Natural language processing devices and programs
By calculating the word segment vectors of sentences associated with classification words, the problem of the inability to fully represent the meaning of sentences associated with classification words in the prior art is solved, and an effective summary and representation of the semantic relationship of sentences is achieved.
Patent Information
- Application Number
- CN202010206312.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Priority Date
- 2020-03-19
- Filing Date
- 2020-03-23
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2040-03-23
AI Technical Summary
The prior art is difficult to express the meaning of sentences associated with classification words in a summary. Even if the word participle vector of a word or sentence is used as the word participle vector of a label, it cannot fully represent the semantics of the sentences associated with the label.
By calculating and managing word participle vectors of words in the database of multiple sentences associated with classification words, a word participle vector of classification words is generated, and then a summary of the meaning of sentences associated with classification words is expressed.
An effective summary of the meaning of sentences associated with classification words is achieved, and a participle vector that can fully represent semantic relations is generated.
Smart Images

Figure CN113495950B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a technology for expressing natural language word segments. Background Art
[0002] Regarding the database that manages sentences, sometimes tags are added to classify sentences. Sentences with common tags belong to the same category, and similar sentences can be collected by extracting sentences using tags as keywords. In other words, tags are classifiers of sentences, and they function as search conditions, for example.
[0003] In addition, in natural language processing, the technology of expressing the meaning of words, phrases and sentences by using high-dimensional real number vectors has become popular. This real number vector is called a word segmentation vector. If the word segmentation vector is used, the meaning of natural language can be compared.
[0004] Prior art literature
[0005] Patent Literature
[0006] Patent Document 1: Japanese Patent Application Publication No. 2019-149161
[0007] Patent Document 2: Japanese Patent Application Publication No. 2018-073429 Summary of the invention
[0008] Problems to be solved by the invention
[0009] If the labels are also expressed as word vectors, it is possible to mathematically compare the semantic relationship between labels and the semantic relationship between labels and sentences.
[0010] However, the label name is used to represent the classification, such as the background and characteristics of the classified sentence, which is only one aspect of the meaning of the sentence. Therefore, even if the word or sentence vector of the label name is used as the label vector, it will not comprehensively represent the meaning of the sentence associated with the label.
[0011] The present invention has been accomplished based on the above-mentioned problem recognition, and its main object is to obtain a word segmentation vector of a classifier that comprehensively expresses the meaning of a sentence associated with the classifier.
[0012] Methods for solving problems
[0013] A natural language processing device according to a certain embodiment of the present invention is characterized in that it includes: a first calculation unit, which calculates the word segmentation vectors of the words contained in a database managing multiple sentences associated with classifiers; a second calculation unit, which calculates the word segmentation vector of each sentence based on the word segmentation vector of the word contained in the sentence; and a third calculation unit, which calculates the word segmentation vector of the classifier based on the word segmentation vectors of each sentence associated with the same classifier.
[0014] Effects of the Invention
[0015] According to the present invention, it is possible to obtain a word segmentation vector of a classifier that comprehensively expresses the meaning of a sentence associated with the classifier. BRIEF DESCRIPTION OF THE DRAWINGS
[0016] Figure 1 This is a diagram of the network structure in the search system of this embodiment.
[0017] Figure 2 It is the data structure diagram of the enterprise table.
[0018] Figure 3 This is a data structure diagram of the business characteristics vocabulary.
[0019] Figure 4 This is the data structure diagram of the industry category table.
[0020] Figure 5 It is a data structure diagram of the word list.
[0021] Figure 6 This is a data structure diagram of the enterprise vector storage department.
[0022] Figure 7 This is a data structure diagram of the business characteristic word vector storage unit.
[0023] Figure 8 This is a data structure diagram of the industry category vector storage unit.
[0024] Fig. 9 This is a picture of the search query acceptance screen.
[0025] Fig.10 This is a diagram of a company information box included in the search result screen.
[0026] Fig.11 This is a picture of the search query acceptance screen.
[0027] Fig.12 This is a picture of the search query acceptance screen.
[0028] Fig.13 It is a data structure diagram of a similar word list.
[0029] Fig.14 This is a data structure diagram of a business characteristic word list.
[0030] Fig.15 This is a data structure diagram of a list of industry categories.
[0031] Fig.16 This is a diagram of a company information box included in the search result screen.
[0032] Fig.17 This is a diagram of a company information box included in the search result screen.
[0033] Fig.18 It is a functional block diagram of the server.
[0034] Fig.19 This is a functional block diagram of the data storage unit.
[0035] Fig. 20 It is a flowchart showing the process of the preparation phase.
[0036] Fig.21 It is a flowchart showing the process of generating a word segmentation expression model.
[0037] Fig. 22 is a flow chart showing the process of enterprise vector calculation processing.
