the converted search query, and the step of searching the literature database by the search query. It is characterized by including a step of displaying the search result obtained by the search.

In the above-mentioned document search method, preferably, the document database stores the results vectorized by learning in advance, and the conversion table stores similar search keywords similar to the search keywords. Further, preferably, when there are a plurality of the similar search keywords, the replacement is performed with the most similar similar search keyword. Further, preferably, a step of adding a new similar search keyword similar to the search keyword to the conversion table and updating the conversion table is further provided. Further, preferably, even when the search keyword included in the search query exists in the literature database, the step of obtaining a similar search keyword similar to the search keyword by referring to the conversion table is included in the search query. It further includes a step of displaying the search result by the acquired similar search keyword together with the search result by the search keyword. Further, preferably, it further includes a step of displaying the conversion table on a display device and making it possible to select a search keyword to be specified. Further, preferably, the search keyword includes at least one of a compound name, a chemical formula, a chemical reaction formula, a chemical equation, an analyzer name, and an analysis method, and the literature database includes an analysis report and analysis for the search keyword. At least one of related papers and analysis-related patent documents is included.

According to the present invention, by converting a word contained in literature data into a vector expressing a meaning automatically acquired by machine learning, similar words can be interpreted as having a short distance between vectors, and combinations, substitutions, and index terms can be used. Search time can be shortened and search accuracy can be improved by eliminating the need for setting.

Further, according to the present invention, when a word included in a search query is an unknown word, a conversion table that converts the word into the closest word included in the vector space is used to support a wide variety of search queries and perform a search. As a result, useful literature information can be provided.

It is a block diagram which shows an example of the structure of the document search system which concerns on embodiment of this invention. It is a flowchart which shows the search processing procedure about the search target in the document search system. It is a figure which shows the example which displays a plurality of search targets acquired from a conversion table on a display in a document search system. It is a flowchart which shows the example which expands the search range and searches. It is a figure which shows the configuration example which connected the document search system via a network.

Hereinafter, embodiments according to the present invention will be described in detail with reference to the accompanying drawings. FIG. 1 is a block diagram showing an example of the configuration of a document search system according to an embodiment of the present invention. The document search system 10 of FIG. 1 includes a search processing unit 40 that searches for a search target (case, document, etc.) based on a search query input by a user or the like who is a searcher, and outputs the search result.

The search processing unit 40 includes an input unit 1, an analysis unit 11, a feature extraction unit 13, a search unit 15, a display control unit 17, and an output unit 5. The document search system 10 further includes a document database (hereinafter, also simply referred to as a database) 20, a conversion table 30, and a data update unit 19 that updates the data stored in the conversion table 30.

The search target in the document search system according to the present embodiment includes, for example, an analysis report, an analysis-related paper, an analysis-related patent document, and the like. The analysis report is, for example, a report of analysis cases and analysis results such as "Analysis of residual pesticides in sample △△ using apparatus 〇〇", and application news and application notes are also included in the analysis report.

A search query (also referred to as a search keyword or a search character string) in which a search target is represented by a character string, a sentence, or the like is input to the input unit 1 of the document search system 10 by a user or the like. The search query consists of a search sentence, a search keyword, a character string indicating a search condition, etc., and if the search target is an analysis report, the analysis-related search keyword, the analysis-related compound name, the analysis-related analysis target object name, etc. are included. ..

The input unit 1 includes, for example, an input device such as a keyboard and a mouse, a display, and the like, and a search query is input using an input screen on the display. The display is also an output unit 5 on which search results and the like are displayed.

The analysis unit 11 performs morphological analysis on the search query input from the input unit 1 based on a predefined search dictionary, and divides the search query into the smallest unit of words. Then, the feature extraction unit 13 calculates a vector (also referred to as a word vector or a feature vector) expressing the search query. For example, the word vector is learned and acquired based on the literature data stored in the database 20, general data collected from the Web (Web), and the like.

On the other hand, when the word vector cannot be calculated from the word of the search query, the feature extraction unit 13 extracts a similar search target with reference to the conversion table 30, converts the search query, and then converts the word vector into a word vector. .. The conversion table 30 is a list of words included in the word vector and words not included in the word vector calculated by different distance scales. For example, when a compound name is input as a search query, physical quantities such as compound similarity, use, and toxicity can be mentioned as a distance measure of the conversion table.

The search unit 15 searches the document data stored in the database 20 based on the word vector of the search keyword acquired from the feature extraction unit 13, and outputs it to the display control unit 17.

