JP2021157220A - Natural language processing device and program - Google Patents

Abstract

To provide a natural language processing device and a program which compensate insufficient comprehensiveness in keyword retrieval.SOLUTION: In a retrieval system, a natural language processing device (server 100) includes: a first accepting unit (retrieval query receiving unit 172) which accepts a retrieval query for a data base which manages a plurality of sentences associated with classification words; and a retrieval unit 162 which retrieves a sentence which matches a condition of the retrieval query or a condition of the classification words of a dispersed vector similar to the dispersed vector of the retrieval query.SELECTED DRAWING: Figure 18

Description

The present invention relates to a text data retrieval technique.

In the method of collecting information in various fields, it is widely practiced to search a database by inputting a keyword as a search query. Generally, keywords are selected based on the user's judgment, so the search query set according to the user's sense is different. Also, if there are many synonyms or synonyms, the search results will differ depending on which word is used.

JP-A-2002-297636

When trying to extract text that suits the purpose with conventional keyword search, there is an uncertain aspect because it depends on human factors. In other words, there is a problem of lack of completeness that the text suitable for the purpose is omitted.

The present invention is an invention completed based on the above-mentioned problem recognition, and its main purpose is to make up for the lack of completeness in keyword search.

The natural language processing device in a certain aspect of the present invention is similar to the first reception unit that accepts a search query and the condition of the search query or the distribution vector of the search query with respect to the database that manages a plurality of sentences associated with the classification words. It is characterized by including a search unit for searching a sentence that matches the condition of the classification word of the dispersion vector to be used.

According to the present invention, it is possible to make up for the lack of completeness in keyword search.

It is a network configuration diagram in the search system of this embodiment. It is a data structure diagram of a company table. It is a data structure diagram of the business characteristic word table. It is a data structure diagram of the industry table. It is a data structure diagram of a word table. It is a data structure diagram of a company vector storage part. It is a data structure diagram of the business characteristic word vector storage part. It is a data composition diagram of the industry vector storage unit. It is a reception screen figure of a search query. It is a figure of the company information box included in the search result screen. It is a reception screen figure of a search query. It is a reception screen figure of a search query. It is a data structure diagram of a similar word list. It is a data structure diagram of the business characteristic word list. It is a data composition diagram of the industry list. It is a figure of the company information box included in the search result screen. It is a figure of the company information box included in the search result screen. It is a functional block diagram of a server. It is a functional block diagram of a data storage part. It is a flowchart which shows the process of the preparation phase processing. It is a flowchart which shows the process of the distributed representation model generation processing. It is a flowchart which shows the process of a company vector calculation process. It is a flowchart which shows the process of business characteristic word vector calculation processing. It is a flowchart which shows the process of the industry vector calculation process. It is a flowchart which shows the process of the utilization phase processing. It is a flowchart which shows the process of a list generation process. It is a flowchart which shows the process of a screen generation process. It is a figure of the company information box included in the search result screen.

FIG. 1 is a search system network configuration diagram according to the present embodiment.
The search system in this embodiment includes a server 100 and a user terminal 200. The user terminal 200 connects to the server 100 via a network (for example, the Internet, a LAN, a dedicated line, etc.). The server 100 has a corporate database 120.

When a user searches for company information managed in the company database 120, a search query is input by a user operation on the user terminal 200, and the search query is sent to the server 100. Then, the company information extracted based on the search query is returned to the user terminal 200 and displayed on the user terminal 200.

In this embodiment, the search function is enhanced by using the distributed representation model. In the distributed representation model, words are represented by high-dimensional real vectors. The distributed representation model is obtained by learning processing based on a database called a corpus, which is a large-scale collection of structured natural language sentences. This technique is well known in the field of natural language processing. The real number vector of a word obtained as a distributed representation model is sometimes called a distributed vector or a word vector. In the following, it is called a word vector.

Word2vec is known as a typical example. In Word2vec, based on the premise that "words that appear in the vicinity are similar to each other are semantically close", search for another word that appears in the vicinity of the target word (for example, 5 words before and after or 10 words before and after) in the sentence. Solve the problem of doing. The five words before and after here are the range of a total of 10 words, from the word 5 before the target word to the word before 1 and the word after 1 to 5 after the target word. be. A neural network is used as the learning means. In the process of learning using a large corpus, the word vectors of words that appear in the vicinity and have similar words change to show similar values.