[0038] Fig.23 This is a flowchart showing the process of calculating the business characteristic word vector. Fig.24 It is a flowchart showing the process of industry category vector calculation processing.
[0039] Fig.25 It is a flowchart showing the process of utilizing the phase processing.
[0040] Fig.26 This is a flowchart showing the procedure of the list generation process.
[0041] Fig. 27 This is a flowchart showing the procedure of screen generation processing.
[0042] Description of Reference Numerals
[0043] 100 servers
[0044] 110 Data Storage Department
[0045] 120 Enterprise Database
[0046] 122 Enterprise Table Preservation Department
[0047] 124 Business characteristics vocabulary storage department
[0048] 126 Industry Category List Storage Department
[0049] 130 Word segmentation expression model storage unit
[0050] 132 Word list storage unit
[0051] 134 Enterprise Vector Storage Department
[0052] 136 Business Characteristic Word Vector Storage Department
[0053] 138 Industry Type Vector Storage Department
[0054] 140 Association statement storage unit
[0055] 142 Similar word list storage unit
[0056] 144 Business characteristic word list storage unit
[0057] 146 Industry Category List Storage Department
[0058] 150 Data Processing Department
[0059] 152 Model Generation Department
[0060] 154 Enterprise Vector Computing Department
[0061] 156 Business Characteristic Word Vector Calculation Department
[0062] 158 Industry Category Vector Calculation Department
[0063] 160 Screen data generation unit
[0064] 162 Search Department
[0065] 170 Receiving Department
[0066] 172 Search Query Receiving Department
[0067] 174 Slider value receiving unit
[0068] 180 Sending Department
[0069] 182 Screen data transmission unit
[0070] 190 Ministry of Communications
[0071] 200 User Terminals
[0072] 300 Input Area
[0073] 302 Slider
[0074] 304 Display Area
[0075] 306 Display Area
[0076] 308 display area
[0077] 320 Enterprise Information Box
[0078] 324 display area
[0079] 326 Display Area
[0080] 328 display area
[0081] 330 display area DETAILED DESCRIPTION
[0082] Figure 1 This is a network structure diagram of the retrieval system of this embodiment.
[0083] The search system of this embodiment includes a server 100 and a user terminal 200. The user terminal 200 is connected to the server 100 via a network (for example, the Internet, a LAN, or a dedicated line). The server 100 has a company database 120.
[0084] When the user searches for enterprise information managed by enterprise database 120, the user inputs a search query in user terminal 200 and transmits the search query to server 100. Then, enterprise information extracted based on the search query is returned to user terminal 200 and displayed on user terminal 200.
[0085] In this embodiment, a word segmentation expression model is used to improve the search function. In the word segmentation expression model, a word is represented by a high-dimensional real number vector. The word segmentation expression model is obtained by a learning process based on a database called a corpus, which is a structured and large-scale aggregation of natural language articles. This technology is well known in the field of natural language processing. Sometimes the real number vector of a word obtained as a word segmentation expression model is called a word segmentation vector or a word vector. Hereinafter, it is referred to as a word vector.
[0086] As a representative example, Word2vec is known. In Word2vec, under the premise that "similar words appearing nearby have similar meanings", the problem of searching for other words that appear near the target word in a sentence (for example, 5 words before and after or 10 words before and after) is solved. Here, the 5 words before and after are a total of 10 words ranging from 5 words before the target word to 1 word before the target word and from 1 word after the target word to 5 words after the target word. A neural network is used as a learning unit. In the process of learning using a large corpus, the word vectors of words that are similar to words appearing nearby change in a way that shows similar values.
[0087] In the word segmentation expression model, not only the similarity between words is represented by word vectors, but also the similarity between sentence sets. In a sentence set containing multiple words, the sentence vector of the sentence set is generated from the word vectors of these words. When the sentence vector of a certain sentence set is similar to the sentence vector of another sentence set, it means that the contents of these sentence sets are similar. In this example, the word segmentation expression model is stored in the word segmentation expression model storage unit 130.
[0088] In this embodiment, a label that is strongly related to a search query is selected based on a label vector of a label of a business content sentence set attached to corporate information, and a character string representing the label is used as a condition to supplement the search query.
[0089] As a rough order, it is divided into a preparation stage and a utilization stage. In the preparation stage, a word segmentation expression model based on the enterprise database 120 is generated using existing technology. That is, word vectors of words in the enterprise database 120 are generated. Then, label vectors are generated for the labels used in the enterprise database 120 and added to the word segmentation expression model. In this way, the word segmentation expression model is extended to the expression of labels.
[0090] In the utilization phase, a specific search query is received, and a tag associated with the search query is selected by referring to the word segmentation expression 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 a related sentence and stored in the related sentence storage unit 140. The specific processing is described in detail below.
