Information processing system, information processing method, and information processing program
The information processing system uses distributed representation vectors and machine learning models to enhance the accuracy of estimating the relationship between a target text and a specific concept by calculating similarity and association, addressing the limitations of existing text processing technologies.
Patent Information
- Application Number
- JP2023077464
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2023-05-09
- Publication Date
- 2024-11-21
- Estimated Expiration
- 2043-05-09
AI Technical Summary
Existing text processing technologies struggle to accurately estimate the relationship between a target text and a specific concept using machine learning models.
An information processing system that utilizes distributed representation vectors for target texts and background knowledge related to a concept, calculating similarity and association through machine learning models to determine the degree of relevance between the two.
Accurately estimates the degree to which a target text is related to a specific concept by leveraging background knowledge and machine learning models, enhancing the precision of text analysis.
Smart Images

Figure 2024162168000001_ABST
Abstract
Description
[Technical field]
[0001] One aspect of the present disclosure relates to an information processing system, an information processing method, and an information processing program. [Background technology]
[0002] There is known a technique for processing text using a machine learning model. For example, Patent Document 1 describes a category estimation device that can improve the accuracy of estimating a category of a sentence that includes an unknown word. [Prior art documents] [Patent documents]
[0003] [Patent Document 1] JP 2020-161075 A Summary of the Invention [Problem to be solved by the invention]
[0004] It is desirable to accurately estimate how relevant a piece of text is to a particular concept. [Means for solving the problem]
[0005] An information processing system according to an aspect of the present disclosure includes at least one processor, which acquires a distributed representation vector of a target text as a target vector, acquires a distributed representation vector of background knowledge related to a target concept as a knowledge vector, calculates a similarity between the target vector and the knowledge vector, inputs the calculated similarity to a machine learning model that calculates a relevance between the target text and the target concept based on the similarity, and acquires a relevance between the target text and the target concept from the machine learning model.
[0006] An information processing method according to an aspect of the present disclosure is executed by an information processing system including at least one processor, and includes the steps of: acquiring a distributed representation vector of a target text as a target vector; acquiring a distributed representation vector of background knowledge related to a target concept as a knowledge vector; calculating a similarity between the target vector and the knowledge vector; and inputting the calculated similarity into a machine learning model that calculates a relevance between the target text and the target concept based on the similarity, and acquiring a relevance between the target text and the target concept from the machine learning model.
[0007] An information processing program according to one aspect of the present disclosure causes a computer to execute the steps of obtaining a distributed representation vector of a target text as a target vector, obtaining a distributed representation vector of background knowledge related to the target concept as a knowledge vector, calculating a similarity between the target vector and the knowledge vector, and inputting the calculated similarity into a machine learning model that calculates a degree of relevance between the target text and the target concept based on the similarity, and obtaining a degree of relevance between the target text and the target concept from the machine learning model.
[0008] In this aspect, the similarity between the target text and background knowledge related to the target concept is first calculated, and the relevance between the target text and the target concept is obtained from the similarity by a machine learning model. By using the similarity with the background knowledge related to the target concept and the machine learning model in this way, it is possible to accurately estimate the degree to which the target text is related to the target concept. Effect of the Invention
[0009] According to one aspect of the present disclosure, it is possible to accurately estimate the degree to which text is related to a particular concept. [Brief description of the drawings]
[0010] [Figure 1] FIG. 1 is a diagram illustrating an example of an outline of an operation of an information processing system. [Diagram 2]FIG. 2 is a diagram illustrating an example of a functional configuration of an information processing system. [Diagram 3] FIG. 2 is a diagram illustrating an example of a hardware configuration of a computer that constitutes the information processing system. [Figure 4] 11 is a flowchart illustrating an example of processing in an information processing system. [Diagram 5] 13 is a flowchart illustrating an example of a process for acquiring a target vector. [Figure 6] 13 is a flowchart illustrating an example of a process for acquiring a knowledge vector. [Figure 7] FIG. 11 is a diagram illustrating an example of a process for acquiring a target vector. [Figure 8] FIG. 13 is a diagram illustrating an example of a process for acquiring a knowledge vector. [Figure 9] FIG. 13 is a diagram illustrating an example of a process for acquiring a knowledge vector. [Figure 10] FIG. 13 is a diagram illustrating an example of a process for acquiring a knowledge vector. [Figure 11] FIG. 13 is a diagram illustrating an example of a process for acquiring a knowledge vector. [Figure 12] FIG. 11 is a diagram illustrating an example of a process for acquiring a similarity. [Figure 13] FIG. 11 is a diagram illustrating an example of a process for acquiring a degree of association. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0011] Hereinafter, embodiments of the present disclosure will be described in detail with reference to the accompanying drawings. In the description of the drawings, the same or equivalent elements are denoted by the same reference numerals, and duplicated descriptions will be omitted.
