Knowledge base enhancement method based on statement entity link

Abstract

The invention relates to a knowledge base enhancement method based on statement entity links, which adopts a brand-new logic design, introduces and combines a first knowledge base and a second knowledge base corresponding to each preset judgment rule, associates contexts, performs consistency comparison on each entity link of a to-be-analyzed statement with the first knowledge base, and enhances the statement entity links according to a comparison result. According to the method, entity linking of a to-be-analyzed statement is realized, or data supplementation is performed for corresponding entity words and corresponding description information in the first knowledge base, enhancement of the first knowledge base is realized, and through continuous data enhancement and accumulation of the first knowledge base under continuous application, the entity linking accuracy of the first knowledge base is continuously improved; and thus, the accuracy of statement analysis and question and answer interaction in practical application is improved.

Description

Technical Field

[0001] The invention relates to a knowledge base enhancement method based on sentence entity linking, and belongs to the technical field of sentence entity linking analysis. Background Art

[0002] Entity linking (EL), also known as named entity linking (NEL), refers to identifying entities with specific meanings in text and assigning them unique IDs in the concept dictionary. For example, a dialogue system receives a natural language question, understands the intent of the natural language question, and outputs an answer that matches the natural language question. Identifying entities in natural language questions and linking them to specific entities in the knowledge base is the key to obtaining appropriate answers. This step is called entity linking. In related technologies, the research method of entity linking can usually be regarded as a similarity match between input entities and knowledge base entities, and the most appropriate link entity is obtained by comparing the input entity itself with the entities in the knowledge base.

[0003] In a vertical field dialogue system such as medical care, there is relatively little data of all kinds, so the performance bottleneck of entity linking does not lie in how to choose a more sophisticated word embedding method or similarity comparison function, but in the lack of necessary input entities and knowledge base entity context data. The lack of necessary data leads to low accuracy of entity linking. In addition, due to the large number of entities in vertical fields and their distribution tends to be long-tailed, centralized data supplementation is very labor-intensive and resource-intensive, and it is also difficult to accumulate the appropriate amount of data for all entities.

[0004] In addition, the medical scenario mainly includes diseases, symptoms, examinations, indicators, proper nouns, etc. The entity links in the existing technical dialogue system are all single sentences, and do not use the context dialogue and other historical dialogue information. Especially in the question-and-answer scenario, the question-and-answer users usually use simple sentences to express their intentions and needs. Therefore, when performing entity linking, it is difficult to identify the unique ID of some proper words from the current sentence alone. For example, there is an abbreviation called RA, which can mean "rheumatoid arthritis" or "right atrium". So when a dialogue contains "RA", is the "RA" here the unique ID given to rheumatoid arthritis? Or is it the unique ID of the right atrium? In this case, it is impossible to find a unique concept ID that matches the context through the word itself. And from the short sentences in the dialogue, it is also very difficult to select the correct unique ID from a set of hundreds of thousands of concepts, especially in vertical fields such as the medical industry, which lacks the accumulation of infrastructure. At present, no effective technical solutions have been seen to solve these problems. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a knowledge base enhancement method based on sentence entity linking, which adopts a new design, connects the context, and introduces knowledge base supplementation and enhancement, which can effectively improve the accuracy of sentence content entity linking in a single-round dialogue.