[0091] Figure 2 It is the data structure diagram of the enterprise table.
[0092] The enterprise table is managed in the enterprise database 120. The enterprise table has a record for each enterprise. In the records of the enterprise table, enterprise ID, enterprise name, business content sentence set, business characteristic word ID, main industry category ID and sub-industry category ID are set. In the field of the business content sentence set, only one sentence can be set or multiple sentences can be set. Business characteristic words and industry categories are equivalent to labels. In this example, for one enterprise, up to two industry category labels can be set. When only one industry category label is set, the main industry category is used. Prepare the enterprise table before the preparation stage.
[0093] Figure 3 This is a data structure diagram of the business characteristics vocabulary.
[0094] The business characteristic word table is managed in the enterprise 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 a record of the business characteristic word table. The business characteristic word table is prepared before the preparation stage.
[0095] Figure 4 This is the data structure diagram of the industry category table.
[0096] The industry category table is managed in the enterprise database 120. The industry category table has a record for each industry category. The industry category ID and the industry category name are set in the record of the industry category table. The industry category table is prepared before the preparation stage.
[0097] Figure 5 It is a data structure diagram of the word list.
[0098] The word list is included in the word segmentation expression model. The word list has a record for each word. The word and word vector are set in the record of the word list. The word list is generated in the preparation stage.
[0099] Figure 6 This is a data structure diagram of the enterprise vector storage department.
[0100] The enterprise vector is equivalent to the word segmentation expression vector of the entire business content sentence set of the enterprise. The enterprise vector is generated from the word vectors of the words contained in the business content sentence set. The enterprise vector storage unit is included in the word segmentation expression model storage unit 130. The enterprise vector is calculated in the preparation stage. The enterprise vector becomes the basis for the calculation of the business characteristic word vector and the industry category vector.
[0101] Figure 7 This is a data structure diagram of the business characteristic word vector storage unit.
[0102] The business characteristic word vector represents the practical meaning of the business characteristic word. The business characteristic word vector is a vector generated from the enterprise vector of the enterprise that has set the business characteristic word, and is different from the word vector of the word equivalent to the business characteristic word (for example, "region"). The operational characteristics of the business characteristic word, which determine its meaning depending on which enterprise it is likely to be set by, are digitized by the business characteristic word vector. The business characteristic word vector is calculated in the preparation stage.
[0103] 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 word segmentation expression model storage unit 130. The business characteristic word ID is an example of a label ID, and the business characteristic word vector is an example of a label vector.
[0104] Figure 8 This is a data structure diagram of the industry category vector storage unit.
[0105] The industry category vector represents the practical meaning of the industry category label. The industry category vector is a vector generated from the enterprise vector of the enterprise that has set the industry category label, and is different from the semantic vector of a phrase equivalent to the industry category name (for example, "car dealer"). The operational characteristic that the industry category label is determined by which enterprise it is likely to be set by is digitized by the industry category vector. The industry category vector is calculated in the preparation stage.
[0106] The industry category vector storage unit stores the industry category ID and the industry category vector in association with each other. The industry category vector storage unit is included in the word segmentation expression model storage unit 130. The industry category ID is an example of a label ID, and the industry category vector is an example of a label vector.
[0107] Fig. 9This is a picture of the search query acceptance screen.
[0108] The search query acceptance screen is generated by the server 100 and displayed on the user terminal 200. The acceptance screen includes a slider 302, a display area 304 for similar words, a display area 306 for business-specific words, and a display area 308 for industry categories in addition to a search query input area 300.
[0109] In the display area 304 for similar words, similar words whose similarity to the sentence input as the search query is above the reference value are displayed. Similar words are selected from all words contained in the business content sentence set. This reference value is called the first reference value. In the display area 306 for business characteristic words, business characteristic words whose similarity to the sentence input as the search query is above the reference value are displayed. This reference value is called the second reference value. The business characteristic words are preset. Similar words are selected from them. In the display area 308 for industry categories, industry categories whose similarity to the sentence input as the search query is above the reference value are displayed. This reference value is called the third reference value. The industry categories are also preset. Similar words are selected from them.
[0110] The slider value indicated by the slider 302 indicates the first reference value, the second reference value, and the 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 with a similarity of more than the first reference value, business characteristic words with the same second reference value, and industry categories with the same third reference value as "car", and none of them are displayed. In this case, the search query input in the input area 300 is used for search.
[0111] Fig.10 This is a diagram of a company information box included in the search result screen.