[0012] [System Overview] The information processing system according to the present disclosure is a computer system that estimates the relevance between a target text and a target concept. The target concept refers to a common meaning for a specific similar thing. The target concept may be, for example, a specific brand, product, service, company, group, team, organization, person, publication, work of art, paper, variety, region, event, experience, right, information, or idea. The target text refers to a text to be examined for its relevance to the target concept. The target text is expressed by a character string and is composed of, for example, one or more sentences. The target text may be obtained from various information sources such as a website or a social networking site. Both the target concept and the target text may be specified by a user of the information processing system. The relevance refers to an index indicating the degree to which the target text is related to the target concept.
[0013] The information processing system acquires the relevance between the target text and the target concept by utilizing background knowledge of the target concept. Background knowledge refers to information related to the target concept. Background knowledge includes, for example, the formal name, abbreviation, or colloquial name of the target concept, or homographs, homonyms, synonyms, or co-occurring words of such names. Here, a synonym refers to another word that is similar in meaning to a certain word. A synonym may be another word that is close to a certain word in the word embedding space or has a close cosine similarity. A synonym may also be another word obtained from a certain word by a thesaurus. Background knowledge may also include a superordinate concept of the target concept, a concept that is similar in concept to the target concept, or a person, organization, or region associated with the target concept. As an example, if the target concept is a specific brand, the background knowledge may include the product name of the brand, keywords or topics of the brand, and the founder or birthplace of the brand. In the information processing system, information constituting the background knowledge may be grouped into a plurality of information categories, and in this disclosure, each information category is also referred to as a knowledge domain. For example, if the target concept is a particular brand, one or more knowledge domains may be established, such as product names, brand keywords, brand topics, and the like.
[0014] The information processing system performs the calculations related to the relevance using a machine learning model. Machine learning is a method of autonomously finding laws or rules by repeatedly learning based on given information. A machine learning model is a computational model obtained by machine learning, and is therefore also called a trained model.
[0015] The information processing system acquires a distributed representation vector of a target text as a target vector, and acquires a distributed representation vector of background knowledge related to a target concept as a knowledge vector. The distributed representation vector is a vector that expresses a word or text as a set of multi-dimensional numerical values. The information processing system calculates a similarity between the target vector and a knowledge vector. This similarity is an index showing how similar two vectors are. The information processing system calculates a relevance based on the similarity. In one example, the information processing system uses a machine learning model for acquiring the distributed representation vector and another machine learning model for calculating the relevance from the similarity. As described above, both of these machine learning models are trained models generated in advance by machine learning. Examples of machine learning models for acquiring the distributed representation vector include fastText, Word2vec, GloVe, ELMo, and BERT. Examples of machine learning models for calculating the relevance include logistic regression.
[0016] 1 is a diagram illustrating an example of an outline of the operation of an information processing system as a processing flow S1. In this example, N knowledge domains (N is a natural number) are shown as background knowledge.
[0017] The information processing system converts the distributed representation vector of the target text into the target vector W x As an example, the information processing system acquires the target vector W by using a first vector acquisition model 11a, which is a machine learning model. x Get the.
[0018] The information processing system processes knowledge domains K1 to K2 contained in background knowledge related to the target concept. NEach distributed representation vector of N As an example, the information processing system acquires the knowledge vectors W1 to W2 by using a second vector acquisition model 12a, which is a machine learning model. N The second vector acquisition model 12a may be a machine learning model common to a plurality of knowledge domains that can calculate a distributed representation vector for each of the plurality of knowledge domains. Alternatively, the second vector acquisition model 12a may be provided individually for each knowledge domain.
[0019] The information processing system calculates the knowledge vectors W1 to W N For each of the target vectors W x Similarity with R1~R N (Step S13). Examples of the similarity include cosine similarity, Euclidean distance, and Manhattan distance.
[0020] The information processing system has similarity scores R1 to R N Based on this, the degree of relevance between the target text and the target concept is acquired (step S14). As an example, the information processing system acquires the degree of relevance using a relevance acquisition model 14a, which is a machine learning model.
[0021] In this way, the information processing system can process the target vector W x and knowledge vector W1~W N First, obtain the target vector W x and knowledge vector W1~W N The relevance between the target text and the target concept is obtained from the similarity with each of the above. By using the similarity with the background knowledge related to the target concept and the machine learning model in this way, it is possible to accurately estimate the degree to which the target text is related to the target concept.
[0022] [System configuration] 2 is a diagram showing an example of a functional configuration of the information processing system 10. In one example, the information processing system 10 includes a first vector acquisition unit 11, a second vector acquisition unit 12, a similarity calculation unit 13, and an association degree acquisition unit 14 as functional modules. The information processing system 10 can access a database group 20 via a communication network. The communication network may include at least one of a mobile communication network, the Internet, an intranet, and a WAN (Wide Area Network).
[0023] The first vector acquisition unit 11 is a functional module that acquires a distributed representation vector of a target text as a target vector. As an example, the first vector acquisition unit 11 causes a first vector acquisition model 11a to calculate the target vector.
