Power transformation project review data information extraction method based on AI

Abstract

The invention relates to the technical field of information extraction, in particular to an AI-based power transformation project review data information extraction method. The method comprises the following steps: firstly, acquiring a plurality of review tables of the power transformation project, extracting numeric characters and alphabetic characters from content information of each cell of the review tables, and determining the number of the cells of the review tables according to the difference of the number of the characters among the cells of the review tables and the difference of the number of the numeric characters and the number of the alphabetic characters of all the cells; the method comprises the steps of obtaining the importance degree of a review table according to the review data, screening out a review opinion table from all review tables, screening out a column where a review date is located, a column where a review subject is located and a column where review opinions are located from all columns of the review opinion table, extracting key review information of the review opinion table, inputting the key review information into an AI model, and generating a review abstract. According to the method, the key information in the review data can be accurately extracted, and the effect of extracting the power transformation project review data information is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of information extraction, and in particular to an AI-based method for extracting information from substation engineering review materials. Background Art

[0002] The construction and management of substation projects is a complex task, involving a large number of design, construction, operation and maintenance, monitoring and other links, and a comprehensive review is required for each link. Therefore, it is necessary to extract information from the review materials of the substation project to ensure the quality and safety of each link and the entire substation project construction.

[0003] Since the review materials of substation projects are usually presented in the form of tables, in related technologies, the contents of the entire review form are usually scanned to achieve the extraction of review material information. However, since the review materials of substation projects usually involve a large number of different types of tables, the relevance between the information contained in different tables and the review information is different, and the relevance between the information contained in different parts of the same table and the review information is also different, the existing methods are unable to accurately extract the key information in the review materials, thereby reducing the effect of extracting information from substation project review materials. Summary of the invention

[0004] In order to solve the technical problem that the existing methods cannot accurately extract key information from the review materials and reduce the effect of extracting information from the substation engineering review materials, the purpose of the present invention is to provide an AI-based substation engineering review materials information extraction method, and the technical scheme adopted is as follows:

[0005] The present invention proposes an AI-based method for extracting information from substation engineering review materials, the method comprising:

[0006] Acquire a plurality of review forms on the substation project, wherein each cell of the review form contains content information;

[0007] Taking any review form as the target review form, extracting characters from the content information in each cell of the target review form to obtain characters in each cell, wherein the characters include numeric characters and text characters; obtaining the importance of the target review form according to the difference in the number of characters between cells in the target review form and the difference in the number of numeric characters and text characters in all cells; and screening out the review opinion form from all review forms based on the importance;

[0008] Take any review opinion table as the target review opinion table, take any column of cells in the target review opinion table as the target column, and obtain the review date confidence of the target column according to the difference in the number of characters between the cells of the target column and the number of cells whose content information in the target column contains elements in a preset first keyword set; perform keyword extraction on the content information in each cell of the target column to obtain the keywords in each cell of the target column and the word frequency of each keyword; obtain the review subject confidence of the target column according to the word frequency of the keywords in each cell of the target column; obtain the review opinion confidence of the target column according to the difference in elements between the set composed of all the keywords in each cell of the target column and a preset second keyword set; extract key information from the target review opinion table based on the review date confidence, the review subject confidence and the review opinion confidence to obtain key review information of the target review opinion table;

[0009] Based on the AI ​​model, the key review information of each review opinion form is processed to obtain a review summary.

[0010] Furthermore, the importance of obtaining the target review form includes:

[0011] In the target review table, the product value of the number of the text characters in each cell and the preset first bit number is used as the text content length of each cell, the product value of the number of the digital characters in each cell and the preset second bit number is used as the digital content length of each cell, and the sum of the text content length and the digital content length is used as the actual content length of each cell;

[0012] Take any two cells in the target review table as a reference cell group, and obtain the content length difference value of each reference cell group according to the difference in the actual content length between the two cells in each reference cell group; take the average value of the content length difference values ​​of all reference cell groups as the first important parameter of the target review table;

[0013] The sum of the number of text characters in all cells is used as the numerator, the sum of the number of characters in all cells is used as the denominator, and the ratio is used as the second important parameter of the target review table;

[0014] The first important parameter and the second important parameter are integrated and normalized to obtain the importance of the target review form.