[0112] express Fig. 9 A company information box displayed as a search result in the state. Company information containing "automobile" is extracted from the company database 120 in the text of business content sentences, business characteristic word names, main industry category names and sub-industry category names. The company name contained in the extracted company information is displayed in display area 322. Similarly, the business content sentences are displayed in display area 324. Similarly, the business characteristic words are displayed in display area 326. Similarly, the main industry category is displayed in display area 328. Similarly, the sub-industry category is displayed in display area 330. In addition to this example, if there is company information containing "automobile", it is displayed as a dialog box list on the search result screen. In addition, the search result screen is displayed side by side below the search query acceptance screen. When a search query is entered, the search is automatically performed and the search result screen is displayed.
[0113] Fig.11This is an image of the search query acceptance screen when the slider value is reduced.
[0114] 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-specific words, and industry categories that meet the conditions appear.
[0115] For example, the similarity between the word vector of "car" and the word vector of "bicycle" is greater than the first reference value indicated by the slider value at that time. The similarity between the word vector of "bicycle" and the business characteristic word vector of "second-hand" is greater than the second reference value indicated by the slider value at that time. In addition, the similarity between the word vector of "bicycle" and the industry type vector of "car dealer" is greater than the third reference value indicated by the slider value at that time.
[0116] Here, the similar words, business characteristic words, and industry category names that appear are added as OR conditions in the search for corporate information. Corporate information containing "bicycle", "second-hand" or "car dealer" in the text of business content sentence set, business characteristic word name, main industry category name and sub-industry category name is extracted from the corporate database 120.
[0117] Fig.12 This is a picture of the search query acceptance screen when the slider value is further narrowed.
[0118] 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 further decrease, and the similar words, business-specific words, and industry types that meet the conditions increase.
[0119] Thus, additional "truck", "quality" or "car manufacturer" etc. are also added as OR conditions in the search for corporate information.
[0120] In this way, by operating the slider 302, the strength of association with the search query is adjusted, and similar words, business characteristic words and industry types added to the search conditions are increased or decreased. In this way, the comprehensiveness of the search for corporate information can be adjusted. In the above example, the comprehensiveness is increased in order. This action is explained.
[0121] Fig.13 It is a data structure diagram of a similar word list.
[0122] When a search query is input in the utilization phase, a similar word list is generated accordingly. The similar word list is stored in the associated sentence storage unit 140. The similar word list stores the similarity between each word and the sentence of the search query in association with each other. The similar word list is sorted in descending order of similarity.
[0123] exist Fig.11In the example, the value of slider 302 is 0.7, and the first reference value is also 0.7, so similar words "bicycle", "taxi" and "bus" are selected. Fig.12 In the example, the value of the slider 302 is 0.6 and the first reference value is also 0.6, so similar words "truck", "transportation" and "railway" are added.
[0124] Fig.14 This is a data structure diagram of a business characteristic word list.
[0125] When a search query is input in the utilization phase, a business characteristic word list is generated in response thereto. The business characteristic word list is stored in the associated sentence storage unit 140. The business characteristic word list stores the similarity between each business characteristic word and the sentence of the search query in association with each other. The business characteristic word list is sorted in descending order of similarity.
[0126] exist Fig.11 In the example, the first reference value is also 0.7, so the business characteristic words "semi-used", "repair", and "high-end" are selected. Fig.12 In the example, the first reference value is 0.6, so the business characteristic words "quality", "popular", and "new product" are added.
[0127] Fig.15 This is a data structure diagram of a list of industry categories.
[0128] When a search query is input in the utilization phase, an industry category list is generated accordingly. The industry category list is stored in the associated phrase storage unit 140. The industry category list stores the similarity between each industry category and the phrase of the search query in association with each other. The industry category list is sorted in descending order of similarity.
[0129] exist Fig.11 In the example, the first benchmark value is 0.7, so the industry categories "car dealers", "car repair industry", and "car rental industry" are selected. Fig.12 In the example, the first base value is 0.6, so the industry categories "automakers", "transportation industry", and "house sales" are added.
[0130] Fig.16 is Fig.11 Figure 1 shows the company information box included in the search result screen in the state of .
[0131] In this example, the character string of the business characteristic word "senior" similar to the search query is consistent with the business characteristic word name, so the company information is displayed as a search result. In addition, the character string "senior" is hit when it is included in the business content sentence set, the main industry category name, or the secondary industry category name. That is, the business content sentence set, the main industry category name, or the secondary industry category name is also set as the search scope. However, the business content sentence set, the main industry category name, or the secondary industry category name may not be set as the search scope.
[0132] Fig.17 is Fig.12 Figure 1 shows the company information box included in the search result screen in the state of .
[0133] In this example, the character string of the industry category "housing seller" similar to the search query is consistent with the main industry category name, so the company information is displayed as the search result. In addition, the character string "housing seller" is hit when it is included in the business content sentence set, business characteristic words, and the main industry category name. In other words, the business content sentence set, business characteristic words, or sub-industry category name are also set as the search scope. However, the business content sentence set, business characteristic words, or sub-industry category name may not be set as the search scope.