In the distributed representation model, not only the similarity between words can be shown by the word vector, but also the similarity of sentences can be shown. In a sentence containing a plurality of words, a sentence vector of the sentence is generated from the word vector of these words. When the sentence vector of one sentence is similar to the sentence vector of another sentence, these sentences indicate that the contents are close to each other. In this example, the distributed representation model is stored in the distributed representation model storage unit 130.

In the present embodiment, a tag strongly related to the search query is selected based on the tag vector of the tag attached to the business content sentence of the company information. Then, a character string representing the tag is used as a condition for supplementing the search query.

The general procedure is divided into a preparation phase and a usage phase. In the preparatory phase, a distributed representation model based on the enterprise database 120 is generated using the prior art. That is, the word vector of the word in the company database 120 is generated. Next, a tag vector is generated for the tag used in the enterprise database 120 and added to the distributed representation model. This extends the distributed representation model to tag representation.

In the usage phase, a specific search query is accepted and the tags related to the search query are selected by referring to the distributed representation model. Then, the character expression of the tag is used as a search condition. At this time, the character expression of the tag is selected as the related phrase and is stored in the related phrase storage unit 140. The specific processing will be described in detail below.

FIG. 2 is a data structure diagram of the company table.
The company table is managed in the company database 120. The company table has a record for each company. In the record of the company table, a company ID, a company name, a business content sentence, a business characteristic word ID, a main industry ID, a sub-industry ID, and the like are set. Only one sentence may be set in the field of the business content sentence, or a plurality of sentences may be set. Business characteristics Words and industries correspond to tags. In this example, up to two industry tags can be set for one company. If only one industry tag is set, the main industry is used. The company table is prepared before the preparation phase.

FIG. 3 is a data structure diagram of the business characteristic word table.
The business characteristic word table is managed in the company database 120. The business characteristic word table has a record for each business characteristic word. A business characteristic word ID and a business characteristic word are set in the record of the business characteristic word table. The business characteristic word table is prepared before the preparation phase.

FIG. 4 is a data structure diagram of the industry table.
The industry table is managed in the company database 120. The industry table has records for each industry. An industry ID and an industry name are set in the record of the industry table. The industry table is prepared before the preparation phase.

FIG. 5 is a data structure diagram of a word table.
The word table is included in the distributed representation model. The word table has a record for each word. Words and word vectors are set in the records of the word table. The word table is generated during the preparation phase.

FIG. 6 is a data structure diagram of the company vector storage unit.
The company vector corresponds to the distributed representation vector of the entire business content sentence of the company. The company vector is generated from the word vector of the words contained in the business content sentence. The company vector storage unit is included in the distributed representation model storage unit 130. The enterprise vector is calculated in the preparatory phase. The company vector is the basis for calculating the business characteristic word vector and the industry vector.

FIG. 7 is a data structure diagram of the business characteristic word vector storage unit.
The business characteristic word vector shows the practical meaning of the business characteristic word. The business characteristic word vector is generated from the company vector of the company in which the business characteristic word is set, and is different from the word vector of the word corresponding to the business characteristic word (for example, "region"). The operational characteristic that the meaning is determined by what kind of company the business characteristic word is actually likely to be set in is quantified by the business characteristic word vector. The business characteristic word vector is calculated in the preparatory phase.

The business characteristic word vector storage unit stores the business characteristic word ID and the business characteristic word vector in association with each other. The business characteristic word vector storage unit is included in the distributed representation model storage unit 130. The business characteristic word ID is an example of a tag ID, and the business characteristic word vector is an example of a tag vector.

FIG. 8 is a data structure diagram of the industry vector storage unit.
The industry vector shows the practical significance of the industry tag. The industry vector is generated from the company vector of the company to which the industry tag is set, and is different from the meaning vector of the phrase corresponding to the industry name (for example, "automobile sales business"). The operational characteristic that the meaning is determined by what kind of company the industry tag is actually likely to be set to is quantified by the industry vector. The industry vector is calculated in the preparatory phase.