The output unit 5 displays information according to the control result such as the display content and the display format in the display control unit 17. As a result, the user or the like can acquire a desired search target or a search target similar thereto.

The data update unit 19 performs continuous maintenance by newly registering new words, synonyms, synonyms, etc. that are not in the conversion table 30 in the conversion table 30, so that the latest search targets are accumulated. maintain.

In the document search system 10, the conversion table 30 is provided separately from the database 20 because the database 20 requires re-learning of word vector acquisition by machine learning, whereas the conversion table 30 updates the accumulated data. This is because it is easy to use and the data storage structure is simple.

Next, the search processing procedure in the document search system of the present embodiment will be described. FIG. 2 is a flowchart showing a search processing procedure for a search target in the document search system of the present embodiment. In step S11 of FIG. 2, a search query (search keyword) is input by a user or the like. In the next step S13, the search query is divided into the smallest unit morphemes by morphological analysis, and the word vector is calculated.

In step S15, a search process is performed with reference to the database 20 in which data obtained by vectorizing document data (for example, analysis report, analysis-related paper, analysis-related patent document, etc.) is stored. In this search process, a search target related to a search query (search keyword) is searched, and that is used as a search result.

In step S17, it is determined from the search result in step S15 whether or not the search target exists in the database 20, that is, whether or not the search target is a search target consisting of "unknown words". For example, if the search query includes the name of a compound or the like as a search target, but the compound name does not exist in the database 20, the compound name is recognized as an "unknown word".

If the search target exists in the database 20, the search result is displayed in step S27. On the other hand, when a search query not included in the database 20 is input and the search target is an unknown word and the search target does not exist in the database 20, the conversion table 30 is referred to in step S19.

In the conversion table 30, as described above, the relationships with similar search targets are calculated for each search target, and they are listed. For example, when the conversion table 30 is a compound conversion table, the results of calculating the relationship between the compound and a similar compound are accumulated for each known compound.

In step S21, a search target similar to the search query is acquired. More specifically, if the search query is a compound and a phthalate ester is entered as the compound, the phthalates listed in the conversion table 30 are diprovyl phthalates, dinebutyl phthalates, phthalates. Search queries are for similar compounds such as dinonyl, ventilisoventyl phthalate, and dimethyl phthalate.

Therefore, in the process in step S21, the input search query is replaced with a similar search query to search the conversion table 30. When there are a plurality of similar compounds obtained, the search query is replaced by, for example, the most similar compound among them.

In step S23, from the plurality of similar compounds acquired in step S21, for example, the optimum similar compound in consideration of the characteristics of the search query and the like is selected.

The search target in the conversion table 30 may include not only the compound but also the chemical formula, the name of the analyzer, the chemical equation, the analysis method, and the like. Further, in order to improve the accuracy of determining whether or not the search target is similar to the search target, for example, the structural formula, the polarity, and the like may be additionally input.

In step S25, a search process such as extracting the similar compound selected in step S23 above from the conversion table 30 is performed, and in the subsequent step S27, the search result is displayed.

FIG. 3 is an example of displaying a plurality of search targets acquired from the conversion table 30 on the display in the document search system of the present embodiment. Here, when the search target is compound A and a plurality of similar compounds are obtained from the conversion table 30, the plurality of similar compounds A1, A2, and A3 are displayed on the display 41, and the user can display the similar compounds A1, A2, and A3. By checking the box 43a provided corresponding to the compound, a search target (similar compound) can be selected.

Similar to the compound A, the other compounds B and C to be searched also have similar compounds B1 to B4 and C1 displayed and can be selected in the boxes 43b and 43c.

The search process for the search target in the document search system of the present embodiment is not limited to the example shown in FIG. 2, and a search with an expansion of the search range may be performed as follows.

FIG. 4 is a flowchart showing an example of searching by expanding the search range. In FIG. 4, the same steps as those shown in FIG. 2 are designated by the same reference numerals, and their description will be omitted.

In the search process shown in FIG. 4, if it is determined in step S17 that the search target does not exist in the database 20 from the search result in step S15, the processes of steps S19, S21, and S23 are executed. Even when the search target exists in the database 20, the same processing as in steps S19, S21, and S23 is performed.

That is, even if the determination result in step S17 is NO and the search target exists in the database 20, the conversion table 30 is listed by calculating the relationship with the similar search target for each search target in step S31. Refer to.

In step S33, a search target similar to the search target is acquired. At that time, the input search query is replaced with a similar search query to search the conversion table 30. The search query is replaced by the most similar compound among the obtained plurality of similar compounds.