[0006] In order to solve the above technical problems, the present invention adopts the following technical solution: the present invention designs a knowledge base enhancement method based on sentence entity linking, for a first knowledge base including a preset number of entity words, and each entity word corresponds to preset different types of description information, according to the following steps A to D, by semantically linking the sentences to be parsed, the first knowledge base is enhanced; Step A. extract each entity word in the sentence to be parsed as each entity word to be analyzed, form an entity set, and then proceed to step B; Step B. Based on the first knowledge base containing a preset number of entity words and preset different types of description information corresponding to each entity word, the word embedding model is applied to convert the word into a representation vector, and combined with similarity calculation, the corresponding entity words in the first knowledge base linked to each entity word to be analyzed in the entity set and the corresponding description information are obtained to realize the semantic link of the sentence to be parsed, and then enter step C; Step C. Based on each entity word in the first knowledge base, each different type of description information corresponding to each entity word, and the preset judgment rules corresponding to each different type of description information of each entity word, a second knowledge base is constructed, and for each entity word to be analyzed in the entity set, according to the judgment rules, the context information of the preset character length corresponding to the entity word to be analyzed is linked, and the entity word to be analyzed is analyzed and judged, and the corresponding entity words in the second knowledge base linked to the entity word to be analyzed and the corresponding description information are obtained, and then the corresponding entity words in the second knowledge base linked to each entity word to be analyzed in the entity set and the corresponding description information are obtained, so as to realize the semantic link of the sentence to be parsed, and then enter step D; Step D. For each entity word to be analyzed in the entity set, determine whether the corresponding entity words and corresponding description information in the first knowledge base linked to the entity word to be analyzed are consistent with the corresponding entity words and corresponding description information in the second knowledge base linked to the entity word to be analyzed. If so, determine the entity words and corresponding description information linked to the entity word to be analyzed. Otherwise, supplement the data for the corresponding entity words and corresponding description information in the first knowledge base to enhance the first knowledge base.

[0007] As a preferred technical solution of the present invention: in the step A, a sequence labeling model is used to perform sequence labeling on the character sequence corresponding to the sentence to be parsed, and the entity words corresponding to the sentence to be parsed are obtained by locating and classifying the elements in the sequence, as the entity words to be analyzed, to form an entity set.

[0008] As a preferred technical solution of the present invention: the step B includes executing the following steps B1 to B3; Step B1. For each entity word to be analyzed in the entity set, respectively, for the entity word to be analyzed, in conjunction with the context information of the preset character length corresponding to the entity word to be analyzed, the word embedding model is applied to obtain the representation vector corresponding to the entity word to be analyzed, and then the representation vectors corresponding to each entity word to be analyzed in the entity set are obtained, and then step B2 is entered; Step B2. Based on a first knowledge base including a preset number of entity words and preset different types of description information corresponding to each entity word, a word embedding model is applied to obtain a representation vector corresponding to each combination for each entity word and its corresponding description information, and then the process proceeds to step B3; Step B3. For each entity word to be analyzed in the entity set, calculate the similarity between the representation vector corresponding to the entity word to be analyzed and the representation vector corresponding to each combination in the first knowledge base according to the vector similarity calculation method, and select each combination that is greater than the preset initial similarity threshold as the corresponding entity word in the first knowledge base linked to the entity word to be analyzed, and the corresponding description information.

[0009] As a preferred technical solution of the present invention: the word embedding model is any one of the Fasttext model, the Glove model, the ELMo model, and the BERT model.

[0010] As a preferred technical solution of the present invention: in the step D, if it is determined that the corresponding entity words and corresponding description information in the first knowledge base linked to the entity word to be analyzed are inconsistent with the corresponding entity words and corresponding description information in the second knowledge base linked to the entity word to be analyzed, then the following steps D1 to D3 are executed; Step D1. Determine whether the corresponding entity words and corresponding description information in the first knowledge base linked to the entity word to be analyzed and the corresponding entity words and corresponding description information in the second knowledge base linked to the entity word to be analyzed are consistent with each other, if yes, proceed to step D2; otherwise, proceed to step D3; Step D2. Based on the entity words and corresponding description information that are consistent with each other, as the objects to be supplemented corresponding to the entity words to be analyzed, the description information of the objects to be supplemented corresponding to the entity words to be analyzed in the first knowledge base is supplemented respectively, so as to enhance the first knowledge base; Step D3. The corresponding entity words and corresponding description information in the second knowledge base linked to the entity word to be analyzed are used as the objects to be supplemented corresponding to the entity word to be analyzed, and the description information of the objects to be supplemented corresponding to the entity word to be analyzed in the first knowledge base is supplemented respectively to enhance the first knowledge base.

[0011] As a preferred technical solution of the present invention: in the steps D2 and D3, the context information of the preset character length corresponding to the entity word to be analyzed is respectively supplemented to the description information of each object to be supplemented corresponding to the entity word to be analyzed in the first knowledge base, so as to enhance the first knowledge base.

