A method and apparatus for semantic retrieval based on a patent document database

Abstract

The present disclosure provides a patent document database-based semantic retrieval method and device, which comprises: performing preset processing on the retrieval text input by the user to obtain preset labels of the retrieval text; converting each preset label into a label word vector, and taking each label word vector as a target word vector in turn to determine m patent numbers with the highest similarity to the target word vector in the patent document database, and determining the score of the patent number according to the similarity between the patent number and the target word vector; constructing a score dictionary of each patent number to determine the total score of each patent number; and feeding back the k patent numbers with the highest total score to the user as the retrieval result of the retrieval text. The present disclosure performs structured splitting on the retrieval text, realizes retrieval in the patent document database-based retrieval process through more fine-grained and more accurate retrieval information, and obtains the retrieval result most meeting the user's demand by combining the score of the retrieval result under different preset labels.

Description

Technical Field

[0001] This disclosure relates to the field of data processing technology, and in particular to a semantic retrieval method, apparatus, storage medium and electronic device based on a patent document database. Background Technology

[0002] Currently, the information in patents that customers care about most and that has the highest information value includes technical field information, technical feature information, technical problem information, and technical effect information.

[0003] Existing platforms and literature describe patent semantic retrieval methods that are essentially black-box approaches. They do not reveal the key information within the input text to the client; the client simply inputs text and the system performs a background search, displaying a certain number of relevant patents as results. However, this approach lacks depth in understanding the client's original text and the relevance of the search results. Furthermore, current retrieval methods primarily rely on keywords within the original text without categorizing and associating different keywords. Consequently, some semantic content is not fully understood or utilized, leading to low semantic retrieval accuracy and poor overall performance. Summary of the Invention

[0004] The purpose of this disclosure is to provide a semantic retrieval method, apparatus, storage medium, and electronic device based on a patent document database, in order to solve the problems of low semantic retrieval accuracy and poor performance in the prior art.

[0005] The embodiments of this disclosure adopt the following technical solution: a semantic retrieval method based on a patent document database, comprising: performing preset processing on the retrieval text input by a user to obtain preset tags for the retrieval text, wherein the preset tags include at least a combination of domain tags and / or technical descriptions; converting each preset tag into a corresponding tag word vector, and using each tag word vector sequentially as a target word vector, determining the m patent numbers with the highest similarity to the target word vectors in the patent document database, and determining the score of each patent number based on the similarity between the patent number and the target word vectors; constructing a score dictionary for each patent number, and determining the total score of each patent number based on the score dictionary; and feeding back the k patent numbers with the highest total scores as the retrieval results of the retrieval text to the user.

[0006] This disclosure also provides a semantic retrieval device based on a patent document database, comprising: a preprocessing module for performing pre-processing on user-input retrieval text to obtain pre-defined tags for the retrieval text, wherein the pre-defined tags include at least a combination of domain tags and / or technical descriptions; a retrieval module for converting each pre-defined tag into a tag word vector, and using each tag word vector sequentially as a target word vector, determining the m patent numbers with the highest similarity to the target word vectors in the patent document database, and determining the score of each patent number based on the similarity between the patent number and the target word vectors; a score statistics module for constructing a score dictionary for each patent number, and determining the total score of each patent number based on the score dictionary; and a feedback module for feeding back the k patent numbers with the highest total scores as retrieval results for the retrieval text to the user.

[0007] This disclosure also provides a storage medium storing a computer program that, when executed by a processor, implements the steps of the semantic retrieval method based on a patent document database described above.

[0008] This disclosure also provides an electronic device, including at least a memory and a processor. The memory stores a computer program, and the processor executes the computer program in the memory to implement the steps of the semantic retrieval method based on the patent document database described above.

