An intelligent question-answering system for power grid planning based on text parsing

Abstract

The present invention relates to an intelligent question-answering system for power grid planning based on text parsing, which includes an application layer, a support layer, a knowledge layer and a data layer. The knowledge layer is equipped with a text parsing module, and a text parsing module dedicated to parsing text information is set up. Through the support of two algorithms, information structure parsing and semantic parsing, the key problem of the difficulty in extracting knowledge information from policy texts is solved. First, by establishing an industry term library to mark the magnetic characteristics of each power grid industry term, the accuracy of the system's analysis of power grid planning professional knowledge is improved. Second, by associating knowledge information elements through structural information, the entire text information forms a mesh topology structure. Based on this type of mesh topology structure, the correlation between text information can be analyzed and new correlation relationships can be generated, so that the knowledge graph has accuracy and avoids errors in answer information due to the invalidation of some clauses in the text information, time coverage, and different priorities.

Description

Technical Field

[0001] The present invention relates to the field of smart grid technology, and more specifically, to an intelligent question-answering system for grid planning based on text parsing. Background Art

[0002] At present, with the popularization of smart grids, all kinds of grid information are summarized and analyzed in the form of grid data to support decision-making. For electricity development planning projects, a large amount of technical materials, policy texts, etc. need to be mobilized. The collection of relevant information of the original grid development plan has always been the current pain point. Due to the different timeliness and effectiveness of a large number of policy texts, and the high professionalism and large number of details of the policy texts, the text types, principles to follow, scope of application, subjects, requirements, indicators and other information are all different. At this time, a large number of professionals are needed to read, archive and understand. Although electronic data of all texts are retained, these electronic data have not been analyzed. In actual operation, when problems are found, a large number of documents related to the problems are started to be found. This method can obtain the answer to the question, but each text contains a large amount of content that is irrelevant to the question, which will generate a lot of invalid reading. In response to the above problems, the authorization announcement number CN110909140B discloses a central routing method and device that combines graph question answering with retrieval question answering. It answers user questions by constructing a knowledge graph combined with semantic recognition of users. Although this method is suitable for ordinary power grid problems, if it is migrated to information retrieval for power grid planning, there is still a very large gap in information granularity. Because the text information related to power grid planning is highly professional, and the power grid development planning department has a large demand for information accuracy and relevance, there is a need for an intelligent question-answering system that can perform more detailed analysis of power grid text. Summary of the Invention

[0003] In view of this, the object of the present invention is to provide an intelligent question-answering system for power grid planning based on text parsing.

[0004] In order to solve the above technical problems, the technical solution of the present invention is: a power grid planning intelligent question-answering system based on text analysis, including an application layer, a support layer, a knowledge layer and a data layer.

[0005] The knowledge layer is configured with a text parsing module, which is used to parse the text information input by the data layer and generate a text element map;

[0006] The text parsing module includes a structure parsing unit, a semantic matching unit, and an information association unit; the structure parsing unit is associated with a structure feature library and a structure type library, the structure feature library stores a plurality of structure features, and the structure type library stores structure type information; the structure parsing unit traverses the corresponding text information to identify the same structure features through the structure feature library, and matches the closest structure type information from the structure type library based on the order and position relationship of the corresponding structure features in the text information;

[0007] The semantic matching unit is associated with an industry terminology library, which stores power grid industry terms and configures a part-of-speech feature for each power grid industry term. The semantic matching unit marks words corresponding to power grid industry terms in text information using the part-of-speech features, and is configured with a semantic recognition algorithm to perform semantic recognition on the marked text information to generate a plurality of knowledge information elements.

[0008] The information association unit includes a first association strategy and a second association strategy. The first association strategy establishes a first association mark between knowledge information elements based on structural type information. The second association strategy screens knowledge information elements with association characteristics from knowledge information elements based on the first association mark, compares knowledge information elements with association characteristics of different text information to determine the text association relationship between text information, establishes a second association mark between knowledge information elements of different subordinate text information based on the text association relationship, and forms the text element map based on the first association mark, the second association mark and the knowledge information elements.