[0012] As a preferred technical solution of the present invention: in the steps D2 and D3, based on the entire network, the description information of each object to be supplemented corresponding to the entity word to be analyzed is searched respectively, and the description information of each object to be supplemented corresponding to the entity word to be analyzed in the first knowledge base is supplemented respectively, so as to enhance the first knowledge base.

[0013] As a preferred technical solution of the present invention: while enhancing the first knowledge base according to steps A to D, based on step B, the corresponding entity words and corresponding description information in the first knowledge base to which each entity word to be analyzed in the entity set corresponding to the sentence to be parsed is linked are obtained, and the entity words and corresponding description information with the highest similarity in the first knowledge base to which each entity word to be analyzed is linked are selected to form the entity words and corresponding description information to which each entity word to be analyzed in the sentence to be parsed is linked, so as to realize the semantic link of the sentence to be parsed.

[0014] As a preferred technical solution of the present invention: the sentences to be parsed are derived from the conversation and communication process between the doctor and the patient during consultation. By capturing the voice of the conversation and communication process, various sentences are obtained and used as the sentences to be parsed respectively. Steps A to D are executed to perform semantic links on the sentences to be parsed. The entity words in the first knowledge base and the second knowledge base are all preset entity words in the medical field.

[0015] Corresponding to the above, the technical problem that the present invention also needs to solve is to provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the steps of the knowledge base enhancement method based on sentence entity linking are implemented.

[0016] The knowledge base enhancement method based on sentence entity linking described in the present invention adopts the above technical solution and has the following technical effects compared with the prior art: The present invention designs a knowledge base enhancement method based on sentence entity linking, adopts a new logical design, introduces a second knowledge base that combines a first knowledge base with corresponding preset judgment rules, connects the context, and performs consistency comparison on the entity links of the sentence to be parsed with respect to the first knowledge base. Based on the comparison result, the entity linking of the sentence to be parsed is realized, or data is supplemented for the corresponding entity words and corresponding description information in the first knowledge base to realize the enhancement of the first knowledge base. Through continuous data enhancement and accumulation of the first knowledge base under continuous application, the accuracy of the entity linking of the first knowledge base is continuously improved, thereby improving the accuracy of sentence parsing and question-answering interaction under actual application. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a flow chart of the knowledge base enhancement method designed by the present invention based on sentence entity linking; Figure 2 It is a system architecture diagram of the knowledge base enhancement method based on sentence entity linking designed by the present invention. DETAILED DESCRIPTION

[0018] The specific implementation modes of the present invention will be further described in detail below in conjunction with the accompanying drawings.

[0019] The present invention designs a knowledge base enhancement method based on sentence entity linking. In actual application design, for a first knowledge base including a preset number of entity words and each entity word corresponding to preset different types of description information, such as Figure 1 As shown, according to the following steps A to D, the first knowledge base is enhanced by performing semantic linking on the sentences to be parsed.

[0020] Step A. For the character sequence corresponding to the sentence to be parsed, use the sequence labeling model to perform sequence labeling. By locating and classifying the elements in the sequence, the entity words corresponding to the sentence to be parsed are obtained as the entity words to be analyzed to form an entity set, and then go to step B.

[0021] In practical applications, for sentences to be parsed, such as "Can you explain to me what RA means in the report form?", the candidate entities obtained are "report form" and "RA". Unlike general classification problems, the sequence labeling model outputs a label sequence. Generally speaking, the labels are interconnected, forming the structural information between the labels. Using this structural information, the sequence labeling model can often achieve higher performance than traditional classification methods in sequence labeling problems. The sequence labeling models that can be used include one or more of the following: Conditional Random Field (CRF), BiLSTM-CRF (Bi-directional Long Short-Term Memory-CRF), BERT-CRF (Bidirectional Encoder Representations from Transformers-CRF).

[0022] Step B. Based on a first knowledge base that contains a preset number of entity words and preset different types of description information corresponding to each entity word, apply a word embedding model to convert the word into a representation vector, and combine it with similarity calculation to obtain the corresponding entity words in the first knowledge base to which each entity word to be analyzed in the entity set is linked, and the corresponding description information, so as to realize the semantic link of the sentence to be parsed, and then enter step C.