[0009] The beneficial effects of this disclosure are as follows: the search text input by the user is identified and extracted based on a combination of domain tags and technical descriptions, thereby achieving a structured breakdown of the search text. The most critical data in the search text, which corresponds to the main content modules of the patent, is extracted for subsequent retrieval. In the process of searching based on the patent document database, more granular and accurate search information is used. By combining the scores of the search results under different preset tags, the search results that best meet the user's needs are obtained, thereby improving the accuracy of the search. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in one or more embodiments of this specification or in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a flowchart of the semantic retrieval method based on a patent document database in the first embodiment of this disclosure;

[0012] Figure 2This is a schematic diagram of the architecture of the patent document database in the first embodiment of this disclosure;

[0013] Figure 3 This is a schematic diagram of the semantic retrieval device based on a patent document database in the second embodiment of this disclosure. Detailed Implementation

[0014] To enable those skilled in the art to better understand the technical solutions in one or more embodiments of this specification, the technical solutions in one or more embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this specification, and not all of the embodiments. Based on one or more embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of this document.

[0015] Existing platforms and literature describe patent semantic retrieval methods that are essentially black-box approaches. They do not reveal the key information within the input text to the client; the client simply inputs text and the system performs a background search, displaying a certain number of relevant patents as results. However, this approach lacks depth in understanding the client's original text and the relevance of the search results. Furthermore, current retrieval methods primarily rely on keywords within the original text without categorizing and associating different keywords. Consequently, some semantic content is not fully understood or utilized, leading to low semantic retrieval accuracy and poor overall performance.

[0016] To address the aforementioned problems, the first embodiment of this disclosure provides a semantic retrieval method based on a patent document database, the flowchart of which is shown below. Figure 1 As shown, it includes at least steps S10 to S40:

[0017] S10: Perform preset processing on the search text input by the user to obtain preset tags for the search text.

[0018] When performing semantic search, users input relevant search text into the search box. This search text can include the technical features the user wishes to retrieve, the technical problems these features solve, the technical effects they achieve, and the technical fields in which they are applied. It also contains redundant information and distractors used in the textual description. After receiving the user's input search text, the search engine first performs pre-processing to extract pre-defined tags. In this embodiment, the pre-defined tags include at least domain tags and / or combinations of technical descriptions. Domain tags primarily describe the technical field information corresponding to the search text, while technical descriptions primarily characterize the descriptions of technical features, technical problems, and technical effects within the search text. Examples include the subject of a technical feature and the relationships between subjects, the subject of a technical problem and its description, and the subject of a technical effect and its description. By extracting pre-defined tags from the search text, a structured breakdown of the search text is achieved. This allows for the extraction of more refined domain information and descriptive information corresponding to the main content of the patent document. This removes the influence of redundant information during subsequent searches, improving the accuracy of semantic search.

[0019] In some embodiments, after obtaining the preset tags of the search text, all preset tags can be displayed to the user so that the user can have a deeper understanding of the search text they entered, and facilitate the user to filter and use the final search results.

[0020] In some embodiments, the process of pre-processing the user-input search text to obtain pre-defined tags for the search text mainly includes: performing first named entity recognition on the search text using a first named entity recognition model to identify domain tags in the search text. Specifically, the first named entity recognition model is a pre-trained model for entity recognition of domain information in text, which can extract domain tags by using specific keywords or key entities in the search text. In actual use, there are also situations where domain tags cannot be extracted from the search text, such as when the search text content is short, does not clearly define the user's expected usage scenario, or does not identify entities with specific functions.

[0021] In some embodiments, the process of performing preset processing on the search text input by the user to obtain preset tags for the search text may further include: performing second named entity recognition on the search text using a second named entity recognition model to identify the subject tags and description tags in the search text; performing feature sentence segmentation on the search text based on punctuation marks, and combining all subject tags and / or all description tags belonging to the same feature sentence to obtain a technical description combination.

[0022] Specifically, the second named entity recognition model is a pre-trained model used to recognize entities based on subject information and descriptive information in text. Subject information can include subjects representing features, subjects representing technical problems, and subjects representing technical effects. Descriptive information includes feature relationships between subject subjects, specific descriptions of the subject representing technical problems, and specific descriptions of the subject representing technical effects. Corresponding to the retrieved text, it may have multiple different subject tags and descriptive tags. These tags are actually a simplified extraction of the technical description content of the retrieved text, filtering out redundant information and interference words.