[0009] Furthermore, the knowledge layer is further configured with a data configuration module, which includes a feature configuration unit. The feature configuration unit splits each structural type information in the structural type library into a number of structural features. Each structural feature in the structural feature library is configured with a feature priority value, which is D=a / (A1α1 / β1+K+A n α n / β n ), where D is the feature priority value, a is the preset priority value configuration parameter, and A n is the matching association value of the nth structural type information with the structural feature, and the matching association value reflects the matching reliability of the structural type information. n is the number of all structural features in the nth structural type information with this structural feature, β n is the number of the structural feature in the n-th structural type information having the structural feature;

[0010] The structure parsing unit determines the matching order of the structural features in the structure type information library according to the priority value.

[0011] Furthermore, the data configuration module further includes a type association unit, which configures a matching association value for each structural type information according to the input text sample, and A=1 / [(t0-t1)-cM] 2 χ1+K+1 / [(t0-t m )-cM] 2 χ m , where A is the matching association value, t0 is the current time, t m is the creation time of the mth text sample, c is the preset sensitive adjustment variable, M is the total amount of text samples, χ m is the known matching degree between the mth text sample and the structural type information;

[0012] The structure analysis unit calculates the matching degree of each structure type information according to the matching correlation value, and Δχ=A(h1+K+h g ), where Δχ is the matching degree, h g is the similarity of the g-th structural feature, and the structure analysis unit determines the structural type information with the highest matching degree as the closest structural type information.

[0013] Furthermore, the knowledge layer also includes a data extraction module, which includes a vocabulary extraction unit and a part-of-speech tagging unit. The vocabulary extraction unit is associated with several industry terminology databases of the data layer and extracts power grid industry terms from the industry terminology database. The part-of-speech tagging unit is used to mark the part-of-speech features of the extracted power grid industry terms, and the part-of-speech features include the type of industry terminology database.

[0014] Furthermore, the data extraction module further includes a feature configuration unit, the feature configuration unit being used to configure the recognition priority value of each part-of-speech feature;

[0015] The semantic recognition algorithm determines the part-of-speech features as indexes in order of recognition priority values, and performs semantic recognition on text information based on the determined part-of-speech features. When the recognition result meets the first recognition condition, the corresponding knowledge information element is output.

[0016] Furthermore, the semantic recognition algorithm includes a plurality of structured segments, each structured segment includes fixed items and parameter items, and each structured segment is indexed by part-of-speech features. The semantic recognition algorithm includes

[0017] Step A1: selecting part-of-speech features in the order of recognition priority values ​​to determine the corresponding structured segment;

[0018] Step A2: determining fixed items in the target segment and calculating their association degree to generate a first recognition value;

[0019] Step A3: determining a parameter item in the target segment according to the position of the determined fixed item in the target segment, and verifying its data format to generate a second recognition value;

[0020] Step A4: determining the remaining information in the target segment to generate a third recognition value;

[0021] Step A5: summing the first recognition value, the second recognition value, and the third recognition value to obtain a semantic recognition value;

[0022] Step A6: Determine whether the semantic recognition value meets the first recognition condition. If so, extract parameter items from the target segment using the structured segment to generate the knowledge information element. If not, re-enter step A1.

[0023] Furthermore, the first recognition condition is configured with a first recognition threshold and a second recognition threshold. If the semantic recognition value exceeds the first recognition threshold, or the semantic recognition value of the structured segment is greater than the semantic recognition mean second recognition threshold, the first recognition condition is deemed to be met. The semantic recognition mean is the average value of the semantic recognition thresholds obtained for all previously recognized structured segments.