[0024] The second vector acquisition unit 12 is a functional module that acquires a distributed representation vector of background knowledge related to a target concept as a knowledge vector. As an example, the second vector acquisition unit 12 causes a second vector acquisition model 12a to calculate the knowledge vector.
[0025] The similarity calculation unit 13 is a functional module that calculates the similarity between the target vector acquired by the first vector acquisition unit 11 and the knowledge vector acquired by the second vector acquisition unit 12.
[0026] The relevance acquisition unit 14 is a functional module that acquires the relevance between the target text and the target concept based on the similarity calculated by the similarity calculation unit 13. As an example, the relevance acquisition unit 14 causes the relevance acquisition model 14a to calculate the relevance.
[0027] The database group 20 is a collection of databases that store data required by the information processing system 10. As an example, the database group 20 includes a text database 21 and a background knowledge database 22. Each database may be configured as a part of the information processing system 10, or may be provided in a computer system separate from the information processing system 10. The information processing system 10 can access each database to read and write data.
[0028] The text database 21 is a non-transitory storage medium or storage device that stores data of the target text. A collection of multiple databases may be used as the text database 21.
[0029] The background knowledge database 22 is a non-transitory storage medium or storage device that stores background knowledge data. A collection of multiple databases may be used as the background knowledge database 22.
[0030] FIG. 3 is a diagram showing a general hardware configuration of a computer 100 that can function as the information processing system 10. For example, the computer 100 includes a processor 101, a main memory unit 102, an auxiliary memory unit 103, a communication control unit 104, an input device 105, and an output device 106. The processor 101 executes an operating system and an application program. The main memory unit 102 is composed of, for example, a ROM and a RAM. The auxiliary memory unit 103 is composed of, for example, a hard disk or a flash memory, and generally stores a larger amount of data than the main memory unit 102. The auxiliary memory unit 103 non-temporarily stores an information processing program 110 for causing at least one computer to function as the information processing system 10. The communication control unit 104 is composed of, for example, a network card or a wireless communication module. The input device 105 is composed of, for example, a keyboard, a mouse, a touch panel, etc. The output device 106 is composed of, for example, a monitor and a speaker.
[0031] Each functional element of the information processing system 10 is realized by loading an information processing program 110 onto the processor 101 or the main memory unit 102 and having the processor 101 execute the information processing program 110. The information processing program 110 includes code for realizing each functional element of the information processing system 10. The processor 101 operates the communication control unit 104, the input device 105, or the output device 106 in accordance with the information processing program 110, and reads and writes data in the main memory unit 102 or the auxiliary memory unit 103. Each functional element of the information processing system 10 is realized by this processing. Data or a database required for processing may be stored in the main memory unit 102 or the auxiliary memory unit 103.
[0032] The information processing program 110 may be provided after being non-temporarily recorded on a tangible recording medium such as a CD-ROM, a DVD-ROM, a semiconductor memory, etc. Alternatively, the information processing program 110 may be provided via a communication network as a data signal superimposed on a carrier wave.
[0033] The information processing system 10 may be configured with one or more computers. When multiple computers are used, the information processing system 10 is configured by connecting these computers to each other via a communication network.
[0034] [System Processing] The processes executed by the information processing system 10 will be described in detail with reference to Figs. 4 to 6. Fig. 4 is a flowchart showing process flow S1. The information processing system 10 executes process flow S1. Fig. 5 is a flowchart showing an example of step S11. Fig. 6 is a flowchart showing an example of step S12.
[0035] In step S11, the first vector acquisition unit 11 converts the distributed representation vector of the target text into a target vector W xIn one example, the first vector acquisition unit 11 accesses the text database 21 to read the target text, processes the target text using the first vector acquisition model 11a, and acquires the target vector W x The details of step S11 will be described with reference to FIG.
[0036] In step S111, the first vector acquisition unit 11 executes morphological analysis on the target text to acquire one or more words from the target text. The first vector acquisition unit 11 acquires one or more words based on one or more morphemes acquired by the morphological analysis. A morpheme is the smallest unit of an expression element having meaning. Morphological analysis is a process of breaking down a character string into morphemes. In one example, the first vector acquisition unit 11 acquires one or more nouns as the one or more words, and does not acquire words of other parts of speech.
[0037] In step S112, the first vector acquisition unit 11 acquires a distributed representation vector for each of the one or more acquired words. The first vector acquisition unit 11 executes the following process for each word. That is, the first vector acquisition unit 11 inputs the word to the first vector acquisition model 11a. The first vector acquisition model 11a calculates and outputs a distributed representation vector based on the word. The first vector acquisition unit 11 acquires the distributed representation vector.