In step S35, the optimum similar search target is selected from the search targets similar to the searched search target. As a result, it is possible to obtain search results with an expanded search range by adding to the search results of the search target entered by the user, acquiring search results for similar search targets, and displaying them. ..

FIG. 5 shows a configuration example in which the document search system of the present embodiment is connected via a network (public communication network). As shown in FIG. 5, the document search system 10 according to the present embodiment enables a case search in response to a request from a user or the like in, for example, an Internet environment, and provides the search result to the user or the like.

Therefore, the document search system 10 and the plurality of user terminals 60a ... 60n are communicably connected via an information communication network 70 such as the Internet. The communication unit 61 in the document search system 10 is an interface with the network 70. The control unit 65 is composed of, for example, a microprocessor or the like that controls the entire document search system 10 including the search processing unit 40.

The memory 67 stores a search processing program or the like to be searched by the search processing unit 40 described above. The control unit 65 reads out these programs and executes predetermined processing and the like shown in FIGS. 2 and 4. The entire document search system 10 can be positioned as a server device connected to the network 70.

As described above, according to the present embodiment, a database storing document data and a conversion table storing similar search targets similar to the search target are prepared, and even when an unknown word is input as a search query, the conversion table By acquiring a similar search target by referring to and displaying it as a search result, useful information can be surely presented to the user.

Therefore, it is allowed to include unknown words when the learning data is small, and even when unknown words are input, similar search targets (similar compounds) using known relationships of search targets (compounds). ) Is selected, and the search accuracy can be improved by performing a search based on the similar search target (similar compound).

Further, by providing a conversion table separately from the database, it becomes easy to update the data stored in the conversion table having a simple data storage structure. As a result, it is possible to avoid updating the database, which requires a large calculation cost, as in the case where the database has similar word information.

Furthermore, by continuously registering new words, their synonyms, synonyms, etc. in the conversion table, the conversion table can be maintained in a state in which the latest search targets are accumulated.

Even if the search query (compound) entered by the user is not an unknown word, a similar compound is searched by the conversion table and added to the search result of the search query (compound) entered by the user to perform a similar search. Search results for the target (similar compound) can be presented.

1 Input unit 5 Output unit 10 Document search system 11 Analysis unit 13 Feature extraction unit 15 Search unit 17 Display control unit 19 Data update unit 20 Database 30 Conversion table 40 Search processing unit 41 Display 60a ... 60n User terminal 61 Communication unit 65 Control unit 67 Memory 70 Information and Communication Network

Claims (7)

This is a document search method that searches the document data stored in the database based on the entered search query.
A step of determining whether or not the search target included in the search query exists in the database, and
When it is determined that the search target does not exist in the database, a step of referring to the conversion table and acquiring a similar search target similar to the search target, and
The step of replacing the search target included in the search query with the acquired similar search target, and
The step of searching the database by the search query after the replacement, and
The process of displaying the search results obtained by the above search and
A document search method characterized by comprising. In the database, a vector expressing the meaning of a word is acquired by machine learning, and data obtained by mapping literature data into a vector space is stored, and the conversion table is similar to a search target based on known information. The document search method according to claim 1, wherein a similar search target is stored. The document search method according to claim 1, wherein when there are a plurality of similar search targets, the replacement is performed with the most similar similar search target. The document search method according to claim 1, further comprising a step of adding a new similar search target similar to the search target to the conversion table and updating the conversion table. A step of acquiring a similar search target similar to the search target by referring to the conversion table even when the search target included in the search query exists in the database.
A process of displaying the search results of the acquired similar search targets together with the search results of the search targets included in the search query, and
The document search method according to claim 1, further comprising. The search query includes at least one of a compound name, a chemical formula, a chemical reaction formula, a chemical equation, an analyzer name, and an analysis method, and the literature data includes an analysis report, an analysis-related article, and an analysis-related item for the search target. The document search method according to any one of claims 1 to 5, wherein at least one of the patent documents is included. A literature search system that searches for literature data stored in a database based on the entered search query.
A means for determining whether or not the search target included in the search query exists in the database, and
When it is determined that the search target does not exist in the database, a means for obtaining a similar search target similar to the search target by referring to the conversion table, and
A means for replacing the search target included in the search query with the acquired similar search target, and
A means for searching the database by the search query after the replacement, and
A means of displaying the search results obtained by the above search,
A literature search system characterized by being equipped with.

Images