[0134] Based on these screen examples, an application example is shown. For example, suppose an insurance company that wants to find new contract candidates for automobile insurance uses this search system. The insurance company first enters "automobile" as a search query and obtains Fig.10 's search results. Fig.10 The car dealers shown are unlikely to become new transaction channels because the promotion of car insurance through car dealers has always been common.
[0135] So, if Fig.11 As shown, operating the slider 302 increases the comprehensiveness and obtains the Fig.16 The company information of "Red Rose" does not contain the character string "car", but it is actually related to "car" through the business characteristic word "luxury" which is highly related to "car". The insurance company observed the business content of "Red Rose" and believed that the assistants dispatched by "Red Rose" might cause high compensation when driving luxury cars of wealthy customers. Then, the insurance company found that it could propose an insurance contract to "Red Rose".
[0136] Then, if Fig.12 As shown, the slider 302 is operated to further improve the comprehensiveness and obtain the search results. Fig.17Although the character string "car" is not included in the company information of "yellow lemon", it is actually related to "car" through the business "housing seller" which is highly related to "car". Insurance companies observe the content of this business and believe that the occupants must have a car because they buy a house with a parking space. In addition, choosing employees who are familiar with the local conditions means that the occupants are not familiar with the traffic conditions in the area and have few acquaintances. Therefore, if they are worried about traffic accidents and disputes with other residents in an unknown area, they will realize that this is a motivation for joining the insurance. In this way, they can constantly find clues to expand their business.
[0137] In this example, it is difficult for the insurance company to notice that "luxury" or "housing seller" is a keyword for expanding business. However, according to this embodiment, the target company information can be found from the common word "car".
[0138] Since there are many differences in the tags used by people to describe content, the label information assigned manually cannot guarantee comprehensiveness, so according to this embodiment, comprehensiveness can be supplemented by auxiliary search. The following is a detailed description of the process.
[0139] Fig.18 is a functional block diagram of the server 100 .
[0140] The components of the server 100 are implemented by a CPU (Central Processing Unit) and various coprocessors, memory and storage devices, hardware including wired or wireless communication lines connecting them, and software stored in the storage device and supplying processing commands to the arithmetic unit. Computer programs can also be composed of device drivers, operating systems, various application programs located above them, and program libraries that provide common functions to these programs. The blocks shown in the figure are not the structure of hardware units, but blocks representing functional units.
[0141] The server 100 includes a data storage unit 110, a data processing unit 150, and a communication unit 190. The communication unit 190 is responsible for communication processing via a network. The data storage unit 110 stores various data. The data processing unit 150 performs various processes based on the data obtained through 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.
[0142] The communication unit 190 includes a transmission unit 180 for transmitting data and a reception unit 170 for receiving data.
[0143] The transmitting unit 180 includes a screen data transmitting unit 182 for transmitting 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 a search query. The slider value receiving unit 174 receives a slider value.
[0144] The data processing unit 150 includes a model generating unit 152 , a company vector calculating unit 154 , a business characteristic word vector calculating unit 156 , an industry category vector calculating unit 158 , a screen data generating unit 160 , and a searching unit 162 .
[0145] The model generation unit 152 performs word segmentation expression model generation processing for calculating word vectors. The enterprise vector calculation unit 154 performs enterprise vector calculation processing. The business characteristic word vector calculation unit 156 performs business characteristic word vector calculation processing. The industry category vector calculation unit 158 performs industry category vector calculation processing. The screen data generation unit 160 generates various screen data. The search unit 162 searches for enterprise information.
[0146] The data storage unit 110 includes a company database 120 , a word segmentation expression model storage unit 130 , and a related sentence storage unit 140 .
[0147] Fig.19 This is a functional block diagram of the data storage unit 110 .
[0148] The enterprise database 120 includes an enterprise table storage unit 122, a business characteristic word table storage unit 124, and an industry category table storage unit 126. The enterprise table storage unit 122 stores the enterprise table ( Figure 2 The business characteristic word list storage unit 124 stores the business characteristic word list ( Figure 3 The industry category table storage unit 126 stores the industry category table ( Figure 4 ).
[0149] The word segmentation expression model storage unit 130 includes a word list storage unit 132 , a business vector storage unit 134 , a business characteristic word vector storage unit 136 , and an industry category vector storage unit 138 .
[0150] The word list storage unit 132 stores the word list ( Figure 5 ). For the enterprise vector storage unit 134 and Figure 6 The business characteristic word vector storage unit 136 and Figure 7 The industry category vector storage unit 138 and Figure 8 Explain in relation.