[0023] In practical application, the above step B is specifically designed to execute the following steps B1 to B3.

[0024] Step B1. For each entity word to be analyzed in the entity set, respectively, the context information of the preset character length corresponding to the entity word to be analyzed is connected, and any word embedding model such as the Fasttext model, the Glove model, the ELMo model, and the BERT model is applied to obtain the representation vector corresponding to the entity word to be analyzed, and then obtain the representation vectors corresponding to each entity word to be analyzed in the entity set, and then enter step B2.

[0025] Step B2. Based on a first knowledge base that includes a preset number of entity words and preset different types of description information corresponding to each entity word, apply any word embedding model such as the Fasttext model, the Glove model, the ELMo model, and the BERT model to the combinations between each entity word and its corresponding description information, obtain the representation vectors corresponding to each combination, and then proceed to step B3.

[0026] Step B3. For each entity word to be analyzed in the entity set, calculate the similarity between the representation vector corresponding to the entity word to be analyzed and the representation vector corresponding to each combination in the first knowledge base according to a vector similarity calculation method such as the cosine distance function, and select each combination that is greater than a preset initial similarity threshold as the corresponding entity word in the first knowledge base linked to the entity word to be analyzed, and the corresponding description information.

[0027] In practical applications, if it is clear that RA stands for rheumatoid arthritis, then the answers in the conversation will be more inclined towards answers related to rheumatoid arthritis rather than others.

[0028] Step C. Based on each entity word in the first knowledge base, each different type of description information corresponding to each entity word, and the preset judgment rules corresponding to each different type of description information of each entity word, a second knowledge base is constructed. For each entity word to be analyzed in the entity set, according to the judgment rules, the context information of the preset character length corresponding to the entity word to be analyzed is linked to the entity word to be analyzed, and the corresponding entity words in the second knowledge base linked to the entity word to be analyzed, and the corresponding description information are obtained, and then the corresponding entity words in the second knowledge base linked to each entity word to be analyzed in the entity set, and the corresponding description information are obtained, so as to realize the semantic link of the sentence to be parsed, and then enter step D.

[0029] Step D. For each entity word to be analyzed in the entity set, determine whether the corresponding entity words and corresponding description information in the first knowledge base linked to the entity word to be analyzed are consistent with the corresponding entity words and corresponding description information in the second knowledge base linked to the entity word to be analyzed. If so, determine the entity words and corresponding description information linked to the entity word to be analyzed. Otherwise, supplement the data for the corresponding entity words and corresponding description information in the first knowledge base to enhance the first knowledge base.

[0030] In actual application of the above step D, if it is determined that the corresponding entity words and corresponding description information in the first knowledge base linked to the entity word to be analyzed are inconsistent with the corresponding entity words and corresponding description information in the second knowledge base linked to the entity word to be analyzed, then execute the following steps D1 to D3.

[0031] Step D1. Determine whether the corresponding entity words and corresponding description information in the first knowledge base linked to the entity word to be analyzed are consistent with the corresponding entity words and corresponding description information in the second knowledge base linked to the entity word to be analyzed. If so, proceed to step D2; otherwise, proceed to step D3.

[0032] Step D2. Taking the entity words and corresponding description information that are consistent with each other as the objects to be supplemented corresponding to the entity words to be analyzed, the description information of the objects to be supplemented corresponding to the entity words to be analyzed in the first knowledge base is supplemented respectively to enhance the first knowledge base.

[0033] Step D3. The corresponding entity words and corresponding description information in the second knowledge base linked to the entity word to be analyzed are used as the objects to be supplemented corresponding to the entity word to be analyzed, and the description information of the objects to be supplemented corresponding to the entity word to be analyzed in the first knowledge base is supplemented respectively to enhance the first knowledge base.

[0034] In the design of the enhancement of the first knowledge base involved in the above steps D2 and D3, two specific designs are implemented in actual application. First, the context information of the preset character length corresponding to the entity word to be analyzed is supplemented to the description information of each object to be supplemented corresponding to the entity word to be analyzed in the first knowledge base, so as to enhance the first knowledge base.