[0023] After extracting the tags, it is necessary to associate and combine the main tags and descriptive tags to clarify the true content disclosed in the search text and represent it through concise and accurate tag combinations. Generally, the main subjects described in the same sentence are related. Therefore, this embodiment uses the position of punctuation marks as a basis to divide the search text into feature sentences, and uses these feature sentences as the basic unit for determining the association between tags to determine the combination of technical descriptions.

[0024] For each segmented feature statement, a statement start index and a statement end index are determined. These can be the position information of the characters marking the beginning and end of the feature statement. Similarly, for subject tags and description tags, when determining which feature statement they belong to, corresponding word start indices and word end indices need to be determined. Then, based on these indices, all subject tags and / or description tags within the statement start and end indices of each feature statement are identified. This is equivalent to identifying a group of related subject tags and description tags. Arranging these subject tags and description tags belonging to the same feature statement in ascending order of word start indices yields a set of technical description combinations. This set of technical descriptions records all the key subjects in the feature statement and the specific relationships between them, achieving the goal of simplifying the technical solution. In the specific implementation, each feature statement is traversed, and the main labels and / or description labels that have a word start index greater than or equal to the statement start index and a word end index less than or equal to the statement end index are determined as the main labels and / or description labels located within the statement start index and statement end index of each feature statement.

[0025] S20: Convert each preset label into a label word vector, and use each label word vector as a target word vector in turn. Determine the m patent numbers with the highest similarity to the target word vectors in the patent document database, and determine the score of the patent number based on the similarity between the patent number and the target word vector.

[0026] After step S10 is executed, one or more preset tags can be extracted from the search text. Based on all these preset tags, a search can be performed in the patent document database. During the search, each preset tag is converted into a corresponding tag word vector. Each tag word vector is then used as a target word vector to determine the m patent numbers with the highest similarity in the patent document database. These m patent numbers represent the search results for the preset tag currently used as the target word vector in the database. The similarity between the patent number and the target word vector is the score of the patent number under the corresponding preset tag. In actual implementation, m patent numbers and their scores are output for each preset tag. Therefore, the actual output of this step may be m times the number of preset tags and their scores.

[0027] Specifically, corresponding to the combination of domain tags and technical descriptions, the patent document database has different indexes to record the association between all patent numbers and their respective domains and technical descriptions. That is, the domain index records the correspondence between all patent numbers and their corresponding domain indexes, while the technical description index records the correspondence between all patent numbers and their associated technical description indexes. Based on the identification and extraction results of the preset tags in step S10, different indexes are used for retrieval when the preset tags include different types.

[0028] In some embodiments, when the preset label includes a domain label, the domain label is first converted into a corresponding domain word vector. Then, the distance between each domain index and the domain word vector is determined in the domain index library. Based on the calculation results of the distances between all domain indices and domain word vectors, the m domain indices closest to the domain word vectors are determined. The distance between each of the m domain indices and the domain word vectors is normalized to obtain the scores of the m domain indices. Finally, the patent numbers corresponding to the m domain indices are determined in the domain index library, and the scores of the domain indices are used as the scores of their corresponding patent numbers under the current domain label.

[0029] In some embodiments, when the preset label includes a technical description label, the technical description label is first converted into a corresponding technical description word vector. Then, the distance between each technical description index and the technical description word vector is determined in the technical description index library. Based on the calculation results of the distances between all technical description indices and technical description word vectors, the m technical description indices closest to the technical description word vectors are determined. The distance between each of the m technical description indices and the technical description word vectors is normalized to obtain the scores of the m technical description indices. Finally, the patent numbers corresponding to the m technical description indices are determined in the technical description index library, and the scores of the technical description indices are used as the scores of their corresponding patent numbers under the current technical description label.

[0030] Typically, a single domain tag can be extracted from a search text, representing the domain most desired for application or retrieval. However, technical description tags can reflect three aspects: technical features, technical problems, and technical effects. Furthermore, each aspect may have multiple technical subjects and descriptions. Therefore, the number of technical description tags is usually not unique. In the process of determining the corresponding patent number and calculating the score, it is necessary to determine the corresponding patent number and score for each technical description tag. The m patent numbers corresponding to different technical description tags may contain the same patent or different patents.