[0151] The related sentence storage unit 140 includes a similar word list storage unit 142 , a business characteristic word list storage unit 144 , and an industry category list storage unit 146 .
[0152] The similar word list storage unit 142 stores a similar word list ( Fig.13 The business characteristic word list storage unit 144 stores a business characteristic word list ( Fig.14 The industry category list storage unit 146 stores the industry category list ( Fig.15 ).
[0153] Fig. 20 It is a flowchart showing the process of the preparation phase.
[0154] First, the model generation unit 152 performs a word segmentation expression model generation process (S20). In the word segmentation expression model generation process, a word list ( Figure 5 ) as the word segmentation expression model. Fig.21 This will be described later in connection with the above.
[0155] Next, the enterprise vector calculation unit 154 performs enterprise vector calculation processing (S22). Fig. 22 This will be described later in connection with the above.
[0156] Next, the business-characteristic word vector calculation unit 156 performs a business-characteristic word vector calculation process (S24). Fig.23 This will be described later in connection with the above.
[0157] Finally, the industry category vector calculation unit 158 performs industry category vector calculation processing (S26). Fig.24 This will be described later in connection with the above.
[0158] Fig.21 It is a flowchart showing the process of generating a word segmentation expression model.
[0159] The model generation unit 152 generates a corpus from the enterprise table (S30). The corpus is a large-scale data formed by structuring articles in natural language. The corpus is stored in the data storage unit 110. Specifically, the model generation unit 152 performs morphological analysis on each sentence extracted from the business content sentence set of each enterprise and converts it into a delimited writing format. Delimited writing is a character string in which words are separated by spaces. Data in the delimited writing format becomes a corpus.
[0160] The model generation unit 152 generates word vectors using the corpus (S32). Specifically, the model generation unit 152 obtains word vectors for each word included in the corpus through, for example, Word2vec learning processing. The generated word vectors are stored in the word table storage unit 132. Then, the process returns to S22.
[0161] Fig. 22 is a flow chart showing the process of enterprise vector calculation processing.
[0162] The enterprise vector calculation unit 154 repeatedly performs the following processing for each enterprise (S40). The enterprise vector calculation unit 154 calculates the sentence vector of each sentence contained in the enterprise's business content sentence set (S42). The sentence vector is generated from the word vectors 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.
[0163] The enterprise vector calculation unit 154 calculates the average of the sentence vectors of each sentence included in the business content sentence set, and uses the calculated average vector as the enterprise vector (S44). The enterprise vector calculation unit 154 may generate an enterprise vector from the sentence vectors by a method other than averaging. The calculated enterprise vector is stored in the enterprise vector storage unit 134.
[0164] If there are still unprocessed companies (N in S46), the company vector calculation unit 154 returns to the process of S40. When the processing of all companies is completed (Y in S46), the process returns to S24.
[0165] Fig.23 This is a flowchart showing the process of calculating the business characteristic word vector.
[0166] The business characteristic word vector calculation unit 156 repeatedly performs 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 for which the business characteristic word is set (S52).
[0167] The business characteristic word vector calculation unit 156 calculates the average vector of the selected enterprise vectors, and uses the calculated average vector as the business characteristic word vector of the business characteristic word (S54). The business characteristic word vector calculation unit 156 may generate a business characteristic word vector from the enterprise vector by a method other than averaging. The calculated business characteristic word vector is stored in the business characteristic word vector storage unit 136.
[0168] If there are still unprocessed business characteristic words (N in S56), the business characteristic word vector calculation unit 156 returns to the process of S50. When the processing of all business characteristic words is completed (Y in S56), the process returns to S26.
[0169] Fig.24 It is a flowchart showing the process of industry category vector calculation processing.
[0170] The industry category vector calculation unit 158 repeats the following processing for each industry category (S60). The industry category vector calculation unit 158 refers to the company table and selects a company for which the industry category is set (S62).
[0171] The industry category vector calculation unit 158 calculates the average vector of the selected enterprise vectors and uses the calculated average vector as the industry category vector of the industry category. The industry category vector calculation unit 158 may generate the industry category vector from the enterprise vectors by a method other than averaging. The calculated industry category vector is stored in the industry category vector storage unit 138.
[0172] If there is still an unprocessed industry category (N in S66), the industry category vector calculation unit 158 returns to the process of S60. When the process is completed for all industry categories (Y in S66), the preparation phase process is terminated.
[0173] Fig.25 It is a flowchart showing the process of utilizing the phase processing.
[0174] When the search query receiving unit 172 receives a search query from the user terminal 200 displaying the search query acceptance screen (Y in S70), the screen data generating unit 160 executes a list generating process (S72). In the list generating process, a similar word list ( Fig.13 )、Business Characteristic Word List( Fig.14 ) and industry category list ( Fig.15 ). For list generation processing and Fig.26 This will be described later in connection with the above.