[0015] Furthermore, based on the importance, the review opinion form selected from all review forms includes:

[0016] The review forms whose importance is greater than the preset importance threshold are used as review opinion forms.

[0017] Furthermore, obtaining the review date confidence of the target column includes:

[0018] Take any two cells in the target column as a cell group to be tested, and based on the calculation method of the content length difference value of each reference cell group, obtain the content length difference value of each cell group to be tested according to the difference of the real content length between the two cells in each cell group to be tested;

[0019] The cumulative values ​​of the content length differences of all the tested cell groups are negatively correlated to obtain the possibility of the review date of the target column;

[0020] Using the elements in the preset first keyword set, keyword matching is performed on the content information in each cell of the target column, and the cell whose content information contains the elements in the preset first keyword set is used as the review date cell of the target column;

[0021] The review date possibility of the target column and the number of the review date cells are integrated and normalized to obtain the review date confidence of the target column.

[0022] Furthermore, obtaining the review subject confidence of the target column includes:

[0023] Taking the average of the word frequencies of all the keywords in each cell of the target column as the overall word frequency of each cell of the target column;

[0024] The accumulated values ​​of the overall word frequencies of all cells of the target column are normalized to obtain the review subject confidence of the target column.

[0025] Furthermore, obtaining the confidence of the review opinions of the target column includes:

[0026] The set consisting of all the keywords in each cell of the target column is used as the keyword set of each cell of the target column;

[0027] The Jaccard coefficient between the keyword set of each cell of the target column and the preset second keyword set is used as the possibility of the review opinion of each cell of the target column;

[0028] The accumulated values ​​of the likelihood of the review opinions of all cells in the target column are normalized to obtain the confidence of the review opinions of the target column.

[0029] Furthermore, the key review information of the target review opinion form includes:

[0030] In the target review opinion table, the column corresponding to the maximum value of the review date confidence is used as the review date column, the column corresponding to the maximum value of the review subject confidence is used as the review subject column, and the column corresponding to the maximum value of the review opinion confidence is used as the review opinion column;

[0031] The content information of all cells in the review date column, the review subject column and the review opinion column in the target review opinion table is used as the key review information of the target review opinion table.

[0032] Further, obtaining the review summary includes:

[0033] The key review information of all review opinion forms is input into the trained AI language model, and the AI ​​language model generates a review summary.

[0034] Further, obtaining the characters in each cell includes:

[0035] The content information in each cell of the target review table is processed using a regular expression matching method to obtain the characters in each cell.

[0036] Furthermore, obtaining the keyword in each cell of the target column and the word frequency of each keyword includes:

[0037] Use the TF-IDF algorithm to process the content information in each cell of the target column to obtain the keywords in each cell of the target column and the word frequency of each keyword.

[0038] The present invention has the following beneficial effects:

[0039] The present invention takes into account that the existing methods cannot accurately extract key information from the review materials, which reduces the effect of extracting information from the review materials of the substation project. Therefore, multiple review forms of the substation project are first obtained. Since there are many types of review forms, the information contained in different types of review forms has different connections with the review information. Among the many review forms, the information contained in the review opinion form has the strongest connection with the review information. Compared with other types of review forms, the content information in the cells of the review opinion form is more in the form of text and less in the form of numbers, and the length of the content information between each cell varies greatly. Therefore, the characters in each cell can be extracted first to provide a data basis for the subsequent importance analysis of the review form, and the importance of the target review form is reflected by the obtained importance, thereby screening the review opinions. See the table, so as to screen out the review forms with weak connection with the review information, so as to facilitate the subsequent extraction of key review information from the review opinion table. Considering that there are usually multiple different types of column fields in the review opinion table, each cell in each column contains the same type of content information, and the information contained in different columns has different connections with the review information. Among the many column fields, the review date, review subject and review opinion are the three most important columns in the review opinion table. Therefore, the target column can be analyzed, and the key information of the target review opinion table can be extracted through the review date confidence, review subject confidence and review opinion confidence. The key review information of the target review opinion table can be accurately extracted, and the review summary can be generated through the AI ​​model, so as to further refine the core content of the key review information and improve the effect of extracting information from substation engineering review materials. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0041] Figure 1 A flowchart of a method for extracting information from substation engineering review materials based on AI provided by one embodiment of the present invention;

[0042] Figure 2 A flow chart of a method for obtaining the importance of a target review form provided in one embodiment of the present invention. DETAILED DESCRIPTION

[0043] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the following is a detailed description of the specific implementation method, structure, features and effects of a method for extracting information from substation engineering review materials based on AI proposed by the present invention in combination with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics in one or more embodiments may be combined in any suitable form.