[0031] S30, construct a score dictionary for each patent number, and determine the total score for each patent number based on the score dictionary.

[0032] The scoring dictionary is constructed primarily based on the patent number, recording the scores of each patent under different preset tags. It can be constructed in the following format: {'Patent Number': [{'Technical Field': []}, 'Technical Description': []]}, where the square brackets after 'Technical Field' and 'Technical Description' record the score value of the current patent number under the combination of the technical field tag and the technical description. The technical description score, in particular, is recorded as the sum of the scores of the patent number under different combinations of technical descriptions. Corresponding to the calculation of the total score, firstly, the product of the sum of the scores of the patent number under different combinations of technical descriptions and the preset technical weight is calculated. The technical weight serves to balance the quantitative difference between the technical field score and the technical description score. Since the technical description score is usually the sum of the scores of the patent number under multiple combinations of technical descriptions, its specific value is much larger than the technical field score, thus it is adjusted through the technical weight. Then, the product is added to the technical field score of the patent number, and the sum is recorded as the total score of the patent number. It should be noted that the specific value of the above technical weight can be set based on actual needs and historical experience; this embodiment does not impose specific restrictions.

[0033] S40 returns the k patent numbers with the highest total scores as the search results to the user.

[0034] After calculating the total score of all patent numbers, the patent document search results based on the preset tag extraction results of the search text can be obtained. The k patent numbers with the highest scores are selected as the search results and fed back to the user. The solutions of the above k patent numbers correspond to the technical content that is most similar to the actual description of the search text.

[0035] In practice, the total score of all patent numbers can be normalized and sorted, and the total score can be limited to between score_low and score_high. Then, the top k patent numbers can be selected and returned.

[0036] This embodiment performs identification and extraction on the user-input search text based on a combination of domain tags and technical descriptions, achieving structured decomposition of the search text. It extracts the most critical data in the search text that corresponds to the main content modules of the patent for subsequent retrieval. This enables more granular and precise search information in the process of searching based on the patent document database. By combining the scores of the search results under different preset tags, the most suitable search results are obtained, thereby improving the search accuracy.

[0037] In actual implementation, before the user performs semantic search, the process may also include building a database of patent documents. During the construction of the patent document database, it is also necessary to use more refined and relevant information such as domain tags and combinations of technical descriptions to achieve more accurate search results in conjunction with the semantic search method provided in this embodiment.

[0038] A patent document typically includes the following parts: the technical field section, the technical features section, the technical problem section, and the technical effect section. The technical field section refers to the technical field to which the patent document belongs, as indicated in the specification. This technical field can be described by specific field information or by entities with specific functions included in the patent document, such as a method or apparatus. The technical features section describes the actual technical solution of the patent document, such as the claims or the invention description section of the specification. The technical problem section describes the technical problem that the patent document aims to solve, usually described in the background section of the specification. The technical effect section describes the beneficial effects achieved by the technical solution, typically found in the latter part of the invention description.

[0039] In some embodiments, a database of patent documents can be constructed first based on the technical field. For example, the technical field portion of the patent documents can be named entity recognized using a first named entity recognition model. This includes the technical field information and descriptions of the specific apparatus or method involved in the patent document. The corresponding field tag and subject tag can be identified using a NER model, corresponding to the technical field information and the specific apparatus or method of the patent, respectively. In practical applications, some patent documents may not describe the technical field information in the technical field portion. Therefore, the corresponding field tag may not be identifiable for these patent documents. However, the technical field portion will certainly record the subject of the disclosed solution. Therefore, each patent document should be able to identify the subject tag. Based on the named entity recognition results, all patent documents are divided into a first category of patent documents and a second category of patent documents. The first category of patent documents contains patent documents with identified field tags and subject tags, while the second category of patent documents contains patent documents with only identified subject tags.

