A knowledge graph retrieval method integrating entity category features and target context

Abstract

The present invention relates to the field of natural language processing, and provides a knowledge graph retrieval method integrating entity category features and target context, comprising: extracting entity categories in an existing knowledge base and entity sets corresponding to each category, preprocessing a text to be detected, and obtaining a phrase consisting of a to-be-matched word and its context; calculating the similarity between each entity category word and the to-be-matched word and its context, and selecting p entity category words with the highest similarity scores; calculating the similarity between the entity words corresponding to the p entity category words with the highest similarity scores and the corresponding entity feature words and the context; combining the two similarities to comprehensively calculate the similarity score of each entity word to the context of the to-be-detected word, and returning the entity with the highest score as the correct entity. The present invention adopts a method of first matching entity categories, first excluding most entities with low similarity, and then matching entities, optimizing matching time, and making calculation faster.

Description

Technical Field

[0001] The present invention relates to the field of natural language processing, and more specifically, to a knowledge graph retrieval method that integrates entity category features and target context. Background Art

[0002] Knowledge graph is a technology that uses graph structure to represent entities and the relationships between entities. It uses big data and artificial intelligence technology to express Internet information in a mode that is closer to the human cognitive world. The knowledge graph question-answering system combines natural language processing technology and knowledge graphs to understand and analyze user questions and extract relevant information from the knowledge graph to generate answers. The system has powerful knowledge representation and logical reasoning capabilities, and can provide users with intelligent and personalized question-answering services. It is widely used in enterprise knowledge management, intelligent customer service and other fields. Knowledge graph retrieval is based on structured data in the knowledge graph, and provides users with more accurate and comprehensive search results through technologies such as entity resolution and relationship reasoning. It solves the limitation of traditional retrieval methods that cannot understand semantics and relationships, and improves the efficiency and accuracy of information retrieval.

[0003] Knowledge graph retrieval algorithms are mainly divided into three categories: graph-based retrieval algorithms, vector space model-based retrieval algorithms, and deep learning-based retrieval algorithms. Different knowledge graph retrieval algorithms are suitable for different application scenarios and data sets. In practical applications, it is necessary to select appropriate algorithms or algorithm combinations based on specific needs and data characteristics. At present, the retrieval question-answering system based on knowledge graph mainly uses retrieval algorithms based on vector space models and retrieval algorithms based on deep learning. When the amount of knowledge in the graph is insufficient, the deep learning retrieval algorithm does not have enough samples to train the model, so the vector space model retrieval algorithm will be selected. Generally, the entities and relationships in the knowledge graph are mapped to a low-dimensional vector space, and the correlation between entities is evaluated by calculating the similarity between vectors.

[0004] The current retrieval algorithm of the vector space model has problems when the vector dimension is too high. The large number of vectors affects the calculation speed, and simply matching the target entity with the knowledge graph library will lose a lot of information in the original text, affecting the model's understanding of the entire article. At the same time, there may be entities with the same name, which affects the accuracy of the retrieval match, resulting in the subsequent question-and-answer system being unable to give accurate answers. Summary of the invention

[0005] In view of the technical problems existing in the prior art, the present invention provides a knowledge graph retrieval method that integrates entity category features and target context, aiming to solve the above-mentioned problems.

[0006] The present invention provides a knowledge graph retrieval method integrating entity category features and target context, comprising:

[0007] Step 1: extract entity categories in the existing knowledge graph library and entity sets corresponding to each entity category, wherein the entity set corresponding to each entity category includes multiple entity words, and each entity word includes multiple entity feature words;

[0008] Step 2: preprocess the text to be detected to obtain a phrase consisting of the word to be matched and its context;

[0009] Step 3, calculating the similarity between each entity category word and the word to be matched and its context, and selecting p entity category words with the highest similarity to the word to be matched and its context;

[0010] Step 4, obtaining entity sets corresponding to p entity category words, obtaining all entity words in the p entity sets as candidate entity words, and obtaining entity feature words included in each candidate entity word;

[0011] Step 5, based on the similarity between the context of each candidate entity word and the word to be matched and the similarity between the entity feature word of each candidate entity word and the context of the word to be matched, comprehensively calculate the similarity between the context of each candidate entity word and the word to be matched;

[0012] Step 6: The candidate entity word with the highest similarity to the context of the word to be matched is used as the entity word to be matched with the word to be matched.