[0024] Furthermore, the part-of-speech priority value is a weighted sum of a word priority value and a lexicon priority value, and the feature configuration unit includes a word priority algorithm and a lexicon priority algorithm. The word priority algorithm is to add a preset first priority increment to the word priority value of a power grid industry term when it is identified in text information, and at the same time add a preset second priority increment to the word priority values ​​of other power grid industry terms according to the similarity weight between the power grid industry terms; the lexicon priority algorithm is to add a preset third priority increment to the lexicon priority values ​​of all power grid industry terms that belong to the same industry terminology database as the power grid industry term when it is identified in text information.

[0025] Furthermore, the first association strategy includes determining the subordinate relationships between knowledge information elements based on the structure type information, identifying missing elements of the knowledge information elements, and generating a first association mark based on the identified missing elements and the subordinate relationships.

[0026] Furthermore, the second association strategy includes determining the knowledge information elements as basic information based on the first association mark, and matching the knowledge information elements of different text information to determine the element competing items among the knowledge information elements whose matching results meet the second comparison conditions, retrieving the corresponding knowledge information elements as basic information based on the first association mark to identify the element competing items, and generating corresponding competing conditions based on the element competing items, wherein the element competing items include time competing items, policy level competing items, and regulation priority competing items.

[0027] The technical effects of the present invention are mainly reflected in the following aspects: a text parsing module dedicated to parsing text information is set up, and through the support of two algorithms, information structure parsing and semantic parsing, the key problem of the difficulty in extracting knowledge information from policy texts is solved, and the magnetic characteristics of each power grid industry term are marked by establishing an industry term library, which provides the possibility of implementation for the semantic recognition algorithm. On the other hand, the knowledge information elements are associated through structural information, so that the entire text information forms a mesh topology structure. According to this type of mesh topology structure, the correlation between text information can be analyzed and new correlation relationships can be generated, so that the knowledge graph has accuracy, avoiding errors in answer information due to the invalidity of some clauses in the text information, time coverage, and different priorities. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 : Schematic diagram of the architecture of an intelligent question-answering system for power grid planning based on text parsing in the present invention;

[0029] Figure 2 : The functional module architecture principle diagram of a power grid planning intelligent question-answering system based on text parsing of the present invention. DETAILED DESCRIPTION

[0030] The specific embodiments of the present invention are further described below in conjunction with the accompanying drawings to make the technical solutions of the present invention easier to understand and grasp.

[0031] Reference Figure 1As shown, a text-parsing-based intelligent question-and-answer system for power grid planning consists of an application layer, a support layer, a knowledge layer, and a data layer. Generally speaking, existing question-and-answer systems also have a four-layer architecture. The application layer's main function is to enable user interaction, such as questions and answers about policies, regulations, standards, system functions, and business indicators, as well as in-depth analysis of energy policies and development standards. It then performs monitoring and represents the definition and maintenance of the text framework. The support layer provides services for the application layer, such as semantic parsing services to parse user questions, information extraction services to extract information based on parsing results, image content extraction services to extract information based on input images, text analysis services to analyze content based on uploaded text, and knowledge inference services to infer corresponding elements based on the obtained content. The support layer also provides basic services such as messaging, workflow, security, permissions, and monitoring. The data layer mainly provides a data foundation, such as news, consultation, system standards, analysis reports, topological data, project data, map data, policies and regulations, experience cases, professional libraries, training videos, electricity data, archival data, practical data, etc. The knowledge layer mainly extracts data, establishes a basic knowledge base, and then constructs a knowledge graph based on the basic knowledge base, and provides support for intelligent question and answer through an algorithm model. The core of this invention lies in how to accurately extract information elements such as policy texts and generate a knowledge graph that can distinguish the competitive and cooperative relationships between texts, so that when users ask questions, they can accurately find the corresponding elements in the policy text to answer. The specific plan is as follows:

[0032] The knowledge layer is configured with a text parsing module, which is used to parse the text information input by the data layer and generate a text element map;