[0044] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0045] The following is a detailed description of a specific scheme of an AI-based substation engineering review document information extraction method provided by the present invention in conjunction with the accompanying drawings.

[0046] See also Figure 1 , which shows a flow chart of a method for extracting information from substation engineering review materials based on AI provided by an embodiment of the present invention, the method comprising:

[0047] Step S1: Acquire multiple review forms for substation engineering, each cell of the review form containing content information.

[0048] The review materials of substation projects usually include multiple aspects, such as basic project information, equipment specifications and parameters, review opinions, feasibility reports, risk assessments, etc. The main function of the review materials is to review the project as a whole to ensure the feasibility and compliance of the project. Generally speaking, the original review materials of substation projects are usually in paper form and recorded in a table. Therefore, the embodiment of the present invention first uses OCR technology to scan the paper review materials and convert the paper review materials into electronic versions, thereby obtaining multiple review forms for substation projects, wherein there are multiple cells in the review form, each cell contains relevant content information, and the content information is recorded in text form. For example, the cell contains content information about the review date or content information about the review unit, etc. It should be noted that OCR technology is a technical means well known to technicians in this field and will not be elaborated here.

[0049] So far, several review forms for substation projects have been obtained.

[0050] Step S2: Take any review form as the target review form, extract characters from the content information in each cell of the target review form, and obtain the characters in each cell, which include numeric characters and text characters; obtain the importance of the target review form based on the difference in the number of characters between cells in the target review form, and the difference in the number of numeric characters and text characters in all cells; based on the importance, filter out the review opinion form from all review forms.

[0051] Since there are many types of review forms such as budget cost tables, technical parameter tables or review opinion tables, the information contained in different types of review forms has different connections with the review information. Among the many review forms, the information contained in the review opinion table has the strongest connection with the review information. In other words, the review opinion table contains key information for the substation project review. Therefore, it is necessary to screen out the review opinion table from the many types of review forms in order to extract the key review information later. Therefore, firstly analyze any review form and take any review form as the target review form. Considering that compared with other types of review forms, the content information in the cells of the review opinion table is more in text form and less in digital form, and the length of the content information between cells varies greatly. Therefore, characters can be extracted from each cell of the target review form first, where the characters in each cell include text characters and digital characters, providing a data basis for the subsequent accurate analysis of the importance of the target review form.

[0052] Preferably, in one embodiment of the present invention, the method for obtaining the characters in each cell specifically includes:

[0053] Since the regular expression matching method can not only extract characters in the text, but also distinguish between numeric characters and text characters, the regular expression matching method can be used to process the content information in each cell of the target review table to obtain the characters in each cell. In other embodiments of the present invention, other methods can also be used to extract characters in the cells, which is not limited here.

[0054] From the above analysis, it can be seen that the length of the content information between the cells of the review opinion form varies greatly, and the content information in the cells is mostly in text form and less in digital form. Therefore, the difference in the number of characters between the cells in the target review form, as well as the difference in the number of numeric characters and text characters in all cells can be analyzed. The importance of the target review form can be reflected by the obtained importance. The greater the importance, the more likely the target review form is to be a review opinion form. Subsequently, the review opinion form can be accurately screened out based on the importance.

[0055] Preferably, in one embodiment of the present invention, the method for obtaining the importance of the target review form specifically includes:

[0056] See also Figure 2 , which shows a flow chart of a method for obtaining the importance of a target review form provided by an embodiment of the present invention.