[0013] The present invention provides a knowledge graph retrieval method that integrates entity category features and target context, which has the following beneficial effects:

[0014] (1) The existing knowledge graph retrieval and matching mechanism uses a direct matching method to match all graph entities. If the number of entities is large, there is a problem of low matching efficiency. The present invention adopts a method of matching tags first, first excluding most entities with low similarity, optimizing the matching time and making the calculation faster.

[0015] (2) The existing knowledge graph retrieval and matching mechanism directly uses similarity calculation to match the target entity with the entities in the graph, and selects the entity with the highest similarity as the correct entity. This may result in successful matching of entities with the same name but different meanings, which is easy to cause ambiguity and ignores the context information of the target entity and the feature information of the graph entity, resulting in a low matching accuracy rate. The similarity matching of the present invention also matches the context information of the target entity and the feature information of the graph entity, taking into account the full-text information and entity features, reducing the probability of successful matching of entities with the same name but different meanings, further improving the matching accuracy rate, and reducing the occurrence of ambiguity.

[0016] (3) The existing knowledge graph retrieval and matching mechanism directly uses the similarity value as the matching basis, but the effect is not particularly good in the retrieval and matching of some complex knowledge graphs; the present invention adopts a special similarity scoring algorithm as the matching basis, and takes into account some words with high frequency of occurrence but low recognition in the calculation of the similarity score, reduces the scores of these words, and enhances the accuracy of the algorithm. The parameters in the algorithm can also be adjusted according to the current data to enhance the robustness of the algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 A flowchart of a knowledge graph retrieval method that integrates entity category features and target context provided by the present invention;

[0018] Figure 2 A schematic diagram of the algorithm structure involved in a knowledge graph retrieval method that integrates entity category features and target context. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. In addition, the technical features in the various embodiments provided by the present invention can be arbitrarily combined with each other to form a feasible technical solution. This combination is not restricted by the sequence of steps and / or the structural composition mode, but must be based on the ability of ordinary technicians in this field to achieve it. When the combination of technical solutions is contradictory or cannot be achieved, it should be considered that this combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0020] Figure 1 A flow chart of a knowledge graph retrieval method that integrates entity category features and target context provided by the present invention, such as Figure 1 and Figure 2 As shown in Figure 1, the knowledge graph retrieval method includes:

[0021] Step 1: extract entity categories in the existing knowledge graph library and the entity sets corresponding to each entity category, wherein the entity set corresponding to each entity category includes multiple entity words, and each entity word includes multiple entity feature words.

[0022] Among them, suppose that in a knowledge graph library, the entity label set contains Entity category labels:

[0023] ;

[0024] Entity category label The corresponding entity set is:

[0025] ;

[0026] The set of all entities in the knowledge graph is:

[0027] ;

[0028] Each entity Included The entity feature words are:

[0029] ;

[0030] in, Represents the number of entity sets, Represents entity category label The number of entities in the entity set, Indicates that the entity category label is No. entities, Represents the rth entity in the set The qth entity feature word of an entity, Represents the number of entities in the r-th entity set.

[0031] Step 2: preprocess the text to be detected to obtain a phrase consisting of the word to be matched and its context.

[0032] It is understandable that the text to be detected is preprocessed, including operations such as word segmentation and removal of stop words, to obtain a phrase consisting of the word to be matched and its context.

[0033] Now there is a piece of text data to be detected. According to the application requirements, it is preprocessed to obtain text :

[0034] ;

[0035] Contains entities to be matched , and the target entity context :

[0036] .

[0037] Assuming entity category labels ,category The entity in ,category The entity in ,category The entity in The entities to be matched and their context are ,in is the entity to be matched, and the context is And the calculation parameter K=3, .

[0038] Subsequently, the entity words, entity feature words, entity category words extracted from the existing knowledge graph and the words to be matched and the context extracted from the text to be detected are embedded to obtain the entity word embedding vector, entity feature word embedding vector, entity category embedding vector, word vector to be matched and context word vector respectively.

[0039] In this invention, the dimension of the word embedding is set to three dimensions. The specific word embedding method is adjusted according to the actual situation. It is designed here only to facilitate the understanding of the calculation process. The following is the vector of the entity category, entity, entity to be matched, and context word embedding:

[0040] Entity category word vector: ; ; .

[0041] Entity word vector: ; ; ; ; .

[0042] The entities to be matched and the context are:

[0043] ; ; ; ; ; ; .