[0033] The text parsing module includes a structure parsing unit, a semantic matching unit and an information association unit; the structure parsing unit is associated with a structure feature library and a structure type library, the structure feature library stores a number of structure features, the structure type library stores structure type information, the structure parsing unit traverses the corresponding text information to identify the same structure features through the structure feature library, and matches the closest structure type information from the structure type library according to the order and position relationship of the corresponding structure features in the text information; the knowledge layer is also configured with a data configuration module, the data configuration module includes a feature configuration unit, the feature configuration unit splits each structure type information in the structure type library into a number of structure features, and each structure feature in the structure feature library is configured with a feature priority value, D=a / (A1α1 / β1+K+A n α n / β n ), where D is the feature priority value, a is the preset priority value configuration parameter, and A nis the matching association value of the nth structural type information with the structural feature, and the matching association value reflects the matching reliability of the structural type information. n is the number of all structural features in the nth structural type information with this structural feature, β nIt is the number of structural features in the nth structural type information with this structural feature; first, the content of the structural parsing unit is described in detail. The purpose of the structural parsing unit is to identify the structural features of text information. On the one hand, the structural features will be the same as general texts, such as format, font, size, label, etc. There will also be features unique to the power grid development text, such as graphic features such as tables, charts, distribution maps, or special text format features. The above features are stored in advance through the structural feature library, and then the entire text information is traversed to mark the positions of these features. After the marking is completed, a relationship graph is generated according to the relationship between the features. This relationship graph is generated based on the paragraph relationship. Since the present invention is more complex and has more features than general text structure recognition, the text recognition error rate may also increase. Therefore, the present invention also proposes a verification algorithm. The main purpose of this method is to filter invalid rules. How to define invalid rules? Invalid rules are reflected in some rules that appear in unstructured text content, such as 150 million. This number can be matched by rule 1, but it is followed by numbers and unit content. And some characteristics that non-structural templates should have: For the reverse detection template rules, we designed the following rule structure and algorithm: In order to meet the algorithm, a structure combining MAP and a three-node linked list was designed: 1. Load Map<String,Chains> Structural rule template; 2. Define the stack; 3. Define the first rule variable and initialize it to true firstRule = true; 4. for traverse the current non-structured document rule set; 5. Get the rule object; 6. According to the rule, get the Chains front and back pointers to point to the three-node linked list; 7. If the rule is a numeric value; 8. Check whether the current number is a number in the content; 9. If it is a number in the content 10. Set state to 1 to mark the current rule as invalid; 11. For non-numeric case, judge the first rule; 12. If the front pointer of the first rule is not null; 13. Set state to 1 to mark the current rule as invalid; 14. else set the first rule to false firstRule = false, and add the current rule to the stack; 15. If it is not the first rule, judge; 16. for traverse the stack; 17. If the current rule is the same type as the most recent rule in the stack, it is a forward pointing relationship; 18. Add the current rule to the stack; 19. else set state to 1 to mark the current rule as invalid.The purpose is to eliminate invalid structural features. After marking all structural features, the present invention also configures a type feature database to determine the format type of the text information, further avoiding the inability to determine the format type due to the information features being too complex. The type feature database is also pre-configured. Although the feature elements are relatively complex, the general policy texts have a high similarity in format and the format features are relatively uniform. For example, after marks one and two, there will be digital marks 1 and 2 as sub-items, followed by marks 1.1. This type of mark is more common in a policy text, so the structural type is entered in advance to match the closest structural type, which can further reduce the error rate.

[0034] The data configuration module also includes a type association unit, which configures a matching association value for each structural type information according to the input text sample, and A=1 / [(t0-t1)-cM] 2 χ1+K+1 / [(t0-t m )-cM] 2 χ m , where A is the matching association value, t0 is the current time, t m is the creation time of the mth text sample, c is the preset sensitive adjustment variable, M is the total amount of text samples, χ m The mth text sample corresponds to the known matching degree of the structural type information. Specifically, the structural parsing unit determines the matching order of the structural features in the structural type information library based on the priority value. This configuration allows the matching association value of each structural type information to be calculated based on the way the text sample is imported, reflecting the frequency of occurrence of the structural type information. If a combination of structural features is similar to multiple structural type information, the optimal structural type information can be determined based on the matching association value. The time variable ensures that the model can self-learn and optimize when the text sample format is updated, so that text structural types with a time advantage will correspond to higher matching association values.