[0175] Next, the screen data generation unit 160 executes screen generation processing (S74). In the screen generation processing, similar words, business characteristic words, and industry categories are displayed again on the search query acceptance screen. In addition, the company information is searched and the search result screen is displayed. Fig. 27 This will be described later in connection with the above.
[0176] Next, 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 S70. The user terminal 200 displays the search query acceptance screen and the search result screen based on the received screen data.
[0177] When the slider value receiving unit 174 receives the slider value from the user terminal 200 displaying the search query acceptance screen (Y in S78), the first reference value, the second reference value, and the third reference value are changed based on the slider value. In this example, the slider value is used as the new first reference value, the new second reference value, and the new third reference value.
[0178] When the first reference value, the second reference value, and the third reference value are changed, the screen data generation unit 160 performs screen generation processing based on the changed first reference value, the second reference value, and the third reference value (S82). In addition, the screen data transmission unit 182 transmits the newly generated screen data in the screen data transmission processing (S84). The user terminal 200 displays the acceptance screen of the search query and the screen of the search result based on the received screen data. Then, the process returns to S70.
[0179] If an end instruction is received from the user terminal 200 (Y in S86), the use phase process is ended. If an end instruction is not received from the user terminal 200, the process returns to S70.
[0180] Fig.26 This is a flowchart showing the procedure of the list generation process.
[0181] The screen data generation unit 160 calculates the similarity between the whole word and the search query (S90). If the search query is a single word, the word vector of the word is set as the semantic vector of the search query. If the search query is a phrase, the semantic vector of the search query is generated from the vectors of the words contained in the phrase. The generation method is, for example, averaging. Then, the similarity between the word vector of each word and the semantic vector of the search query is calculated. The screen data generation unit 160 calculates, for example, cosine similarity. The same applies to the following similarities.
[0182] The screen data generation unit 160 sorts the words in the order of high similarity and generates a similar word list ( S92 ). The generated similar word list is stored in the similar word list storage unit 142 .
[0183] The screen data generation unit 160 calculates the similarity between all business characteristic words and the search query (S94). Specifically, the screen data generation unit 160 obtains the similarity between the business characteristic word vector of each business characteristic word and the semantic vector of the search query.
[0184] The screen data generation unit 160 sorts the business characteristic words in the order of high similarity, and generates a business characteristic word list ( S96 ). The generated business characteristic word list is stored in the business characteristic word list storage unit 144 .
[0185] The screen data generation unit 160 calculates the similarity between all industry categories and the search query (S98). Specifically, the screen data generation unit 160 obtains the similarity between the industry category vector of each industry category and the semantic vector of the search query.
[0186] The screen data generation unit 160 sorts the industry categories in the order of high similarity and generates an industry category list (S100). The generated industry category list is stored in the industry category list storage unit 146. Then, the process returns to S74.
[0187] Fig. 27 This is a flowchart showing the procedure of screen generation processing.
[0188] The screen data generation unit 160 arranges similar words having a similarity greater than or equal to the first reference value in the display area 304 of the similar words in the search query acceptance screen (S110). The screen data generation unit 160 arranges business characteristic words having a similarity greater than or equal to the second reference value in the business characteristic word area 306 in the search query acceptance screen (S112). In addition, the screen data generation unit 160 arranges industry categories having a similarity greater than or equal to the third reference value in the industry category area 308 in the search query acceptance screen (S114).
[0189] The screen data generation unit 160 searches for enterprise information under the OR condition of these similar words, business characteristic words, and industry categories (S116). For example, a new search query is generated as a character string for searching these similar words, business characteristic words, and industry categories, and the search query is input into a database management system (DataBase Management System). The database management system compares the enterprise information of each enterprise (a group of texts representing a business content sentence set, business characteristic words, main industry category, and sub-industry category) and selects the corresponding enterprise information. The database management system is software that manages the enterprise database 120 and operates the enterprise database 120.
[0190] The screen data generation unit 160 arranges a company information frame displaying the hit company information on the search result screen ( S118 ). Then, the process returns to S70 .
[0191] [Modifications]
[0192] It is also possible to delete unnecessary words from the similar words, business characteristic words, and industry categories displayed on the acceptance screen of the search query. When unnecessary similar words, business characteristic words, or industry categories are indicated by user operation, a deletion request for the similar words, business characteristic words, or industry categories is sent 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 the unnecessary similar words, business characteristic words, or industry categories from the similar word list, the business characteristic word list, or the industry category list. Then, the server 100 performs the list generation process, the screen generation process, and the screen data transmission process again.
[0193] Although an example is shown in which the company information is automatically searched in conjunction with the reception of a search query, the company information may be searched after an instruction to execute a search is received through a user operation.