[0044] Step 3: Calculate the similarity between each entity category word and the word to be matched and its context, and select p entity category words with the highest similarity to the word to be matched and its context.

[0045] It is understandable that the similarity between each entity category word and the word to be matched and its context is calculated first, that is, the average similarity between each entity category word and each entity word in the text to be detected is calculated. The specific idea is to first calculate the similarity between the entity category word and each entity word in the text to be detected, sort the similarities, select the top K largest similarities and average them, and obtain the average similarity between the entity category word and all entity words in the text to be detected.

[0046] The process of calculating the average similarity between the entity category word and all entity words in the text to be detected is:

[0047] Calculate the average similarity between each entity category label and the words to be matched in the text to be detected and their context to form an average similarity set:

[0048] ;

[0049] , ;

[0050] in, Represents all entity category labels relative to the text to be detected The average similarity set of the words to be matched and their contexts in , Represents entity category label The average similarity with the text V to be detected, Entity category label Text to be detected The similarity set of each entity word in , Retains the array The first K largest values.

[0051] in:

[0052] ;

[0053] , ;

[0054] ;

[0055] in, Represents entity category label For the text to be detected Entity words in The similarity of is the cosine similarity, Represents the total number of entities in the text to be detected, function Represents entity category label In the text to be detected The number of times it appears in Represents the entity in the text V to be detected The second norm of Represents entity category label The second norm of . Among them, It means that while ensuring that its value is greater than or equal to 1, the weights of some frequently appearing but unrecognizable words are reduced to enhance the recognition ability.

[0056] The above method is used to calculate the average similarity between each entity category word and the words to be matched in the text to be detected and their contexts, and the p entity category words with the highest similarity are selected.

[0057] Step 4: obtain entity sets corresponding to the p entity category words, obtain all entity words in the p entity sets as candidate entity words, and obtain entity feature words included in each candidate entity word.

[0058] It is understandable that, for the selected p entity category words, the entity word corresponding to each entity category word is obtained as a candidate entity word, and the entity feature words contained in each candidate entity word are obtained.

[0059] Specifically, choose The P entity category labels with the highest average similarity are:

[0060] ;

[0061] The entity set corresponding to p entity category labels is , where all entities contained in these P entity category labels are:

[0062] ;

[0063] in Indicates the p entity category labels with the highest similarity The entity set corresponding to the entity category label The entity.

[0064] According to step 3, calculate With entity class , , The similarity is rounded to four decimal places:

[0065] ;

[0066] ;

[0067] ;

[0068] ;

[0069] ;

[0070] Select the P entity category labels with the highest similarity, where P=2, and select the entity category labels as and Entity.

[0071] Entity category label and The entity words and corresponding entity feature words are:

[0072] ;

[0073] ;

[0074] ;

[0075] ;

[0076] ;

[0077] Step 5, based on the similarity between each candidate entity word and the context of the word to be matched and the similarity between the entity feature word of each candidate entity word and the context of the word to be matched, comprehensively calculate the similarity between each candidate entity word and the context of the word to be matched.

[0078] It is understandable that step 4 obtains p entity category words and all candidate entity words contained therein and entity feature words contained in each candidate entity word. The average similarity between the context of each candidate entity word and the word to be matched, and the average similarity between the context of each entity feature word contained in each candidate entity word and the word to be matched are calculated respectively. Then, based on the average similarity between the context of each candidate entity word and the word to be matched and the average similarity between the context of each entity feature word contained in the candidate entity word and the word to be matched, the similarity between the context of each candidate entity word and the word to be matched is comprehensively calculated.

[0079] The average similarity between each candidate entity word and the context of the word to be matched is calculated, including:

[0080] ;

[0081] ;

[0082] ;

[0083] in, For Entity For the set of context similarity, For Entity The average similarity score for the context, For Entity Words in context The similarity.

[0084] Calculate the average similarity of multiple entity feature words of each entity candidate word to the context of the word to be matched, including:

[0085] ;

[0086] ;

[0087] ;

[0088] in, For Entity The set of average similarities of all entity feature words to the context, For Entity Entity feature words The set of similarities to the context, Representing Entities Entity feature words Words in context The similarity.

[0089] According to step S5, the similarity between each candidate entity word and its entity feature word and the context of the word to be matched is calculated, and the result is rounded to four decimal places.