The industry vector storage unit stores the industry ID and the industry vector in association with each other. The industry vector storage unit is included in the distributed representation model storage unit 130. The industry ID is an example of a tag ID, and the industry vector is an example of a tag vector.

FIG. 9 is a screen view of a search query reception screen.
The search query reception screen is generated by the server 100 and displayed on the user terminal 200. In addition to the search query input area 300, this reception screen includes a slider 302, a display area 304 for similar words, a display area 306 for business characteristic words, and a display area 308 for the type of business.

In the similar word display area 304, similar words whose degree of similarity to the phrase input as the search query is equal to or higher than the reference value are displayed. Similar words are selected from all the words contained in the business content sentence. This reference value is called the first reference value. In the display area 306 of the business characteristic word, the business characteristic word whose similarity with the word / phrase input as the search query is equal to or higher than the reference value is displayed. This reference value is called the second reference value. The business characteristic word is set in advance. Similar ones are selected from them. In the industry display area 308, the industry whose similarity with the word / phrase input as the search query is equal to or higher than the reference value is displayed. This reference value is called the third reference value. The type of business is also set in advance. Similar ones are selected from them.

The slider value indicated by the slider 302 indicates a first reference value, a second reference value, and a third reference value. In this example, the slider value is the maximum value, and the first reference value, the second reference value, and the third reference value are also the maximum values. There are no similar words whose degree of similarity to "automobile" is equal to or higher than the first reference value, business characteristic words having the same degree of similarity to the second reference value or higher, and industries having the same degree of similarity to the third reference value or higher, and none of them is displayed. In this case, the search is performed by the search query input in the input area 300.

FIG. 10 is a diagram of a company information box included in the search result screen.
The company information box displayed as the search result in the state of FIG. 9 is shown. Company information including "automobile" is extracted from the company database 120 in the texts of the business content sentence, the business characteristic word name, the main industry name, and the sub-industry name. The company name included in the extracted company information is displayed in the display area 322. Similarly, the business content sentence is displayed in the display area 324. Similarly, the business characteristic word is displayed in the display area 326. Similarly, the main industry is displayed in the display area 328. Similarly, the sub-industry is displayed in the display area 330. If there is company information including "automobile" other than this example, it will be displayed as a box list on the search result screen. The search result screen is displayed side by side at the bottom of the search query reception screen. When a search query is entered, the search is automatically performed and the search result screen is displayed.

FIG. 11 is a screen view of a search query reception screen when the slider value is reduced.
When the user operates the slider 302 to reduce the slider value, the first reference value, the second reference value, and the third reference value become smaller, and similar words, business characteristic words, and industries that satisfy the conditions appear.

For example, the degree of similarity between the word vector of "automobile" and the word vector of "motorcycle" is equal to or higher than the first reference value indicated by the slider value at this time. The similarity between the word vector of "motorcycle" and the business characteristic word vector of "used" is equal to or higher than the second reference value indicated by the slider value at this time. Further, the similarity between the word vector of "motorcycle" and the industry vector of "automobile sales business" is equal to or higher than the third reference value indicated by the slider value at this time.

Similar words, business characteristic words, and industry names that appear here are added as OR conditions in the search for company information. Company information including "motorcycle", "used", "automobile sales business", etc. in the text of the business content sentence, the business characteristic word name, the main industry name, and the sub-industry name is extracted from the company database 120.

FIG. 12 is a screen view of a search query reception screen when the slider value is further reduced.
When the user operates the slider 302 to further reduce the slider value, the first reference value, the second reference value, and the third reference value become smaller, and the number of similar words, business characteristic words, and industries that satisfy the conditions increases.

The increased "truck", "quality", "automobile manufacturing industry", etc. are also added as OR conditions in the search for corporate information.

In this way, by operating the slider 302, the strength of the search query and the association is adjusted to increase or decrease the similar words, business characteristic words, and industries added to the search conditions. Thereby, the completeness in the search of corporate information can be adjusted. In the above example, the completeness increases in order. This operation will be described.

FIG. 13 is a data structure diagram of a similar word list.
When a search query is entered in the usage phase, a similar word list is generated accordingly. The similar word list is stored in the related word / phrase storage unit 140. The similar word list stores the similarity between the word and the phrase of the search query in association with each word. The similar word list is sorted in descending order of similarity.