[0038] In step S113, the first vector acquisition unit 11 calculates a target vector W based on the one or more acquired distributed representation vectors. x For example, the first vector acquisition unit 11 calculates a representative value of one or more distributed representation vectors as a target vector W xThe representative value is a value calculated by statistical processing, and may be, for example, an average value or a median value. The average value may be an arithmetic mean (arithmetic mean), a geometric mean (geometric mean), or a harmonic mean. When one or more words are classified into a plurality of parts of speech, the first vector acquisition unit 11 may weight the distributed representation vectors by weights preset for each part of speech, and calculate a representative value of the weighted distributed representation vectors. For example, the first vector acquisition unit 11 may weight nouns higher than other parts of speech, or may weight proper nouns higher than general nouns.
[0039] Returning to FIG. 4, in step S12, the second vector acquisition unit 12 acquires one or more knowledge domains K1 to K2 included in background knowledge related to the target concept. N One or more distributed representation vectors for each of the knowledge vectors W1 to W N As an example, the second vector acquisition unit 12 uses a second vector acquisition model 12a. Details of step S12 will be described with reference to FIG.
[0040] In step S121, the second vector acquisition unit 12 selects one knowledge domain to be processed from one or more knowledge domains constituting the background knowledge.
[0041] In step S122, the second vector acquisition unit 12 acquires one or more words belonging to the selected knowledge domain. In one example, the second vector acquisition unit 12 accesses the background knowledge database 22 and reads out one or more words corresponding to the knowledge domain. Alternatively, the second vector acquisition unit 12 may read out one or more texts corresponding to the knowledge domain from the background knowledge database 22 and perform morphological analysis on each text to acquire one or more words.
[0042] In step S123, the second vector acquisition unit 12 acquires a distributed representation vector for each of the acquired one or more words. The second vector acquisition unit 12 executes the following process for each word. That is, the second vector acquisition unit 12 inputs the word to the second vector acquisition model 12a. The second vector acquisition model 12a calculates and outputs a distributed representation vector based on the word. The second vector acquisition unit 12 acquires the distributed representation vector.
[0043] In step S124, the second vector acquisition unit 12 calculates a knowledge vector based on the one or more acquired distributed representation vectors. As an example, the second vector acquisition unit 12 calculates a representative value of one or more distributed representation vectors as a knowledge vector. For example, the representative value may be an average value or a median value. The average value may be an arithmetic mean (arithmetic mean), a geometric mean (geometric mean), or a harmonic mean. When one or more words are classified into a plurality of parts of speech, the second vector acquisition unit 12 may weight the distributed representation vectors by a weight preset for each part of speech, and calculate a representative value of the weighted distributed representation vector. For example, the second vector acquisition unit 12 may weight nouns higher than other parts of speech, or may weight proper nouns higher than general nouns.
[0044] If there is a knowledge domain that has not yet been processed (NO in step S125), the process returns to step S121. In the repeated step S121, the second vector acquisition unit 12 selects the next knowledge domain that has not yet been processed. In the repeated steps S122 to S124, the second vector acquisition unit 12 calculates a knowledge vector for that knowledge domain. Knowledge domains K1 to K N When the process is performed for all of the knowledge vectors W1 to W N If all of these have been acquired (YES in step S125), the second vector acquisition unit 12 ends step S12.
[0045] Returning to FIG. 4, in step S13, the similarity calculation unit 13 calculates the similarity between one or more knowledge vectors W1 to W NFor each of the target vectors W x Calculate the similarity with R1 to R2, and set the similarity to R1. N Get the.
[0046] In step S14, the relevance obtaining unit 14 obtains the similarities R1 to R N Based on this, the relevance degree acquisition unit 14 acquires the relevance degree between the target text and the target concept. N The relevance acquisition model 14a inputs the similarities R1 to R N The relevance degree acquisition unit 14 calculates and outputs the relevance degree based on the above. The relevance degree acquisition unit 14 acquires and outputs the relevance degree. For example, the relevance degree acquisition unit 14 may display the relevance degree on a display device, may store it in a given database, or may transmit it to another computer or computer system. The relevance degree acquisition unit 14 may express the relevance degree by a numerical value such as the probability that the target text is related to the target concept, may express it by a character string such as "related" or "not related," or may express it by a symbol indicating the presence or absence of relevance.
[0047] An example of the process flow S1 will be described with reference to Figures 7 to 13. In the examples shown in these figures, the target concept is a cosmetics brand named "XYZ."
[0048] FIG. 7 is a diagram showing an example of a process (step S11) for acquiring a target vector. In this example, the first vector acquisition unit 11 acquires a target text "XYZ eye shadow is monochromatic and has a good shine" from the text database 21. The first vector acquisition unit 11 performs a morphological analysis on the target text to acquire four nouns (words), namely "XYZ," "eye shadow," "monochromatic," and "shiny" (step S111). The first vector acquisition unit 11 acquires a distributed representation vector for each of the four words using the first vector acquisition model 11a (step S112). The first vector acquisition unit 11 inputs "XYZ" to the first vector acquisition model 11a, and acquires the distributed representation vector W of the word from the first vector acquisition model 11a. x1Similarly, the first vector acquisition unit 11 acquires a distributed representation vector W x2 , W x3 and W x4 The first vector acquisition unit 11 acquires the distributed representation vector W x1 ~W x4 Based on the target vector W x is calculated (step S113).