[0035] Secondly, based on the entire network, search for descriptive information about each object to be supplemented corresponding to the entity word to be analyzed, and supplement it to the descriptive information of each object to be supplemented corresponding to the entity word to be analyzed in the first knowledge base, so as to enhance the first knowledge base.

[0036] Steps A to D of the knowledge base enhancement method based on sentence entity linking designed by the present invention, while realizing the enhancement of the first knowledge base in actual application, further obtain the corresponding entity words and corresponding description information in the first knowledge base to which each entity word to be analyzed in the entity set corresponding to the sentence to be parsed is linked based on step B, and design and select the entity words and corresponding description information with the highest similarity in the first knowledge base to which each entity word to be analyzed is linked, so as to constitute the entity words and corresponding description information to which each entity word to be analyzed in the sentence to be parsed is linked, so as to realize the semantic link of the sentence to be parsed.

[0037] The design scheme of the present invention is applied in practice as described above, that is, the first knowledge base is used to realize semantic linking of the sentences to be parsed, and the second knowledge base is used to analyze the entity linking results of the first knowledge base, and the data of the first knowledge base is continuously enhanced. That is, while the first knowledge base is used to continuously realize semantic linking of the sentences to be parsed, its own data is continuously enhanced, and the accuracy of semantic linking for the sentences to be parsed is continuously improved.

[0038] The above design scheme is applied to the actual medical consultation dialogue, that is, the sentence to be parsed comes from the dialogue and communication process between the doctor and the patient during the consultation. The corresponding entity words in the first knowledge base and the second knowledge base are all preset entity words in the medical field, and the first knowledge base also includes preset entity words in the medical field corresponding to preset different types of description information, such as the "cold" entity word, and the corresponding type description information such as the symptoms corresponding to different stages of the cold, the "fever" entity word, and the corresponding type description information such as the symptoms corresponding to different stages of the fever. At the same time, the second knowledge base includes the entities of the first knowledge base and the corresponding type description information, and also includes the preset judgment rules corresponding to the different types of description information of each entity word. For example, for a cold or a fever, the preset judgment rules corresponding to different stages include the duration of the symptoms involved in the context of the consultation session, the patient's age, whether there are related medications, whether there are other symptoms, etc.

[0039] During the actual execution of the above-mentioned scenario of the embodiment of the patient-doctor consultation conversation, the voice is captured by an audio detection sensor such as a microphone during the conversation communication process, that is, each sentence in the conversation process is obtained, and then each sentence is used as a sentence to be parsed, and steps A to D are executed to perform semantic links on the sentences to be parsed. At the same time, while performing entity semantic recognition on the consultation conversation process, the first knowledge base data is continuously enhanced, and the accuracy of the semantic link feedback of the first knowledge base on the sentences to be parsed is continuously improved. That is, through technical design, each sentence of communication in the conversation process is analyzed more accurately and comprehensively, thereby improving the consultation efficiency in medical scenarios.

[0040] In practical applications, the above-mentioned design scheme is implemented by a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the knowledge base enhancement method based on sentence entity linking are implemented.

[0041] Modern intelligent question-answering systems usually generate answers through retrieval or text generation. In the retrieval and answer generation scenario, the system is responsible for identifying the question and its context, and uses this to search the answers in the question-answering database to find the answer that best matches the question and context. The system processes the context + entity link information of the answer and the standard questions and answers in the question-answering database through language models such as word2vec, LSTM, and BERT and converts them into representation vectors. The most similar answer can be obtained by comparing the representation vectors of the two. However, in actual question-answering systems, due to the ambiguity of the words, the same question, in the absence of the correct entity link, will have answers related to the correct word and answers related to the ambiguous word, and the two will be difficult to distinguish from the representation vectors that lack entity link information. For example, if the question contains RA, the answer about "Rheumatoid arthritis" and the answer about "Right Atrium" are similar to the question. By introducing entity link information, the answer about joint care will be more similar to the entity link of rheumatoid arthritis, and the answer about heart will be more similar to the entity link of right atrium. Therefore, the link object information is used to optimize the generation of answers in the retrieval-based question-answering system, which reflects the value of the design scheme of the present invention in practical application.