[0035] The structure analysis unit calculates the matching degree of each structure type information according to the matching correlation value, and Δχ=A(h1+K+h g ), where Δχ is the matching degree, h gThe structural analysis unit determines the structural type information with the highest matching degree as the closest structural type information based on the similarity of the g-th structural feature. The matching degree can be calculated based on the actual matching results. Since the structural type information is also composed of several structural features, the matching result can be determined through feature matching. Specifically, a similarity relationship is set for the structural features of each structural type information. For example, if the Chinese symbols are exactly the same, the similarity value is 1, while the similarity value of (2,) and (2:) is 0.8, and the similarity value of (2,) and (2) is 0.6. The similarity relationship can be manually configured, so that different structures can be obtained based on the different structural features, and the matching degree of each text information and each structural type information is determined. A reflects the frequency of occurrence of this structural type information, that is, the trust value. The above method can complete the identification of structural type information. In this way, the structural type of the text information can be determined, and the entire text can be divided according to the structural type, so that targeted identification can be carried out. The advantage of this is that, for example, for the recognition of information in the power grid industry, the responsible party may be required under the recognition standard requirements for a sentence. However, since the responsible party has been introduced in the previous text of the paragraph, the corresponding semantics can still be recognized even if the responsible party is missing during semantic recognition. Therefore, pre-division of the text supports semantic recognition. However, the complexity of the text in the power grid industry determines that the above technology alone cannot complete the construction of the knowledge graph.

[0036] The semantic matching unit is associated with an industry terminology library, which stores power grid industry terms and configures part-of-speech features for each power grid industry term. The semantic matching unit tags words corresponding to power grid industry terms in text information using part-of-speech features, and is configured with a semantic recognition algorithm to perform semantic recognition on the marked text information to generate a number of knowledge information elements. The purpose of the industry terminology library is to store industry terms. Because general natural language semantic recognition algorithms cannot be effectively applied to highly professional texts, it is necessary to build an industry terminology library to increase the adaptability of the semantic recognition algorithm. The industry terminology library is pre-loaded with part-of-speech features of corresponding power grid industry terms, and then tags corresponding words based on these part-of-speech features. In this way, semantic recognition can be performed based on part-of-speech features. Specifically, the knowledge layer also includes a data extraction module, which includes a vocabulary extraction unit and a part-of-speech tagging unit. The vocabulary extraction unit is associated with several industry terminology databases in the data layer and extracts power grid industry terms from the industry terminology databases. The part-of-speech tagging unit is used to tag the extracted power grid industry terms with part-of-speech features, and the part-of-speech features include the type of industry terminology database. The industry terminology database is extracted from several industry terminology databases in the data layer, such as the responsible entity vocabulary database, the information system vocabulary database, the power vocabulary database, the transformer industry professional vocabulary database, the electrician common vocabulary database, the power and electrical vocabulary database, the motor and electrical vocabulary database, the motor design vocabulary database, the motor professional terminology database, the power installation vocabulary database, the power system vocabulary database, the power industry equipment vocabulary database, the electrical automation professional vocabulary database, the power station construction vocabulary database, and so on. The power industry professional vocabulary is formed by summarizing the above vocabulary information, but each vocabulary is assigned a corresponding part-of-speech feature, such as noun, equipment name, or transformer-related, or status word, equipment status, or power distribution-related. Each vocabulary can have multiple part-of-speech features, and the part-of-speech features can be used to perform semantic recognition of text sentences. Specifically, the data extraction module also includes a feature configuration unit, which is used to configure the recognition priority value of each part-of-speech feature. The semantic recognition algorithm determines the part-of-speech features as indexes in order of recognition priority, and performs semantic recognition on the text information based on the determined part-of-speech features. When the recognition result meets the first recognition condition, the corresponding knowledge information element is output. The semantic recognition algorithm includes a plurality of structured segments, each of which includes fixed items and parameter items, and each of which is indexed by part-of-speech features. The semantic recognition algorithm includes