In the example of FIG. 11, since the value of the slider 302 is 0.7 and the first reference value is also 0.7, the similar terms “motorcycle”, “taxi”, and “bus” are selected. In the example of FIG. 12, since the value of the slider 302 is 0.6 and the first reference value is also 0.6, the similar words “truck”, “traffic” and “railway” are added.

FIG. 14 is a data structure diagram of a business characteristic word list.
When a search query is entered in the usage phase, a business characteristic word list is generated accordingly. The business characteristic word list is stored in the related word / phrase storage unit 140. The business characteristic word list is associated with each business characteristic word and stores the degree of similarity between the business characteristic word and the phrase of the search query. The business characteristic word list is sorted in descending order of similarity.

In the example of FIG. 11, since the first reference value is also 0.7, the business characteristic words “used”, “repair”, and “luxury” are selected. In the example of FIG. 12, since the first reference value is 0.6, the business characteristic words “quality”, “popular”, and “new” are added.

FIG. 15 is a data structure diagram of the industry list.
When a search query is entered in the usage phase, an industry list is generated accordingly. The industry list is stored in the related word / phrase storage unit 140. The industry list stores the similarity between the industry and the words and phrases of the search query in association with each industry. The industry list is sorted in descending order of similarity.

In the example of FIG. 11, since the first reference value is 0.7, the “automobile sales industry”, “automobile maintenance industry”, and “automobile rental industry” of the industry are selected. In the example of FIG. 12, since the first reference value is 0.6, the industries “automobile manufacturing”, “transportation”, and “house sales” are added.

FIG. 16 is a diagram of a company information box included in the search result screen in the state of FIG.
In this example, since the character string of the business characteristic word "luxury" similar to the word of the search query matches the business characteristic word name, this company information is displayed as the search result. It should also be hit when the character string "luxury" is included in the business content sentence, main industry name or sub-industry name. In other words, the business content sentence, the name of the main industry or the name of the sub-industry are also included in the search range. However, it is not necessary to include the business content sentence, the name of the main industry, or the name of the sub-industry as the search range.

FIG. 17 is a diagram of a company information box included in the search result screen in the state of FIG.
In this example, since the character string of the business type "house sales business" similar to the word of the search query matches the main business type name, this company information is displayed as the search result. It also hits when the character string "house sales business" is included in the business content sentence, business characteristic word, or main industry name. In other words, the business content sentence, business characteristic word, or sub-industry name is also included in the search range. However, the search range does not have to be the business content sentence, business characteristic word, or sub-industry name.

An operation example based on these screen examples is shown. For example, suppose an insurer who wants to find a new car insurance policy candidate uses this search system. The insurer first enters "car" as the search query to get the search results in FIG. The automobile dealer shown in FIG. 10 is unlikely to become a new transaction route. This is because it has been widely practiced to recommend car insurance through a car dealer.

Therefore, as shown in FIG. 11, the slider 302 is operated to improve the completeness, and the re-searched company information of FIG. 16 is obtained. The corporate information of "Red Rose" does not include the character string of "automobile", but is substantially related to "automobile" through the business characteristic word "luxury" which is highly related to "automobile". The insurer looks at the business of "Red Rose" and prepares for a large amount of compensation when a helper dispatched by "Red Rose" causes an accident while driving a luxury car of a wealthy customer. I think there is a possibility. Then, the insurer realizes that he can propose an insurance policy to "Red Rose".

Next, as shown in FIG. 12, the slider 302 is operated to further enhance the completeness, and the re-searched company information of FIG. 17 is obtained. Although the corporate information of "Yellow Lemon" does not include the character string "automobile", it is substantially related to "automobile" through the business "house sales business" which is highly related to "automobile". Insurers look at this business and buy a house with a parking lot, so they think that the resident always has a car. Also, choosing a contractor who is familiar with the local circumstances means that the residents are not familiar with the traffic conditions in the area and there are few acquaintances. Therefore, if you are worried about having a traffic accident in an unfamiliar area and having trouble with other residents, you will find that it will motivate you to take out insurance. In this way, clues for business expansion can be found more and more.

In this example, it is unlikely that insurers will realize that "luxury" and "home sales" are key to business expansion. However, according to this embodiment, it becomes possible to find the target company information from the common word "automobile".