[0057] Step S201: In the target review table, the product value of the number of text characters in each cell and the preset first bit number is used as the text content length of each cell, the product value of the number of numeric characters in each cell and the preset second bit number is used as the numeric content length of each cell, and the sum of the text content length and the numeric content length is used as the actual content length of each cell.

[0058] First, it is necessary to analyze the length of the content information in each cell of the target review form. Since each cell may contain both text characters and numeric characters, and the number of bits occupied by text characters and numeric characters in the document is usually different, it is necessary to analyze the text characters and numeric characters separately. First, the product value of the number of text characters in each cell and the preset first bit number is used as the text content length of each cell, where the preset first bit number represents the number of bits occupied by a text character, and the text content length can be considered as the number of bits occupied by all text characters in each cell. Then, the product value of the number of numeric characters in each cell and the preset second bit number is used as the numeric content length of each cell, where the preset second bit number represents the number of bits occupied by a numeric character, and the numeric content length can be considered as the number of bits occupied by all numeric characters in each cell, and then the sum of the text content length and the numeric content length is used as the actual content length of each cell.

[0059] It should be noted that the preset first bit number and the preset second bit number are related to the document encoding method, and the values ​​of the two are different in different document encoding methods. In one embodiment of the present invention, the GBK encoding method is used to encode the review form. Under the GBK encoding method, one character occupies 16 bits and one number occupies 8 bits. At this time, the preset first bit number is 16 and the preset second bit number is 8. Other encoding methods may also be used in other embodiments of the present invention, which are not limited here.

[0060] As an example, in one embodiment of the present invention, the expression of the actual content length of each cell of the target review table may be specifically, for example, as follows:

[0061] L i =S (i，1) ×b1+S (i，2) ×b2

[0062] Among them, L i Indicates the actual content length of the i-th cell in the target review table; S (i，1) represents the number of characters in the ith cell; b1 represents the preset first bit; S (i，2) represents the number of numeric characters in the ith cell; b2 represents the preset second bit.

[0063] Step S202: Take any two cells in the target review table as a reference cell group, and obtain the content length difference value of each reference cell group based on the difference in actual content length between the two cells in each reference cell group; take the average value of the content length difference values ​​of all reference cell groups as the first important parameter of the target review table.

[0064] Then take any two cells in the target review form as a reference cell group, and analyze the difference in the actual content length between the two cells in each reference cell group to obtain the content length difference value of each reference cell group. The larger the content length difference value, the greater the difference in the length of the content information between the two cells in the reference cell group. Then, the average value of the content length difference values ​​of all reference cell groups is taken as the first important parameter of the target review form. The larger the first important parameter, the greater the difference in the length of the content information between the cells of the target review form, and then it is more likely that the target review form is a review opinion form.

[0065] In one embodiment of the present invention, the absolute value of the difference between the actual content lengths of two cells in each reference cell group may be used as the content length difference value of each reference cell group to analyze the difference between the two.

[0066] Step S203: The sum of the number of text characters in all cells is used as the numerator, the sum of the number of characters in all cells is used as the denominator, and the ratio is used as the second important parameter of the target review table.

[0067] Since the cells of the review opinion form contain more text characters and fewer numeric characters, the sum of the number of text characters in all cells can be used as the numerator, the sum of the number of characters in all cells can be used as the denominator, and the ratio can be used as the second important parameter of the target review form. The larger the second important parameter is, the greater the proportion of text characters in the target review form, which means that the target review form is more likely to be a review opinion form.

[0068] Step S204: The first important parameter and the second important parameter are integrated and normalized to obtain the importance of the target review form.

[0069] The larger the first important parameter and the second important parameter are, the more likely the target review form is to be a review opinion form. Therefore, the first important parameter and the second important parameter can be combined and normalized to obtain the importance of the target review form.

[0070] In the embodiment of the present invention, the integration of the first important parameter and the second important parameter can be achieved by calculating the sum or product of the two parameters, which is not limited here.

[0071] In one embodiment of the present invention, the normalization processing can be specifically, for example, maximum and minimum value normalization processing, and the normalization in subsequent steps can all adopt maximum and minimum value normalization processing. In other embodiments of the present invention, other normalization methods can be selected according to the specific range of numerical values, which will not be described in detail.