[0049] 8 to 11 are diagrams showing an example of the process (step S12) for acquiring a knowledge vector. In this example, background knowledge about a cosmetic brand "XYZ" is composed of five knowledge domains: the brand name, top related co-occurrence words, top unrelated co-occurrence words, brand keywords, and brand topics.
[0050] 8 is a diagram showing a process for acquiring a knowledge vector W1 of a brand name. Examples of words belonging to a brand name include the brand's official name, abbreviation, and colloquial name, as well as spelling variations and synonyms of such names.
[0051] The second vector acquisition unit 12 acquires one or more words belonging to the brand name from the background knowledge database 22 (step S122). FIG. 8 shows "XYZ", "Xyz", and "xyz" as examples of such words. The second vector acquisition unit 12 acquires a distributed representation vector for each of the acquired one or more words by the second vector acquisition model 12a (step S123). In the example of FIG. 8, the second vector acquisition unit 12 acquires the distributed representation vector W 11 ,W 12 ,W 13 The second vector acquisition unit 12 calculates a knowledge vector W1 for the brand name based on the acquired one or more distributed representation vectors (step S124).
[0052] FIG. 9 is a diagram showing a process of acquiring a knowledge vector W2 of a top related co-occurrence word and a knowledge vector W3 of a top unrelated co-occurrence word. A co-occurrence word is a word that is frequently used together with a word that means the name of a target concept. A unrelated co-occurrence word is a word that is frequently used together with a word that has the same notation as the name of a target concept but means a different concept. In other words, a unrelated co-occurrence word is a word that is frequently used together with a homograph of the name representing the target concept. The "top related co-occurrence word" refers to a group of M1 related co-occurrence words (M1 is a natural number) from the top when the related co-occurrence words are arranged in descending order of frequency. The "top unrelated co-occurrence word" refers to a group of M2 unrelated co-occurrence words (M2 is a natural number) from the top when the unrelated co-occurrence words are arranged in descending order of frequency.
[0053] In order to prepare the top related co-occurring words and the top unrelated co-occurring words, a group of texts including words that mean the name of the target concept and a group of texts including homographs of the name that represents the target concept are collected. Then, a label indicating whether the text is related to the target concept or not is manually assigned to each of the collected texts. In the example of FIG. 9, the label "1" is assigned to texts that are related to the cosmetic brand "XYZ", the label "0" is assigned to texts that are not related to the cosmetic brand "XYZ", and the label "unknown" is assigned to texts whose association with the cosmetic brand "XYZ" cannot be determined. Hereinafter, texts that are assigned the label "1" are referred to as related texts, and texts that are assigned the label "0" are referred to as unrelated texts.
[0054] In one example, the second vector acquisition unit 12 performs a morphological analysis on each of the related texts, and acquires one or more nouns from each of the related texts as related co-occurrence words. The second vector acquisition unit 12 counts the number of occurrences of each related co-occurrence word as a frequency, and arranges the related co-occurrence words in descending order of frequency to acquire the top related co-occurrence words. The second vector acquisition unit 12 also performs a morphological analysis on each of the unrelated texts, and acquires one or more words from each of the unrelated text groups as unrelated co-occurrence words. The second vector acquisition unit 12 counts the number of occurrences of each unrelated co-occurrence word as a frequency, and arranges the unrelated co-occurrence words in descending order of frequency to acquire the top unrelated co-occurrence words. As another example, the top related co-occurrence words and the top unrelated co-occurrence words may be provided from another computer system and stored in advance in the background knowledge database 22.
[0055] The second vector acquisition unit 12 acquires higher-ranked related co-occurring words (step S122). FIG. 9 shows "cosmetics," "spring," and "purchase" as examples of higher-ranked related co-occurring words. The second vector acquisition unit 12 acquires a distributed representation vector for each of the higher-ranked related co-occurring words by the second vector acquisition model 12a (step S123). In the example of FIG. 9, the second vector acquisition unit 12 acquires the distributed representation vector W 21 ~W 23 The second vector acquisition unit 12 calculates a knowledge vector W2 for the top related co-occurring words based on the one or more acquired distributed representation vectors (step S124). The second vector acquisition unit 12 may acquire distributed representation vectors for synonyms of the top related co-occurring words in addition to the distributed representation vectors of the top related co-occurring words, and calculate the knowledge vector W2 for the top related co-occurring words based on the distributed representation vectors calculated from the top related co-occurring words and their synonyms.