Since there are many notational fluctuations in the explanations by humans and the tag information given by hand is often not guaranteed to be complete, the completeness can be complemented by supporting the search by this embodiment. The details of the processing will be described below.

FIG. 18 is a functional block diagram of the server 100.
Each component of the server 100 is stored in a storage device and hardware including a CPU (Central Processing Unit), a computing unit such as various coprocessors, a storage device such as a memory and a storage device, and a wired or wireless communication line connecting them. It is realized by software that supplies processing instructions to the arithmetic unit. A computer program may be composed of a device driver, an operating system, various application programs located on the upper layers thereof, and a library that provides common functions to these programs. Each of the illustrated blocks shows a block for each function, not a configuration for each hardware.

The server 100 includes a data storage unit 110, a data processing unit 150, and a communication unit 190. The communication unit 190 is in charge of communication processing via the network. The data storage unit 110 stores various data. The data processing unit 150 executes various processes based on the data acquired by the communication unit 190 and the data stored in the data storage unit 110. The data processing unit 150 also functions as an interface between the communication unit 190 and the data storage unit 110.

The communication unit 190 includes a transmission unit 180 for transmitting data and a reception unit 170 for receiving data.
The transmission unit 180 includes a screen data transmission unit 182 that transmits various screen data. The receiving unit 170 includes a search query receiving unit 172 and a slider value receiving unit 174. The search query receiving unit 172 receives the search query. The slider value receiving unit 174 receives the slider value.

The data processing unit 150 includes a model generation unit 152, a company vector calculation unit 154, a business characteristic word vector calculation unit 156, an industry vector calculation unit 158, a screen data generation unit 160, and a search unit 162.
The model generation unit 152 executes a distributed representation model generation process for calculating a word vector. The company vector calculation unit 154 executes the company vector calculation process. The business characteristic word vector calculation unit 156 executes the business characteristic word vector calculation process. The industry vector calculation unit 158 executes the industry vector calculation process. The screen data generation unit 160 generates various screen data. The search unit 162 searches for company information.

The data storage unit 110 includes a company database 120, a distributed representation model storage unit 130, and a related phrase storage unit 140.

FIG. 19 is a functional block diagram of the data storage unit 110.
The company database 120 includes a company table storage unit 122, a business characteristic word table storage unit 124, and an industry table storage unit 126. The company table storage unit 122 stores the company table (FIG. 2). The business characteristic word table storage unit 124 stores the business characteristic word table (FIG. 3). The industry table storage unit 126 stores the industry table (FIG. 4).

The distributed representation model storage unit 130 includes a word table storage unit 132, a company vector storage unit 134, a business characteristic word vector storage unit 136, and an industry vector storage unit 138.
The word table storage unit 132 stores the word table (FIG. 5). The company vector storage unit 134 has been described in relation to FIG. The business characteristic word vector storage unit 136 has been described in relation to FIG. The industry vector storage unit 138 has been described in relation to FIG.

The related word / phrase storage unit 140 includes a similar word list storage unit 142, a business characteristic word list storage unit 144, and an industry list storage unit 146.
The similar word list storage unit 142 stores a similar word list (FIG. 13). The business characteristic word list storage unit 144 stores the business characteristic word list (FIG. 14). The industry list storage unit 146 stores the industry list (FIG. 15).

FIG. 20 is a flowchart showing the process of the preparation phase processing.
First, the model generation unit 152 executes the distributed representation model generation process. (S20). In the distributed representation model generation process, a word table (FIG. 5) is generated as a distributed representation model. The distributed representation model generation process will be described later in relation to FIG.

Next, the company vector calculation unit 154 executes the company vector calculation process. (S22). The company vector calculation process will be described later in relation to FIG.

Subsequently, the business characteristic word vector calculation unit 156 executes the business characteristic word vector calculation process (S24). The business characteristic word vector calculation process will be described later in relation to FIG.

Finally, the industry vector calculation unit 158 executes the industry vector calculation process (S26). The industry vector calculation process will be described later in relation to FIG. 24.