[0037] Step A1: selecting part-of-speech features in the order of recognition priority values ​​to determine the corresponding structured segment;

[0038] Step A2: determining fixed items in the target segment and calculating their association degree to generate a first recognition value;

[0039] Step A3: determining a parameter item in the target segment according to the position of the determined fixed item in the target segment, and verifying its data format to generate a second recognition value;

[0040] Step A4: determining the remaining information in the target segment to generate a third recognition value;

[0041] Step A5: summing the first recognition value, the second recognition value, and the third recognition value to obtain a semantic recognition value;

[0042] Step A6: Determine whether the semantic recognition value satisfies the first recognition condition. If so, extract parameter items from the target segment using the structured segment to generate the knowledge information element. If not, re-enter step A1. This configuration allows the meaning of the target segment to be determined in three dimensions. The target segment is determined based on punctuation, which is conventional. The first recognition value of semantic recognition corresponds to a fixed item, which may differ from the semantic recognition standard. The first recognition value can be calculated based on the degree of association of the fixed item. For example, transformers and transformer equipment have corresponding degrees of association. This degree of association can be pre-configured in the power grid industry database. Parameter items, such as specific numerical values, can verify the data format. For example, the number of digits in a data item or the unit following it can determine whether the parameter item exists. The higher the approximation, the higher the reliability. The more residual information, the more information in the target segment, and the larger the deviation. Therefore, the less residual information, the higher the third recognition value. In this way, the semantic recognition algorithm can be used to match the structured segment and complete the recognition. Specifically, the first recognition condition is configured with a first recognition threshold and a second recognition threshold. If the semantic recognition value exceeds the first recognition threshold, or the semantic recognition value of the structured segment is greater than the semantic recognition mean second recognition threshold, the first recognition condition is considered to be met. The semantic recognition mean is the average value of the semantic recognition thresholds obtained by all previously recognized structured segments. Because the present invention is configured with a recognition priority value, generally speaking, the degree of recognition matching in the front will be higher than that in the back. Therefore, if the semantic recognition value feature is high, that is, higher than the preset first recognition threshold, it means that the recognition is successful and no subsequent recognition is required, thereby improving recognition efficiency. If the semantic recognition result is not high, but higher than the previous average and has a certain difference, it means that the target segment is difficult to recognize, but the recognition result at this time should be close to the optimal result. Therefore, the efficiency of parsing and recognition can be improved by the above method. Specifically, the order of part-of-speech priority values ​​is determined. The part-of-speech priority value is the weighted sum of the word priority value and the lexicon priority value. The feature configuration unit includes a word priority algorithm and a lexicon priority algorithm. The word priority algorithm adds a preset first priority increment to the word priority value of a power grid industry term when it is identified in text information, and simultaneously adds a preset second priority increment to the word priority values ​​of other power grid industry terms based on the similarity weight between power grid industry terms. The lexicon priority algorithm adds a preset third priority increment to the lexicon priority values ​​of all power grid industry terms in the same industry terminology database when it is identified in text information. The part-of-speech priority value is calculated based on two dimensions: the frequency of the word appearing in all text information, the frequency of its associated words appearing in text information, and the number of times the corresponding word appears in the corresponding text information in the underlying database.This allows the grid industry database to be optimized by exporting large quantities of external databases, without the imported vocabulary causing data redundancy.