[0040] For all first subject tags that have been identified as having domain and subject tags, a mapping dictionary is constructed based on the extracted domain and corresponding specific device / method information. After the mapping dictionary for all first-class patent documents is constructed, the corresponding domain tags for each second subject tag in the second-class patent documents can be determined based on the mapping dictionary, so as to achieve the allocation of specific domain information for each patent document. Specifically, based on the constructed mapping dictionary, for each second subject tag, the first subject tag most similar to that second subject tag can be determined from the mapping dictionary. Based on the similarity of the two subject tags, their corresponding domain tags should also be similar. Therefore, at least one first domain tag corresponding to the first subject tag most similar to that second subject tag in the mapping dictionary is assigned to that second subject tag as at least one second domain tag corresponding to that second subject tag. After performing the above processing on all second-class patent documents, the domain information applicable to each patent document can be determined.

[0041] After all patent documents have been assigned their corresponding field labels, a mapping between the patent number of each patent document and its corresponding first field label can be established, forming a patent document database based on the technical field. In the actual establishment process, the patent number can be the application number or publication number of the patent document, and the mapping between each patent number and the field can be stored in matrix form.

[0042] In some embodiments, a patent document database can also be constructed based on the technical description. For example, a second named entity recognition model can be used to perform named entity recognition on the technical description portion of the patent document. This involves identifying the descriptive subject and specific descriptive content of the patent document. The NER model can then identify the corresponding subject tag and description tag, which correspond to the descriptive subject and specific descriptive content appearing in the patent document, respectively. In actual implementation, the subject tag and description tag have different representational meanings for different content portions of the patent document. Specifically, when the technical description portion is the claims, the subject tag represents the feature subject in the claims, while the description tag represents the feature relationship between the feature subjects; when the technical description portion is the background art portion, the subject tag represents the technical problem subject in the background art portion, while the description information represents the problem description of the technical problem subject; when the technical description portion is the beneficial effect portion, the subject tag represents the technical effect subject in the beneficial effect portion, while the description information represents the effect description of the technical effect subject.

[0043] After extracting the tags, it is necessary to associate and combine the main tags and description tags to clarify the true content disclosed in the patent document and represent it through concise and accurate tag combinations. Generally, the main entities described in the same sentence are related. Therefore, this embodiment uses the position of punctuation marks as a benchmark to divide the technical description into feature sentences, and uses these feature sentences as the basic unit for determining the association between tags to determine the combination of technical descriptions. For each segmented feature sentence, a sentence start index and a sentence end index are determined. These can be the position information of the character marking the beginning and end of the feature sentence. Correspondingly, when determining the feature sentences to which the main tags and description tags belong, corresponding word start indexes and word end indexes are also needed. This indexing can be performed during entity recognition or after the sentence-related indexes are determined.

[0044] Subsequently, based on the word start index and word end index, all subject tags and / or description tags located within the sentence start index and sentence end index of each feature statement are determined. This is equivalent to determining a group of subject tags and description tags with related relationships. The subject tags and description tags belonging to the same feature statement are arranged in ascending order according to the word start index to obtain a set of technical description combinations. This set of technical description combinations records all the key subjects in the feature statement and the specific relationships between the key subjects, thus achieving the purpose of simplifying the technical solution.

[0045] After traversing all feature statements, all combinations of technical descriptions for the patent document are obtained. This allows for the establishment of a correspondence between the patent number and all corresponding combinations of technical descriptions for each patent document, forming a patent document database based on technical descriptions. After all combinations of technical descriptions for all patent documents have been determined, in the actual database creation process, the patent number can be either the application number or the publication number of the patent document. The correspondence between each patent number and each combination of technical descriptions can be stored in matrix form.

[0046] After the patent document database based on the technical field and the patent document database based on the technical description are constructed, the field tags and technical descriptions corresponding to each patent number can be combined and converted into word vectors. Then, the Faiss library is used to further convert them into corresponding indexes to accelerate subsequent retrieval tasks. Finally, the correspondence between patent numbers, field indexes, and technical description indexes is transformed into a field index library and a technical description index library, thus completing the overall construction of the patent document database.