[0090] Calculate the average similarity score of each candidate entity word to the context:

[0091] ;

[0092] ;

[0093] ;

[0094] ;

[0095] ;

[0096] ;

[0097] ;

[0098] ;

[0099] ;

[0100] ;

[0101] Calculate the cosine similarity between entity feature words and context:

[0102] ;

[0103] ;

[0104] ;

[0105] ;

[0106] ;

[0107] ;

[0108] ;

[0109] ;

[0110] ;

[0111] ;

[0112] ;

[0113] ;

[0114] Calculate the average similarity score set of entity feature words to context:

[0115] ;

[0116] ;

[0117] ;

[0118] ;

[0119] ;

[0120] Based on the average similarity between the context of each candidate entity word and the word to be matched and the average similarity between the context of each entity feature word contained in the candidate entity word and the word to be matched, the similarity between the context of each candidate entity word and the word to be matched is comprehensively calculated, including:

[0121] ;

[0122] ;

[0123] in, Treat the context of the matching word for each candidate entity word and its entity feature word The comprehensive average similarity set of Entity and its entity feature words The average similarity of the contexts.

[0124] According to step 5, the comprehensive average similarity score set of each candidate entity word and entity feature word to the context is finally calculated:

[0125] ;

[0126] ;

[0127] ;

[0128] ;

[0129] ;

[0130] ;

[0131] ;

[0132] ;

[0133] Step 6: The candidate entity word with the highest similarity to the context of the word to be matched is used as the entity word to be matched with the word to be matched.

[0134] It is understandable that, according to the similarity between each candidate entity word and the context of the word to be matched, the candidate entity word with the highest similarity is returned as the correct entity word, that is, the entity word that matches the word to be matched.

[0135] According to step 5, since ,so Select Entity The entities that are matched are returned.

[0136] The present invention provides a knowledge graph retrieval method that integrates entity category features and target context, in order to solve the problems of low matching efficiency, low matching accuracy and easy ambiguity in the existing knowledge graph retrieval matching mechanism. The present invention optimizes the matching time and improves the computing efficiency by matching the labels first. At the same time, in the similarity matching, the present invention not only considers the similarity between the target entity and the graph entity, but also incorporates the context information of the target entity and the feature information of the graph entity, reduces the number of successful matches of entities with the same name but different meanings, further improves the matching accuracy and reduces ambiguity. In addition, the present invention adopts a special similarity scoring algorithm to reduce the scores of high-frequency and low-recognition words, enhances the accuracy of the algorithm, and allows the parameters to be adjusted according to the current data to enhance the robustness of the algorithm.

[0137] It should be noted that in the above embodiments, the description of each embodiment has its own emphasis, and for parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0138] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0139] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded computer, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0140] These computer program instructions may also be stored in a computer readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0141] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0142] Although the preferred embodiments of the present invention have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0143] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A knowledge graph retrieval method integrating entity category features and target context, characterized in that: include: Step 1: extract entity categories in the existing knowledge graph library and entity sets corresponding to each entity category, wherein the entity set corresponding to each entity category includes multiple entity words, and each entity word includes multiple entity feature words; Step 2: preprocess the text to be detected to obtain a phrase consisting of the word to be matched and its context; Step 3, calculating the similarity between each entity category word and the word to be matched and its context, and selecting p entity category words with the highest similarity to the word to be matched and its context; Step 4, obtaining entity sets corresponding to p entity category words, obtaining all entity words in the p entity sets as candidate entity words, and obtaining entity feature words included in each candidate entity word; Step 5, based on the similarity between the context of each candidate entity word and the word to be matched and the similarity between the entity feature word of each candidate entity word and the context of the word to be matched, comprehensively calculate the similarity between the context of each candidate entity word and the word to be matched; Step 6, taking the candidate entity word with the highest similarity to the context of the word to be matched as the entity word to be matched with the word to be matched; The step 1, extracting entity categories in the existing knowledge graph library and entity sets corresponding to each entity category, includes: In the existing knowledge graph library, the entity label set contains h entity category labels: <h2 style=";text-align:left;direction:ltr">Label = {L1,L2,L3,…,L<h2 style=";text-align:left;direction:ltr"> h <h2 style=";text-align:left;direction:ltr">}; Entity category label L i , 1≤i≤h, the corresponding entity set is: The set of all entities in the existing knowledge graph library is: E={E1,E2,…,E i ,…,THE h } Entity e′ rl The q entity feature words included are: F(e rl ′)={f rl1 ,f rl2 ,…,f rlq },(1≤r≤h,1≤l≤n r ) Among them, h represents the number of entity sets, n i Represents the entity category label L i The number of entities in the entity collection, Indicates that the entity category label is L i The nth i Entity, f rlq F(e) represents the qth entity feature word contained in the lth entity in the rth entity set. rl ) represents entity e′ rl A set of q entity feature words.