[0043] The information association unit includes a first association strategy and a second association strategy. The first association strategy establishes a first association mark between knowledge information elements based on structural type information. The second association strategy selects knowledge information elements with association characteristics from the knowledge information elements based on the first association mark, compares knowledge information elements with association characteristics of different text information to determine the text association relationship between the text information, establishes a second association mark between knowledge information elements of different subordinate text information based on the text association relationship, and forms the text element map based on the first association mark, the second association mark, and the knowledge information elements. The first association strategy includes determining the subordinate relationship between knowledge information elements based on structural type information, identifying missing elements of knowledge information elements, and generating a first association mark based on the identified missing elements and the subordinate relationship. Because there are missing items in the knowledge information elements, such as the responsible party mentioned in the title, which may be represented by a pseudonym later, this pseudonym will be presented as a missing element. This missing element can be completed into complete information through the first association mark. The second association strategy includes determining the knowledge information element as basic information based on the first association mark, matching the knowledge information elements of different text information to determine the element competing items among the knowledge information elements whose matching results meet the second comparison conditions, retrieving the corresponding knowledge information element as basic information based on the first association mark to identify the element competing items, and generating corresponding competing conditions based on the element competing items. The element competing items include time competing items, policy level competing items, and regulation priority competing items. Since the first association mark can determine the basic information, and based on the basic information or other knowledge information elements, two knowledge information elements with a competing relationship can be determined, such as policies at different times or policies at different levels, or information clearly indicating that there is a dispute about the adoption of this solution, the relationship between these knowledge information elements can be used to determine the dispute relationship, forming a corresponding knowledge graph, and obtaining accurate answers when asking questions. Based on the above, for example, keyword analysis, time analysis, and year analysis can be constructed.

[0044] Of course, the above are only typical examples of the present invention. In addition, the present invention may also have many other specific implementation methods. Any technical solutions formed by equivalent replacement or equivalent transformation fall within the scope of protection required by the present invention.

Claims

1. An intelligent question-answering system for power grid planning based on text parsing, comprising an application layer, a support layer, a knowledge layer, and a data layer, characterized by: The knowledge layer is configured with a text parsing module, which is used to parse the text information input by the data layer and generate a text element map; The text parsing module includes a structure parsing unit, a semantic matching unit, and an information association unit; the structure parsing unit is associated with a structure feature library and a structure type library, the structure feature library stores a plurality of structure features, and the structure type library stores structure type information; the structure parsing unit traverses the corresponding text information to identify the same structure features through the structure feature library, and matches the closest structure type information from the structure type library based on the order and position relationship of the corresponding structure features in the text information; The semantic matching unit is associated with an industry terminology library, which stores power grid industry terms and configures a part-of-speech feature for each power grid industry term. The semantic matching unit marks words corresponding to power grid industry terms in text information using the part-of-speech features, and is configured with a semantic recognition algorithm to perform semantic recognition on the marked text information to generate a plurality of knowledge information elements. The information association unit includes a first association strategy and a second association strategy, wherein the first association strategy establishes a first association mark between knowledge information elements according to structure type information, and the second association strategy selects knowledge information elements with association characteristics from the knowledge information elements according to the first association mark, compares the knowledge information elements with association characteristics of different text information to determine the text association relationship between the text information, establishes a second association mark between the knowledge information elements of different subordinate text information according to the text association relationship, and forms the text element map according to the first association mark, the second association mark, and the knowledge information elements; The knowledge layer also includes a data extraction module, which includes a vocabulary extraction unit and a part-of-speech tagging unit. The vocabulary extraction unit is associated with several industry terminology databases of the data layer and extracts power grid industry terms from the industry terminology databases. The part-of-speech tagging unit is used to mark the extracted power grid industry terms with part-of-speech features, and the part-of-speech features include the type of the industry terminology database; The data extraction module further includes a feature configuration unit, the feature configuration unit being used to configure the recognition priority value of each part-of-speech feature; The semantic recognition algorithm determines the part-of-speech features as indexes in order of recognition priority values, and performs semantic recognition on text information based on the determined part-of-speech features. When the recognition result meets the first recognition condition, the corresponding knowledge information element is output.