FIG. 21 is a flowchart showing the process of the distributed representation model generation process.
The model generation unit 152 generates a corpus from the company table (S30). A corpus is a large-scale data that structures sentences in natural language. The corpus is held in the data storage unit 110. Specifically, the model generation unit 152 performs morphological analysis for each sentence extracted from the business content sentence of each company and converts it into a word-separated form. A word-separator is a character string that separates words with a margin. The data in the word-separated format becomes the corpus.

The model generation unit 152 generates a word vector using a corpus (S32). Specifically, the model generation unit 152 obtains a word vector for each word included in the corpus by learning processing of Word2vec, for example. The generated word vector is stored in the word table storage unit 132. Then, the process returns to the process of S22.

FIG. 22 is a flowchart showing the process of the company vector calculation process.
The company vector calculation unit 154 repeats the following processing for each company (S40). The company vector calculation unit 154 calculates the sentence vector of each sentence included in the business content sentence of this company (S42). A sentence vector is generated from the word vector of the words contained in the sentence. The generation method is, for example, averaging. That is, the sentence vector is the average vector of the word vectors.

The company vector calculation unit 154 calculates the average of the sentence vectors of each sentence included in the business content sentence, and uses the calculated average vector as the company vector (S44). The company vector calculation unit 154 may generate a company vector from the sentence vector by a method other than averaging. The calculated company vector is stored in the company vector storage unit 134.

If there is a company that has not been processed yet (N in S46), the company vector calculation unit 154 returns to the processing in S40. When the processing for all the companies is completed (Y in S46), the process returns to the processing in S24.

FIG. 23 is a flowchart showing the process of the business characteristic word vector calculation process.
The business characteristic word vector calculation unit 156 repeats the following processing for each business characteristic word (S50). The business characteristic word vector calculation unit 156 refers to the company table and selects a company in which this business characteristic word is set (S52).

The business characteristic word vector calculation unit 156 calculates the average vector of the selected company vectors, and uses the calculated average vector as the business characteristic word vector of this business characteristic word (S54). The business characteristic word vector calculation unit 156 may generate a business characteristic word vector from the company vector by a method other than averaging. The calculated business characteristic word vector is stored in the business characteristic word vector storage unit 136.

If the business characteristic word that has not been processed remains (N in S56), the business characteristic word vector calculation unit 156 returns to the processing in S50. When the processing for all the business characteristic words is completed (Y in S56), the processing returns to the processing in S26.

FIG. 24 is a flowchart showing the process of the industry vector calculation process.
The industry vector calculation unit 158 repeats the following processing for each industry (S60). The industry vector calculation unit 158 refers to the company table and selects a company in which this industry is set (S62).

The industry vector calculation unit 158 calculates the average vector of the selected company vectors, and uses the calculated average vector as the industry vector of this industry. The industry vector calculation unit 158 may generate an industry vector from the company vector by a method other than averaging. The calculated industry vector is stored in the industry vector storage unit 138.

If there is an industry that has not been processed yet (N in S66), the industry vector calculation unit 158 returns to the processing in S60. When the processing for all industries is completed (Y in S66), the preparation phase processing is completed.

FIG. 25 is a flowchart showing the process of the utilization phase processing.
When the search query receiving unit 172 receives the search query from the user terminal 200 displaying the search query reception screen (Y in S70), the screen data generation unit 160 executes the list generation process (S72). ). In the list generation process, a similar word list (FIG. 13), a business characteristic word list (FIG. 14), and an industry list (FIG. 15) are generated. The list generation process will be described later in relation to FIG.

Next, the screen data generation unit 160 executes the screen generation process (S74). In the screen generation process, similar words, business characteristic words, and industries are displayed again on the search query reception screen. Further, the company information is searched and the search result screen is displayed. The screen generation process will be described later in relation to FIG. 27.

Subsequently, the screen data transmission unit 182 transmits the screen data to the user terminal 200 in the screen data transmission process (S76). Then, the process returns to the process of S70. The user terminal 200 displays a search query reception screen and a search result screen based on the received screen data.

When the slider value receiving unit 174 receives the slider value from the user terminal 200 displaying the search query reception screen (Y in S78), the first reference value, the second reference value, and the slider value are based on the slider value. Change the third reference value. In this example, the slider value is used as it is as a new first reference value, a new second reference value, and a new third reference value.