2. The knowledge graph retrieval method according to claim 1, characterized in that: The step 2 is to pre-process the text to be detected to obtain a phrase consisting of the word to be matched and its context, including: Perform word segmentation and stop word removal operations on the text to be detected to obtain a phrase consisting of the word to be matched and its context; The entity category words, entity words and entity feature words extracted from the existing knowledge graph library, as well as the words to be matched and the context extracted from the text to be detected are embedded to obtain the entity category word embedding vector, entity word embedding vector, entity feature word embedding vector, word vector to be matched and context word vector respectively; The word segmentation and stop word removal operations are performed on the text to be detected to obtain a phrase consisting of the word to be matched and its context, including: After preprocessing the text to be detected, the text V is obtained: V={V m-a ,V m-a+1 ,…,V m-1 ,V m ,V m+1 ,…,V m+b-1 ,V m+b }; The text V contains the entity V to be matched m , and the context entity V' of the entity to be matched: V'={V m-a ,V m-a+1 ,…,V m-1 ,V m+1 ,…,V m+b-1 ,V m+b }。 3. The knowledge graph retrieval method according to claim 2, characterized in that: The step 3, calculating the similarity between each entity category word and the word to be matched and its context, includes: Calculate the average similarity between each entity category label and the matching word in the text to be detected and its context to form an average similarity set: S L ={S[V,L1],S[V,L2],…,S[V,L h ]}; Among them, S L represents the average similarity set of all entity category labels relative to the words to be matched in the text to be detected V and its context, S[V,L j ] represents the entity category label L j The average similarity with the text to be detected V, U j is the entity category label L j The similarity set of each entity word in the detected text V, z(U j ,K) means to keep the array U j The first K largest values; in: IN j ={in j1 ,in j2 ,…,in j(a+b) ,in j(a+b+1) } Among them, u jt Represents the entity category label L j For the entity word V in the text to be detected V t The similarity of Sim(V t ,L j ) is the cosine similarity, |V| represents the total number of entities in the text to be detected, and the function fl(L j ,V) represents the entity category label L j The number of times it appears in the text to be detected V, ||V t ||2 represents the entity V in the text V to be detected t The second norm of ||L j ||2 represents the entity category label L j The second norm of .

4. The knowledge graph retrieval method according to claim 3, characterized in that: The step 4, obtaining entity sets corresponding to p entity category words, obtaining all entity words in the p entity sets as candidate entity words, and obtaining entity feature words included in each candidate entity word, includes: Select S L The P entity category labels with the highest average similarity are: z(S L ,P)=Lable'={L1',…,L p '} The entity set corresponding to the p entity category labels is {E1′,…,E p ′}, where all entities contained in P entity category labels are: where e′ rl Represents the entity set E′ corresponding to the rth entity category label among the p entity category labels with the highest similarity r The lth entity in .

5. The knowledge graph retrieval method according to claim 4, characterized in that: In step 5, the average similarity between each candidate entity word and the context of the word to be matched is calculated, including: Among them, B rl For entity e′ rl For the set of context similarity, S[V',e' rl ] is the entity e' rl The average similarity score for the context, For entity e' rl For the word V in the context v The similarity.

6. The knowledge graph retrieval method according to claim 5, characterized in that: In step 5, the average similarity of multiple entity feature words of each entity candidate word to the context of the word to be matched is calculated, including: Among them, W rl For entity e rl The set of average similarities of all entity feature words of ′ to the context, For entity e rl 'The entity feature word f rld The set of similarities to the context, Represents entity e rl 'The entity feature word f rld For the word V in the context c The similarity.

7. The knowledge graph retrieval method according to claim 6, characterized in that: The step 5 comprehensively calculates the similarity between each candidate entity word and the context of the word to be matched, including: Among them, S e For each candidate entity word and its entity feature word, the comprehensive average similarity set of the context V' of the matching word is S[V',fe' rl ] is the entity word e′ rl and its entity feature word F(e rl ′) is the average similarity to the context.

Images