[0047] In some embodiments, when constructing a patent document database based on technical description, the technical features, technical problems, and technical effects can be separated into three parts, and patent document data can be constructed based on these three parts respectively. This results in patent document databases based on technical features, technical problems, and technical effects. The construction process for these three databases can directly adopt the construction method for a patent document database based on technical description, and will not be repeated here. Simultaneously, feature indexes, problem indexes, and effect indexes are constructed and transformed accordingly. Furthermore, during semantic retrieval, the combination of technical descriptions can be determined based on the three parts of technical features, technical problems, and technical effects, and patent number retrieval based on the feature indexes, problem indexes, and effect indexes can be performed accordingly. The resulting patent document database is as follows: Figure 2 As shown. At this point, the formula for calculating the total score for each patent can be:

[0048] Total score = Technical field score * (q1 * Technical feature score + q2 * Technical problem score + q3 * Technical effect score), where q1, q2, and q3 correspond to the trained weights. The technical feature score is the sum of the scores of the current patent under each different technical feature. The technical problem score and technical effect score are calculated similarly.

[0049] Based on the above, this disclosure combines technical field information, technical feature information, technical problem information, and technical effect information as a foundational retrieval database, thereby more comprehensively and accurately representing the key information of the input text. During semantic retrieval, key information from multiple modules is integrated, making the retrieval results more complete and better representing the full key information of the input text. Simultaneously, combined retrieval through weighting and combination methods makes the retrieval results more reasonable and accurate, more comprehensively utilizing the data information in the patent, improving the value and utilization rate of patent information, and enabling its application to patent texts with different fields and technical features, demonstrating strong versatility and applicability.

[0050] Based on the same inventive concept, the second embodiment of this disclosure provides a semantic retrieval device based on a patent document database, the structural schematic diagram of which is shown below. Figure 3 As shown, the system mainly includes a preprocessing module 10, which performs preprocessing on the user-input search text to obtain pre-defined tags for the search text. The pre-defined tags include at least a combination of domain tags and / or technical descriptions. A retrieval module 20 is used to convert each pre-defined tag into a tag word vector, and uses each tag word vector as a target word vector to determine the m patent numbers with the highest similarity to the target word vectors in the patent document database. The score of each patent number is determined based on the similarity between the patent number and the target word vector. A score statistics module 30 is used to construct a score dictionary for each patent number and determine the total score for each patent number based on the score dictionary. A feedback module 40 is used to provide the user with the k patent numbers with the highest total scores as the search results for the search text.

[0051] For a detailed description of the above modules and their corresponding functions, please refer to the content disclosed in the first embodiment of this disclosure. This embodiment will not repeat the description again.

[0052] Based on the same inventive concept, the third embodiment of this disclosure provides a storage medium storing a computer program, which, when executed by a processor, implements the steps of the semantic retrieval method based on a patent document database described in the first embodiment of this disclosure.

[0053] Based on the same inventive concept, the fourth embodiment of this disclosure provides an electronic device, including at least a memory and a processor. The memory stores a computer program, and the processor, when executing the computer program in the memory, implements the steps of the semantic retrieval method based on a patent document database described in the first embodiment of this disclosure.

[0054] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this disclosure, and are not intended to limit them. Although this disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this disclosure.

Claims

1. A semantic retrieval method based on a patent document database, characterized in that, include: The search text input by the user is processed in a preset manner to obtain preset tags for the search text. The preset tags include at least a combination of domain tags and / or technical descriptions. Each preset label is converted into a label word vector, and each label word vector is used as a target word vector in turn. The m patent numbers with the highest similarity to the target word vector are determined in the patent document database. The score of the patent number is determined based on the similarity between the patent number and the target word vector. Construct a score dictionary for each of the patent numbers, and determine the total score for each patent number based on the score dictionary; The k patent numbers with the highest total scores are used as the search results for the search text and fed back to the user. When the preset tags include the domain tags, the step of converting each preset tag into a corresponding tag word vector, and using each tag word vector as a target word vector in sequence, determining the m patent numbers with the highest similarity to the target word vectors in the patent document database, and determining the score of the patent number based on the similarity between the patent number and the target word vectors, includes: The domain labels are converted into corresponding domain word vectors, and the distance between each domain index and the domain word vector is determined in the domain index library; Determine the m domain indices that are closest to the domain word vector, and normalize the distance between each of the m domain indices and the domain word vector to obtain the score of the domain index; Determine the patent number corresponding to each of the m domain indices, and use the score of the domain index as the score of the patent number under the domain label.