When the first reference value, the second reference value, and the third reference value are changed, the screen data generation unit 160 generates a screen based on the changed first reference value, the second reference value, and the third reference value. The process is executed (S82). Further, the screen data transmission unit 182 transmits the regenerated screen data in the screen data transmission process (S84). The user terminal 200 displays a search query reception screen and a search result screen based on the received screen data. Then, the process returns to the process of S70.

When the end instruction is received from the user terminal 200 (Y in S86), the usage phase processing is completed. If the end instruction is not received from the user terminal 200, the process returns to the process of S70.

FIG. 26 is a flowchart showing the process of list generation processing.
The screen data generation unit 160 calculates the similarity with the search query for all words (S90). If the search query is a word, the word vector of that word is used as the meaning vector of the search query. If the search query is a phrase, the meaning vector of the search query is generated from the word vector of the words contained in the phrase. The method of generation is, for example, averaging. Then, the similarity between the word vector of each word and the meaning vector of the search query is obtained. The screen data generation unit 160 calculates, for example, the cosine similarity. The same applies to the following similarities.

The screen data generation unit 160 sorts words in descending order of similarity to generate a similar word list (S92). The generated similar word list is stored in the similar word list storage unit 142.

The screen data generation unit 160 calculates the similarity with the search query for all business characteristic words (S94). Specifically, the screen data generation unit 160 obtains the similarity between the business characteristic word vector of each business characteristic word and the meaning vector of the search query.

The screen data generation unit 160 sorts the business characteristic words in descending order of similarity to generate a business characteristic word list (S96). The generated business characteristic word list is stored in the business characteristic word list storage unit 144.

The screen data generation unit 160 calculates the degree of similarity with the search query for all industries (S98). Specifically, the screen data generation unit 160 obtains the similarity between the industry vector of each industry and the meaning vector of the search query.

The screen data generation unit 160 sorts the industries in descending order of similarity and generates an industry list (S100). The generated industry list is stored in the industry list storage unit 146. Then, the process returns to the process of S74.

FIG. 27 is a flowchart showing the process of screen generation processing.
The screen data generation unit 160 arranges similar words having a degree of similarity equal to or higher than the first reference value in the display area 304 of the similar words on the reception screen of the search query (S110). The screen data generation unit 160 arranges the business characteristic word having a degree of similarity equal to or higher than the second reference value in the area 306 of the business characteristic word on the reception screen of the search query (S112). Further, the screen data generation unit 160 arranges an industry having a degree of similarity equal to or higher than the third reference value in the area 308 of the industry on the search query reception screen (S114).

The screen data generation unit 160 searches for company information based on these similar words, business characteristic words, and OR conditions of the type of business (S116). For example, generate a new search query with these similar words, business characteristic words, and strings to search for the industry, and enter this search query into the database management system (DataBase Management System). The database management system selects the relevant company information by comparing the company information of each company (business content sentence, business characteristic word, set of texts representing the main industry and sub-industry). The database management system is software that manages the corporate database 120 and operates the corporate database 120.

The screen data generation unit 160 arranges a company information box for displaying hit company information on the search result screen (S118). Then, the process returns to the process of S70.

[Modification example]
It may be possible to delete unnecessary words, business characteristic words, and industries displayed on the search query reception screen. When an unnecessary similar word, business characteristic word or industry is instructed by the user operation, a deletion request for the similar word, business characteristic word or industry is transmitted from the user terminal 200 to the server 100. When the request receiving unit (not shown) of the server 100 receives the deletion request, the deletion unit (not shown) deletes unnecessary similar words, business characteristic words, or industries from the similar word list, business characteristic word list, or industry list. .. Then, the server 100 again performs the list generation process, the screen generation process, and the screen data transmission process.

Although the example in which the company information is automatically searched in conjunction with the reception of the search query is shown, the company information may be searched after receiving the instruction to execute the search by the user operation.

The screen data generation unit 160 may arrange the similarity between the search query and the similar word next to the similar word on the search query reception screen. The screen data generation unit 160 may arrange the similarity between the search query and the business characteristic word next to the business characteristic word on the search query reception screen. The screen data generation unit 160 may arrange the similarity between the search query and the industry next to the industry on the search query reception screen.