[0056] The second vector acquisition unit 12 acquires the top unrelated co-occurring words (step S122). FIG. 9 shows "personal computer," "repair," and "potatoes." The second vector acquisition unit 12 acquires a distributed representation vector for each of the top unrelated co-occurring words by the second vector acquisition model 12a (step S123). In the example of FIG. 9, the second vector acquisition unit 12 uses the second vector acquisition model 12a to acquire the distributed representation vector W 31 ~W 33 The second vector acquisition unit 12 acquires one or more distributed representation vectors including: a) a knowledge vector W3 for the top unrelated co-occurring words based on the acquired one or more distributed representation vectors; b) a knowledge vector W4 for the top unrelated co-occurring words based on the distributed representation vectors for the top unrelated co-occurring words; c) a knowledge vector W5 for the top unrelated co-occurring words based on the distributed representation vectors for the top unrelated co-occurring words and the distributed representation vectors for the synonyms of the top unrelated co-occurring words; and
[0057] FIG. 10 is a diagram showing a process of acquiring a knowledge vector W4 of a brand keyword. A brand keyword is a word predefined for a target concept, a brand. For example, a brand keyword may be a word that can be associated with a brand. Alternatively, a brand keyword may be a word used to express or modify the image of a product or service handled by the brand. As an example, a brand keyword may be a catchphrase or catch copy used on a company site, a brand site, a product site, or an advertisement.
[0058] The second vector acquisition unit 12 acquires one or more words belonging to the brand keyword from the background knowledge database 22 (step S122). FIG. 10 shows "studio," "powder kiss," and "bare skin" as examples of such words. The second vector acquisition unit 12 acquires a distributed representation vector for each of the acquired one or more words by the second vector acquisition model 12a (step S123). In the example of FIG. 10, the second vector acquisition unit 12 calculates the distributed representation vector W 41 ~W 43 The second vector acquisition unit 12 calculates a knowledge vector W4 for the brand keyword based on the one or more acquired distributed representation vectors (step S124). The second vector acquisition unit 12 may acquire distributed representation vectors for synonyms of the one or more words belonging to the brand keyword in addition to the distributed representation vectors of the one or more words belonging to the brand keyword, and calculate the knowledge vector W4 for the brand keyword based on the distributed representation vectors calculated from the one or more words belonging to the brand keyword and their synonyms.
[0059] FIG. 11 is a diagram showing the process of acquiring the knowledge vector W5 of a brand topic. A brand topic refers to text contained in an article that mentions a target concept, a brand. Articles can be collected in advance from various information sources such as websites and social media. As an example, a brand topic may be the title or text in the body of an article that mentions a brand. FIG. 11 shows examples of such text, such as "Meets the needs of professional makeup artists," "Eye palette with a star motif," and "Sakura collection is back this spring."
[0060] The second vector acquisition unit 12 performs morphological analysis on each brand topic and acquires one or more words from one or more brand topics (step S122). FIG. 11 shows "make-up", "eye palette", and "collection" as examples of such words. The second vector acquisition unit 12 acquires a distributed representation vector for each of the acquired one or more words by the second vector acquisition model 12a (step S123). In the example of FIG. 11, the second vector acquisition unit 12 acquires the distributed representation vector W 51 ~W 53 The second vector acquisition unit 12 calculates a knowledge vector W5 about the brand topic based on the acquired one or more distributed representation vectors (step S124). The second vector acquisition unit 12 may acquire a distributed representation vector for each synonym of the one or more words in addition to the distributed representation vector of the one or more words acquired from the brand topic, and calculate the knowledge vector W5 about the brand topic based on the distributed representation vector calculated from the one or more words acquired from the brand topic and their synonyms.
[0061] 12 is a diagram showing a process (step S13) for calculating the similarity between the target vector and the knowledge vector. The similarity calculation unit 13 calculates the similarity between the target vector W1 and the knowledge vector W5. x The similarity between the two is calculated, and similarities R1 to R5 are obtained.
[0062] 13 is a diagram showing a process (step S14) for calculating the relevance between a target text and a target concept. The relevance acquisition unit 14 inputs the similarities R1 to R5 to the relevance acquisition model 14a. The relevance acquisition model 14a calculates the relevance based on these similarities. The relevance acquisition unit 14 acquires the relevance.
[0063] In calculating the relevance, the relevance acquisition unit 14 may perform weighting for each similarity according to the properties of the knowledge domains K1 to K5 on which the similarities R1 to R5 are based. For example, the brand name, the top related co-occurring words, the brand keywords, and the brand topics are knowledge domains that can be positive indicators of the relevance with the target concept. That is, the similarities R1, R2, R4, and R5 calculated from these knowledge domains are highly likely to have a positive correlation with the relevance between the target text and the target concept. Therefore, the larger the similarities obtained from these knowledge domains, the greater the relevance between the target text and the target concept is considered to be. On the other hand, the top unrelated co-occurring words are knowledge domains that are composed of words unrelated to the target concept, and therefore are knowledge domains that can be negative indicators of the relevance with the target concept. That is, the similarity calculated from the top unrelated co-occurring words is highly likely to have a negative correlation with the relevance between the target text and the target concept. Therefore, the larger the similarity R3 obtained from this knowledge domain K3, the smaller the relevance between the target text and the target concept is considered to be. For this reason, as an example, the relevance obtaining section 14 may apply positive weighting to the similarities R1, R2, R4, and R5, and may apply negative weighting to the similarity R3.