[0042] Apply the above design scheme to actual system design, such as Figure 2 As shown, the specific design includes a sentence acquisition module for identification, a candidate entity extraction module, a link object determination module, a link object use module, a link error judgment module, and a data supplement module, wherein each module is implemented in actual application as follows.

[0043] A module for obtaining sentences to be recognized, used for obtaining sentences to be parsed; Candidate entity extraction module, used to extract the entity set of the sentence to be parsed; A link object determination module, used to determine the entity link of each entity word to be analyzed in the entity set in the first knowledge base; A link object usage module is used to provide corresponding analysis or answers using the determined entity links in the dialogue; A link error judgment module is used to judge whether the analysis or answer of the sentence to be analyzed meets the preset failure condition according to the second knowledge base, and if it meets the condition, it is determined that the current entity link is wrong; The data supplement module is used to supplement the entity description information in the first knowledge base.

[0044] The above-mentioned design scheme of the present invention adopts a brand-new logical design, introduces a second knowledge base that combines the first knowledge base with the corresponding preset judgment rules, connects the context, and performs consistency comparison on the entity links of the sentence to be parsed with respect to the first knowledge base. Based on the comparison results, the entity link of the sentence to be parsed is realized, or data is supplemented for the corresponding entity words and corresponding description information in the first knowledge base to realize the enhancement of the first knowledge base. Through continuous data enhancement and accumulation of the first knowledge base under continuous application, the accuracy of the entity link of the first knowledge base is continuously improved, thereby improving the accuracy of sentence parsing and question-answering interaction under actual application.

[0045] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the above embodiments, and various changes can be made within the knowledge scope of ordinary technicians in this field without departing from the purpose of the present invention.

Claims

1. A knowledge base enhancement method based on sentence entity linking, characterized in that: For a first knowledge base including a preset number of entity words and each entity word corresponding to preset different types of description information, according to the following steps A to D, the first knowledge base is enhanced by semantically linking the sentences to be parsed; Step A. extract each entity word in the sentence to be parsed as each entity word to be analyzed, form an entity set, and then proceed to step B; Step B. Based on the first knowledge base containing a preset number of entity words and preset different types of description information corresponding to each entity word, the word embedding model is applied to convert the word into a representation vector, and combined with similarity calculation, the corresponding entity words in the first knowledge base linked to each entity word to be analyzed in the entity set and the corresponding description information are obtained to realize the semantic link of the sentence to be parsed, and then enter step C; Step C. Based on each entity word in the first knowledge base, each different type of description information corresponding to each entity word, and the preset judgment rules corresponding to each different type of description information of each entity word, a second knowledge base is constructed, and for each entity word to be analyzed in the entity set, according to the judgment rules, the context information of the preset character length corresponding to the entity word to be analyzed is linked, and the entity word to be analyzed is analyzed and judged, and the corresponding entity words in the second knowledge base linked to the entity word to be analyzed and the corresponding description information are obtained, and then the corresponding entity words in the second knowledge base linked to each entity word to be analyzed in the entity set and the corresponding description information are obtained, so as to realize the semantic link of the sentence to be parsed, and then enter step D; Step D. For each entity word to be analyzed in the entity set, determine whether the corresponding entity words and corresponding description information in the first knowledge base linked to the entity word to be analyzed are consistent with the corresponding entity words and corresponding description information in the second knowledge base linked to the entity word to be analyzed. If so, determine the entity words and corresponding description information linked to the entity word to be analyzed. Otherwise, supplement the data for the corresponding entity words and corresponding description information in the first knowledge base to enhance the first knowledge base.

2. According to claim 1, a knowledge base enhancement method based on sentence entity linking is characterized in that: In step A, a sequence labeling model is used to perform sequence labeling on the character sequence corresponding to the sentence to be parsed, and the entity words corresponding to the sentence to be parsed are obtained by locating and classifying the elements in the sequence, as the entity words to be analyzed, to form an entity set.