Although the example in which the search query reception screen and the search result screen are displayed side by side is shown, the search query reception screen and the search result screen may be switched and displayed.

The server 100 may convert the slider value to obtain the first reference value, the second reference value, and the third reference value. Further, the first reference value, the second reference value and the third reference value may be different values. For example, the screen data generation unit 160 calculates the first reference value, the second reference value, and the third reference value so as to correlate with the slider value. The screen data generation unit 160 may obtain the first reference value by multiplying the slider value by the first coefficient. The screen data generation unit 160 may obtain the second reference value by multiplying the slider value by the second coefficient. The screen data generation unit 160 may obtain the third reference value by multiplying the slider value by the third coefficient.

In vector averaging, an evaluation index such as TF-IDF (Term Frequency-Inverse Document Frequency) may be used to weight each original word vector, phrase vector, sentence vector, or the like. ..

An example of searching for company information is shown, but other types of information may be searched. For example, when searching for product information on an electronic commerce site, a product classification vector or a product characteristic vector may be calculated for a product classification tag or a product characteristic tag attached to a product introduction sentence. Then, the search condition may be supplemented by using a product classification vector or a product characteristic vector similar to the search query.

For example, when searching for a patent document on a site for browsing patent documents, a patent classification vector may be calculated for the patent classification tag attached to the patent document. Then, the search condition may be supplemented by using a patent classification vector similar to the search query.

For example, when searching a posted article on a posting site of SNS (Social Networking Service), an article classification vector may be calculated for an article classification tag such as a hash tag attached to the posted article. Then, the search condition may be supplemented by using an article classification vector similar to the search query.

The present invention is not limited to the above-described embodiment or modification, and the components can be modified and embodied within a range that does not deviate from the gist. Various inventions may be formed by appropriately combining a plurality of components disclosed in the above embodiments and modifications. In addition, some components may be deleted from all the components shown in the above embodiments and modifications.

FIG. 28 is a diagram of a company information box included in the search result screen in the state of FIG.
In this example, the business characteristic word is "quick" and does not match any of the business characteristic keywords "used", "repair" and "luxury" (FIG. 11) that are similar to the words in the search query. However, since the character string of the business characteristic word "luxury" is included in the business content sentence, this company information is displayed as the search result. As described above in relation to FIG. 16, a hit is also made when the character string "luxury" is included in the business content sentence.

100 server, 110 data storage, 120 company database, 122 company table storage, 124 business characteristic word table storage, 126 industry table storage, 130 distributed representation model storage, 132 word table storage, 134 company vector storage 136 Business characteristic word vector storage unit, 138 Industry vector storage unit, 140 Related phrase storage unit, 142 Similar word list storage unit, 144 Business characteristic word list storage unit, 146 Industry list storage unit, 150 Data processing unit, 152 Model generation Department, 154 Company vector calculation unit, 156 Business characteristic word vector calculation unit, 158 Industry vector calculation unit, 160 screen data generation unit, 162 search unit, 170 reception unit, 172 search query reception unit, 174 slider value reception unit, 180 transmission Unit, 182 screen data transmission unit, 190 communication unit, 200 user terminal, 300 input area, 302 slider, 304 display area, 306 display area, 308 display area, 320 company information box, 324 display area, 326 display area, 328 display Area, 330 display area

Claims (5)

Regarding the database that manages multiple sentences associated with the classification word, the first reception section that accepts search queries and
A natural language processing apparatus including: A second reception unit that receives a reference value adjustment instruction regarding the similarity between the variance vectors, and
The natural language according to claim 1, further comprising an extraction unit for extracting the taxonomy of the variance vector similar to the variance vector of the search query based on the adjusted reference value. Processing equipment. The natural language processing apparatus according to claim 1 or 2, wherein the search unit determines that the sentence is suitable when the character string of the sentence includes the character string of the classification word. The description according to any one of claims 1 to 3, wherein the search unit determines that the sentence is suitable when the character string of the classification word associated with the sentence includes the character string of the search query. Natural language processing equipment. A function that accepts search queries for a database that manages multiple sentences associated with a taxonomy,
A program characterized in that an information processing apparatus is provided with a function of searching for a sentence that matches the conditions of the search query or the conditions of a classification term of a dispersion vector similar to the variance vector of the search query.

Images