[0064] [Variations] Various examples of the present disclosure have been described in detail above. However, the present disclosure is not limited to the above examples. Various modifications of the present disclosure are possible without departing from the spirit of the present disclosure.
[0065] The information processing system 10 may acquire a knowledge vector of the background knowledge without using a representative value of a plurality of distributed representation vectors corresponding to a plurality of words belonging to the background knowledge. For example, the information processing system 10 may acquire a knowledge vector for at least one knowledge domain without calculating a representative value. For example, the information processing system 10 may integrate m d-dimensional (m and d are both natural numbers) distributed representation vectors obtained in one knowledge domain into one m×d-dimensional vector or d×m matrix, and divide the integrated form into m d-dimensional vectors before calculating the similarity with the target vector. The information processing system 10 may calculate a plurality of knowledge vectors from one knowledge domain. That is, the information processing system 10 may divide a plurality of distributed representation vectors into two or more groups and calculate a knowledge vector for each group, or may acquire all distributed representation vectors acquired from one knowledge domain as knowledge vectors.
[0066] In the above example, the information processing system 10 calculates both the object vector and the knowledge vector, but at least one of these vectors may be calculated by another computer system. In this case, the information processing system may obtain the object vector or the knowledge vector from that computer system. At least one knowledge vector may be stored in advance in a predetermined database such as the background knowledge database 22 in the above example, and in this case, the information processing system may obtain the knowledge vector from that database.
[0067] The information processing method executed by at least one processor is not limited to the above examples. For example, some of the steps or processes described above may be omitted, or each step may be executed in a different order. Also, any two or more of the steps described above may be combined, or some of the steps may be modified or deleted. Alternatively, other steps may be executed in addition to each of the steps described above.
[0068] In this disclosure, the expression "at least one processor executes a first process, executes a second process, ... executes an nth process" or a corresponding expression indicates a concept including a case where the processor that executes the n processes from the first process to the nth process changes midway. In other words, this expression indicates a concept including both a case where all of the n processes are executed by the same processor and a case where the processor changes among the n processes according to an arbitrary policy.
[0069] [Note] As can be seen from the various examples above, the present disclosure includes the following aspects. (Appendix 1) At least one processor; the at least one processor: Obtain the distributed representation vector of the target text as the target vector; A distributed representation vector of background knowledge related to the target concept is obtained as a knowledge vector. Calculating a similarity between the target vector and the knowledge vector; inputting the calculated similarity into a machine learning model that calculates a degree of relevance between the target text and the target concept based on the similarity, and acquiring a degree of relevance between the target text and the target concept from the machine learning model; Information processing system. (Appendix 2) the at least one processor calculates, as the knowledge vector, a representative value of a plurality of distributed representation vectors corresponding to a plurality of words belonging to the background knowledge; 2. An information processing system according to claim 1. (Appendix 3) The information constituting the background knowledge is grouped into a plurality of knowledge domains; the at least one processor: obtaining the knowledge vector for each of the plurality of knowledge domains; Calculating the similarity for each of the plurality of knowledge domains; inputting the calculated similarities into the machine learning model and obtaining the relevance from the machine learning model; 3. An information processing system according to claim 1 or 2. (Appendix 4) The background knowledge includes related co-occurrence words whose frequency of use together with the name of the target concept meets a predetermined criterion. An information processing system according to any one of claims 1 to 3. (Appendix 5) The background knowledge includes predefined keywords for the target concept. An information processing system according to any one of claims 1 to 4. (Appendix 6) The background knowledge includes words contained in articles that mention the target concept. 6. An information processing system according to any one of claims 1 to 5. (Appendix 7) the background knowledge includes non-related co-occurring words whose frequency of use together with homographs of the name of the target concept meets a predetermined criterion; An information processing system according to any one of appendixes 1 to 6. (Appendix 8) 1. An information processing method executed by an information processing system having at least one processor, comprising: A step of obtaining a distributed representation vector of a target text as a target vector; A step of acquiring a distributed representation vector of background knowledge related to a target concept as a knowledge vector; calculating a similarity between the target vector and the knowledge vector; inputting the calculated similarity into a machine learning model that calculates a degree of relevance between the target text and the target concept based on the similarity, and acquiring a degree of relevance between the target text and the target concept from the machine learning model; An information processing method comprising: (Appendix 9) A step of obtaining a distributed representation vector of a target text as a target vector; A step of acquiring a distributed representation vector of background knowledge related to a target concept as a knowledge vector; calculating a similarity between the target vector and the knowledge vector; inputting the calculated similarity into a machine learning model that calculates a degree of relevance between the target text and the target concept based on the similarity, and acquiring a degree of relevance between the target text and the target concept from the machine learning model; An information processing program that causes a computer to execute the above.