3. The knowledge base enhancement method based on sentence entity linking according to claim 1 is characterized in that: The step B includes executing the following steps B1 to B3; Step B1. For each entity word to be analyzed in the entity set, respectively, for the entity word to be analyzed, in conjunction with the context information of the preset character length corresponding to the entity word to be analyzed, the word embedding model is applied to obtain the representation vector corresponding to the entity word to be analyzed, and then the representation vectors corresponding to each entity word to be analyzed in the entity set are obtained, and then step B2 is entered; Step B2. Based on a first knowledge base including a preset number of entity words and preset different types of description information corresponding to each entity word, a word embedding model is applied to obtain a representation vector corresponding to each combination for each entity word and its corresponding description information, and then the process proceeds to step B3; Step B3. For each entity word to be analyzed in the entity set, calculate the similarity between the representation vector corresponding to the entity word to be analyzed and the representation vector corresponding to each combination in the first knowledge base according to the vector similarity calculation method, and select each combination that is greater than the preset initial similarity threshold as the corresponding entity word in the first knowledge base linked to the entity word to be analyzed, and the corresponding description information.

4. The knowledge base enhancement method based on sentence entity linking according to claim 3 is characterized in that: The word embedding model is any one of the Fasttext model, the Glove model, the ELMo model, and the BERT model.

5. The knowledge base enhancement method based on sentence entity linking according to claim 1 is characterized in that: In the step D, if it is determined that the corresponding entity words and corresponding description information in the first knowledge base linked to the entity word to be analyzed are inconsistent with the corresponding entity words and corresponding description information in the second knowledge base linked to the entity word to be analyzed, then the following steps D1 to D3 are executed; Step D1. Determine whether the corresponding entity words and corresponding description information in the first knowledge base linked to the entity word to be analyzed and the corresponding entity words and corresponding description information in the second knowledge base linked to the entity word to be analyzed are consistent with each other, if yes, proceed to step D2; Otherwise, proceed to step D3; Step D2. Based on the entity words and corresponding description information that are consistent with each other, as the objects to be supplemented corresponding to the entity words to be analyzed, the description information of the objects to be supplemented corresponding to the entity words to be analyzed in the first knowledge base is supplemented respectively, so as to enhance the first knowledge base; Step D3. The corresponding entity words and corresponding description information in the second knowledge base linked to the entity word to be analyzed are used as the objects to be supplemented corresponding to the entity word to be analyzed, and the description information of the objects to be supplemented corresponding to the entity word to be analyzed in the first knowledge base is supplemented respectively to enhance the first knowledge base.

6. The knowledge base enhancement method based on sentence entity linking according to claim 5 is characterized in that: In the steps D2 and D3, the context information of the preset character length corresponding to the entity word to be analyzed is supplemented to the description information of each object to be supplemented corresponding to the entity word to be analyzed in the first knowledge base, thereby enhancing the first knowledge base.

7. The knowledge base enhancement method based on sentence entity linking according to claim 5 is characterized in that: In step D2 and step D3, based on the entire network, the description information of each object to be supplemented corresponding to the entity word to be analyzed is searched, and the description information of each object to be supplemented corresponding to the entity word to be analyzed in the first knowledge base is supplemented respectively, so as to enhance the first knowledge base.

8. A knowledge base enhancement method based on sentence entity linking according to any one of claims 1 to 7, characterized in that: While enhancing the first knowledge base according to steps A to D, based on step B, the corresponding entity words and corresponding description information in the first knowledge base to which each entity word to be analyzed in the entity set corresponding to the sentence to be analyzed is linked are obtained, and the entity words and corresponding description information with the highest similarity in the first knowledge base to which each entity word to be analyzed is linked are selected to form the entity words and corresponding description information to which each entity word to be analyzed in the sentence to be analyzed is linked, so as to realize the semantic link of the sentence to be analyzed.

9. A knowledge base enhancement method based on sentence entity linking according to any one of claims 1 to 7, characterized in that: The sentences to be parsed are derived from the conversation and communication process between the doctor and the patient during consultation. By capturing the voice of the conversation and communication process, various sentences are obtained and used as the sentences to be parsed respectively. Steps A to D are executed to perform semantic links on the sentences to be parsed. The entity words in the first knowledge base and the second knowledge base are all preset entity words in the medical field.

10. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 8 are implemented.

Images