[0194] The screen data generation unit 160 may also configure the similarity between the search query and the similar word near the similar word in the search query acceptance screen. The screen data generation unit 160 may also configure the similarity between the search query and the business characteristic word near the business characteristic word in the search query acceptance screen. The screen data generation unit 160 may also configure the similarity between the search query and the industry category near the industry category in the search query acceptance screen.
[0195] Although an example is shown in which the search query acceptance screen and the search result screen are displayed side by side, the search query acceptance screen and the search result screen may be displayed in a switching manner.
[0196] The server 100 may convert the slider value to obtain the first reference value, the second reference value, and the third reference value. In addition, 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 in a manner related to the slider value. The screen data generation unit 160 may multiply the slider value by the first coefficient to obtain the first reference value. The screen data generation unit 160 may multiply the slider value by the second coefficient to obtain the second reference value. The screen data generation unit 160 may multiply the slider value by the third coefficient to obtain the third reference value.
[0197] In the averaging of vectors, for example, an evaluation index such as TF-IDF (Term Frequency-Inverse Document Frequency) may be used to weight each original word vector, each phrase vector, or each sentence vector.
[0198] Although an example of searching for company information is shown, it is also possible to search for information of other types. For example, when searching for product information on an e-commerce transaction website, it is also possible to calculate a product classification vector and a product characteristic vector for labels of product classification and labels of product characteristics attached to a product introduction sentence set. Furthermore, it is also possible to supplement the search condition using a product classification vector or a product characteristic vector similar to the search query.
[0199] For example, when searching for patent documents on a patent document browsing website, the patent classification vector can be calculated for the patent classification label attached to the patent document. In addition, the patent classification vector similar to the search query can be used to supplement the search condition.
[0200] For example, when searching for posted messages on a posting site of an SNS (Social Networking Service), a message classification vector may be calculated for a message classification tag such as a hash tag attached to the posted message. In addition, a message classification vector similar to a search query may be used to supplement the search condition.
[0201] In addition, the present invention is not limited to the above-mentioned embodiments and variations, and the constituent elements can be deformed and concretized within the scope of the intent. Various inventions can also be formed by appropriately combining a plurality of constituent elements disclosed in the above-mentioned embodiments and variations. In addition, some constituent elements can also be deleted from all the constituent elements shown in the above-mentioned embodiments and variations.
Claims
1. A natural language processing device, characterized in that: include: a first calculation unit for calculating word segmentation vectors of words contained in a database managing a plurality of sentences associated with a label, wherein the label is at least one of a business characteristic word and an industry category related to an enterprise, and the plurality of sentences are a set of business content sentences of the enterprise to which the label is set; a second calculation unit that calculates a word segmentation vector of the sentence by averaging the word segmentation vectors of the words included in each sentence; and A third calculation unit calculates a word segmentation vector of the label by averaging the word segmentation vectors of each sentence associated with the same label, The word segmentation vector of the label is at least one of a business characteristic word vector representing the meaning of the business characteristic word and an industry category vector representing the meaning of the industry category, When searching for corporate information, upon receiving a search query input by a user, business characteristic words or industry categories represented by word vectors having a similarity with a semantic vector of the search query that is greater than a reference value are added to the search query for retrieval.
2. The natural language processing device according to claim 1, characterized in that: Also includes: An input unit capable of inputting sentences; A selection unit capable of selecting a label of a word segmentation vector similar to the word segmentation vector of the input sentence; and An output unit is capable of outputting the selected tag.
3. The natural language processing device according to claim 2, wherein: The selection unit selects a plurality of tags, The output unit arranges and outputs the selected multiple tags in order of similarity with the word segmentation vector of the input sentence.
4. The natural language processing device according to claim 3, wherein: The output unit outputs the similarity between the word segmentation vector of the input sentence and the word segmentation vector of the selected label.
5. A computer storage medium storing a program, characterized in that: The program enables the information processing device to perform the following functions: A first calculation function is to calculate word segmentation vectors of words contained in a database that manages a plurality of sentences associated with a label, wherein the label is at least one of a business characteristic word and an industry category related to an enterprise, and the plurality of sentences are a set of business content sentences of the enterprise to which the label is set; The second calculation function calculates the word segmentation vector of the sentence by averaging the word segmentation vectors of the words contained in each sentence; and The third calculation function calculates the word vector of the label by averaging the word vectors of each sentence associated with the same label. The word segmentation vector of the label is at least one of a business characteristic word vector representing the meaning of the business characteristic word and an industry category vector representing the meaning of the industry category, When searching for corporate information, upon receiving a search query input by a user, business characteristic words or industry categories represented by word vectors having a similarity with a semantic vector of the search query that is greater than a reference value are added to the search query for retrieval.