[0070] According to Supplementary Notes 1, 8, and 9, the similarity between the target text and background knowledge related to the target concept is first calculated, and then the machine learning model obtains the relevance between the target text and the target concept from the similarity. By using the similarity with the background knowledge related to the target concept and the machine learning model in this way, it is possible to accurately estimate the degree to which the target text is related to the target concept.
[0071] According to Supplementary Note 2, by using a representative value as the knowledge vector, a knowledge vector that more appropriately represents the characteristics of the background knowledge can be obtained, so that the similarity can be calculated more accurately. As a result, the relevance can be estimated more accurately.
[0072] According to Supplementary Note 3, since the relevance is obtained based on a plurality of similarities calculated from a plurality of viewpoints of background knowledge, the relevance can be estimated with higher accuracy.
[0073] According to Supplementary Note 4, the relevance is obtained based on the similarity between the target text and related co-occurring words whose frequency of use with the name of the target concept meets a predetermined standard. The similarity is highly likely to have a positive correlation with the relevance, so by using this, the relevance can be estimated more accurately.
[0074] According to Supplementary Note 5, the relevance is obtained based on the similarity between the target text and keywords predefined for the target concept. The similarity is highly likely to have a positive correlation with the relevance, so by using this, the relevance can be estimated more accurately.
[0075] According to Appendix 6, the relevance is obtained based on the similarity between the target text and the words contained in the article that mentioned the target concept. The similarity is highly likely to have a positive correlation with the relevance, so by using this, the relevance can be estimated more accurately.
[0076] According to Appendix 7, the relevance is obtained based on the similarity between the target text and unrelated co-occurring words that meet a certain criterion for frequency of use together with homographs of the name of the target concept. The similarity is highly likely to have a negative correlation with the relevance, so the relevance can be estimated more accurately by using it. [Explanation of symbols]
[0077] 10...information processing system, 11...first vector acquisition unit, 11a...first vector acquisition model, 12...second vector acquisition unit, 12a...second vector acquisition model, 13...similarity calculation unit, 14...association acquisition unit, 14a...association acquisition model.
Claims
1. At least one processor; the at least one processor: Obtain the distributed representation vector of the target text as the target vector; A distributed representation vector of background knowledge related to the target concept is obtained as a knowledge vector. Calculating a similarity between the target vector and the knowledge vector; inputting the calculated similarity into a machine learning model that calculates a degree of relevance between the target text and the target concept based on the similarity, and acquiring a degree of relevance between the target text and the target concept from the machine learning model; Information processing system.
2. the at least one processor calculates, as the knowledge vector, a representative value of a plurality of distributed representation vectors corresponding to a plurality of words belonging to the background knowledge; The information processing system according to claim 1 .
3. The information constituting the background knowledge is grouped into a plurality of knowledge domains; the at least one processor: obtaining the knowledge vector for each of the plurality of knowledge domains; Calculating the similarity for each of the plurality of knowledge domains; inputting the calculated similarities into the machine learning model and obtaining the relevance from the machine learning model; 3. The information processing system according to claim 1 or 2.
4. The background knowledge includes related co-occurrence words whose frequency of use together with the name of the target concept meets a predetermined criterion.
3. The information processing system according to claim 1 or 2.
5. The background knowledge includes predefined keywords for the target concept.
3. The information processing system according to claim 1 or 2.
6. The background knowledge includes words contained in articles that mention the target concept.
3. The information processing system according to claim 1 or 2.
7. the background knowledge includes non-related co-occurring words whose frequency of use together with homographs of the name of the target concept meets a predetermined criterion; 3. The information processing system according to claim 1 or 2.
8. 1. An information processing method executed by an information processing system having at least one processor, comprising: A step of obtaining a distributed representation vector of a target text as a target vector; A step of acquiring a distributed representation vector of background knowledge related to a target concept as a knowledge vector; calculating a similarity between the target vector and the knowledge vector; inputting the calculated similarity into a machine learning model that calculates a degree of relevance between the target text and the target concept based on the similarity, and acquiring a degree of relevance between the target text and the target concept from the machine learning model; An information processing method comprising:
9. A step of obtaining a distributed representation vector of a target text as a target vector; A step of acquiring a distributed representation vector of background knowledge related to a target concept as a knowledge vector; calculating a similarity between the target vector and the knowledge vector; inputting the calculated similarity into a machine learning model that calculates a degree of relevance between the target text and the target concept based on the similarity, and acquiring a degree of relevance between the target text and the target concept from the machine learning model; An information processing program that causes a computer to execute the above.
Citation Information
Patent Citations
Category estimation device and category estimation method
JP2020161075A