2. The intelligent question-answering system for power grid planning based on text parsing according to claim 1, characterized in that: The knowledge layer is also configured with a data configuration module, which includes a feature configuration unit. The feature configuration unit splits each structure type information in the structure type library into several structure features. Each structure feature in the structure feature library is configured with a feature priority value. ,in is the feature priority value, Configure parameters for preset priority values, is the matching association value of the nth structure type information having the structural feature, and the matching association value reflects the matching reliability of the structure type information. is the number of all structural features in the nth structural type information with this structural feature, is the number of the structural feature in the nth structural type information having the structural feature; The structure parsing unit determines the matching order of the structural features in the structure type information library according to the priority value.

3. The intelligent question-answering system for power grid planning based on text parsing according to claim 2, characterized in that: The data configuration module also includes a type association unit, which configures the matching association value of each structure type information according to the input text sample. ,in, To match the associated value, is the current time, is the creation time of the mth text sample, is the preset sensitive adjustment variable, is the total amount of text samples, is the known matching degree between the mth text sample and the structural type information; The structure analysis unit calculates the matching degree of each structure type information according to the matching association value. ,in For matching, is the similarity of the g-th structural feature, and the structure analysis unit determines the structural type information with the highest matching degree as the closest structural type information.

4. The intelligent question-answering system for power grid planning based on text parsing according to claim 3, characterized in that: The semantic recognition algorithm includes a plurality of structured segments, each of which includes fixed items and parameter items, and each of which is indexed by part-of-speech features. The semantic recognition algorithm includes Step A1: selecting part-of-speech features in the order of recognition priority values ​​to determine the corresponding structured segment; Step A2: determining fixed items in the target segment and calculating their association degree to generate a first recognition value; Step A3: determining a parameter item in the target segment according to the position of the determined fixed item in the target segment, and verifying its data format to generate a second recognition value; Step A4: determining the remaining information in the target segment to generate a third recognition value; Step A5: summing the first recognition value, the second recognition value, and the third recognition value to obtain a semantic recognition value; Step A6: Determine whether the semantic recognition value meets the first recognition condition. If so, extract parameter items from the target segment using the structured segment to generate the knowledge information element. If not, re-enter step A1.

5. The intelligent question-answering system for power grid planning based on text parsing according to claim 4, characterized in that: The first recognition condition is configured with a first recognition threshold and a second recognition threshold. If the semantic recognition value exceeds the first recognition threshold, or the semantic recognition value of the structured segment is greater than the semantic recognition mean second recognition threshold, the first recognition condition is deemed to be met. The semantic recognition mean is the average value of the semantic recognition thresholds obtained for all previously recognized structured segments.

6. The intelligent question-answering system for power grid planning based on text parsing according to claim 5, characterized in that: The part-of-speech priority value is the weighted sum of the word priority value and the lexicon priority value. The feature configuration unit includes a word priority algorithm and a lexicon priority algorithm. The word priority algorithm is to add a preset first priority increment to the word priority value of a power grid industry term when it is identified in text information, and at the same time add a preset second priority increment to the word priority values ​​of other power grid industry terms according to the similarity weight between the power grid industry terms; the lexicon priority algorithm is to add a preset third priority increment to the lexicon priority values ​​of all power grid industry terms that belong to the same industry terminology database as the power grid industry term when it is identified in text information.

7. The intelligent question-answering system for power grid planning based on text parsing according to claim 1, characterized in that: The first association strategy includes determining the subordinate relationships between knowledge information elements according to the structure type information, identifying missing elements of the knowledge information elements, and generating a first association mark according to the identified missing elements and the subordinate relationships.

8. The intelligent question-answering system for power grid planning based on text parsing according to claim 1, characterized in that: The second association strategy includes determining the knowledge information elements as basic information based on the first association mark, and matching the knowledge information elements of different text information to determine the element competing items among the knowledge information elements whose matching results meet the second comparison conditions, retrieving the corresponding knowledge information elements as basic information based on the first association mark to identify the element competing items, and generating corresponding competing conditions based on the element competing items, wherein the element competing items include time competing items, policy level competing items, and regulation priority competing items.

Images