2. The semantic retrieval method according to claim 1, characterized in that, After performing preset processing on the user-input search text to obtain preset tags for the search text, the method further includes: Display all the preset labels to the user.

3. The semantic retrieval method according to claim 1, characterized in that, The step of performing preset processing on the user-input search text to obtain preset tags for the search text includes: The first named entity recognition model is used to perform first named entity recognition on the search text in order to identify the domain tags in the search text.

4. The semantic retrieval method according to claim 1, characterized in that, The step of performing preset processing on the user-input search text to obtain preset tags for the search text includes: The search text is subjected to second named entity recognition using a second named entity recognition model to identify the subject tags and description tags in the search text. The retrieved text is segmented into feature sentences based on punctuation marks. All the main tags and / or all the description tags belonging to the same feature sentence are combined to obtain a technical description combination.

5. The semantic retrieval method according to claim 1, characterized in that, When the preset tags include the combination of technical descriptions, the step of converting each preset tag into a corresponding tag word vector, and using each tag word vector as a target word vector in sequence, determining the m patent numbers with the highest similarity to the target word vectors in the patent document database, and determining the score of the patent number based on the similarity between the patent number and the target word vectors, includes: The technical descriptions are combined and converted into corresponding technical description word vectors. The distance between each technical description index and the technical description word vectors is determined in the technical description index library. Determine the m technical description indices that are closest to the technical description word vector, and normalize the distance between each of the m technical description indices and the technical description word vector to obtain the score of the technical description index; Determine the patent number corresponding to each of the m technical description indices, and use the score of the technical description index as the score of the patent number under the technical description combination.

6. The semantic retrieval method according to claim 5, characterized in that, The process of constructing a score dictionary for each patent number and determining the total score for each patent number based on the score dictionary includes: Based on the patent number and the score of the patent number under the field label and / or the score under the combination of technical descriptions, a score dictionary is constructed, wherein the score of the patent number under the combination of technical descriptions is the sum of the scores of the patent number under different combinations of technical descriptions; Calculate the product between the sum and the technology weight, and sum the product with the score of the patent number under the field label to obtain the total score of the patent number.

7. A semantic retrieval device based on a patent document database, characterized in that, include: The preprocessing module is used to perform pre-processing on the search text input by the user to obtain pre-defined tags for the search text. The pre-defined tags include at least a combination of domain tags and / or technical descriptions. The retrieval module is used to convert each of the preset tags into tag word vectors, and use each of the tag word vectors as target word vectors in turn to determine the m patent numbers with the highest similarity to the target word vectors in the patent document database, and determine the score of the patent number based on the similarity between the patent number and the target word vector; The scoring statistics module is used to construct a scoring dictionary for each patent number and determine the total score for each patent number based on the scoring dictionary. The feedback module is used to provide the user with the k patent numbers with the highest total scores as the search results of the search text; When the preset tag includes the domain tag, the retrieval module is specifically used to: convert the domain tag into a corresponding domain word vector, determine the distance between each domain index and the domain word vector in the domain index library; determine the m domain indexes that are closest to the domain word vector, and normalize the distance between each of the m domain indexes and the domain word vector to obtain the score of the domain index; Determine the patent number corresponding to each of the m domain indices, and use the score of the domain index as the score of the patent number under the domain label.

8. A storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the semantic retrieval method based on a patent document database as described in any one of claims 1 to 6.

9. An electronic device, comprising at least a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program on the memory, it implements the steps of the semantic retrieval method based on the patent document database as described in any one of claims 1 to 6.

Images