[0072] As an example, in one embodiment of the present invention, the expression of the importance of the target review table may be specifically, for example:

[0073]

[0074] Among them, P represents the importance of the target review form; L n and L′ n Respectively represent the actual content lengths of the two cells in the nth reference cell group; |L n -L′ n | represents the content length difference value of the nth reference cell group; N represents the number of reference cell groups; The first important parameter of the target review form; S (i，1) Indicates the number of text characters in the i-th cell; S (i，2) represents the number of numeric characters in the ith cell; I represents the number of cells in the target review table; Represents the sum of the number of text characters in all cells; Represents the sum of the number of characters in all cells; It represents the second most important parameter of the target review table; norm() represents the normalization function, which is used to limit the value range of the importance level P to [0, 1].

[0075] The importance of each review form can be obtained through the same method as above. The greater the importance, the more critical the content information contained in the review form, and the more likely it is to be a review opinion form. Therefore, based on the importance, the review opinion form can be screened out from all review forms. Subsequently, the key review information can be extracted from the review opinion form to improve the accuracy of key information extraction of substation project review materials.

[0076] Preferably, in one embodiment of the present invention, the review form with an importance greater than a preset importance threshold is used as a review opinion form, wherein the preset importance threshold is set to 0.85, and the specific value of the preset importance threshold can also be set by the implementer according to the specific implementation scenario, which is not limited here.

[0077] At this point, the review opinion forms have been screened out.

[0078] Step S3: Take any review opinion table as the target review opinion table, and any column of cells in the target review opinion table as the target column, and obtain the review date confidence of the target column based on the difference in the number of characters between the cells of the target column and the number of cells whose content information in the target column contains elements in the preset first keyword set; perform keyword extraction on the content information in each cell of the target column to obtain the keywords in each cell of the target column and the word frequency of each keyword; obtain the review subject confidence of the target column based on the word frequency of the keywords in each cell of the target column; obtain the review opinion confidence of the target column based on the difference between the elements of the set composed of all keywords in each cell of the target column and the preset second keyword set; extract key information from the target review opinion table based on the review date confidence, review subject confidence and review opinion confidence to obtain key review information of the target review opinion table.

[0079] Considering that there are usually multiple different types of column fields in the review opinion table, such as review subject, review date, review opinion and review unit, etc., each cell in each column contains the same type of content information. For example, for the column where the review date is located, the content information in the column cell is usually date information, including the year, month and day information corresponding to the review, and the information contained in different columns has different connections with the review information. Among the many column fields, the review date reflects the time of the review, which is related to the real-time nature of the review work. The review subject reflects the object or project of the review, and the review opinion reflects the problems of the review object and suggestions for solving the problems. Therefore, the review date, review subject and review opinion are the three most important columns in the review opinion table, and the information contained in the three is the most critical. Therefore, in the subsequent steps, it is necessary to filter out the review date, review subject and review opinion columns from the review opinion table to further realize the extraction of key information in the review opinion table.

[0080] The embodiment of the present invention first takes any review opinion form as the target review opinion form, and analyzes any column of the target review opinion form, and takes any column of cells in the target review opinion form as the target column. For the column where the review date is located, since the content format of the review date is relatively fixed, the length of the content information of each cell in the column where the review date is located is relatively consistent, and the content information of each cell in the column where the review date is located usually contains keywords such as "year", "month", and "day". Therefore, the difference in the number of characters between the cells of the target column and the number of cells in the target column whose content information contains elements in the preset first keyword set can be analyzed, and the possibility that the target column is the column where the review date is located is reflected by the obtained review date confidence. Subsequently, based on the review date confidence, the review date column of the target review opinion form can be determined, wherein the preset first keyword set is {"year", "month", "day"}, and the preset first keyword set can also be set by the implementer according to the specific implementation scenario, which is not limited here.

[0081] Preferably, in one embodiment of the present invention, the method for obtaining the review date confidence of the target column specifically includes:

[0082] Take any two cells in the target column as a cell group to be tested, and based on the calculation method of the content length difference value of each reference cell group, obtain the content length difference value of each cell group to be tested according to the difference in the actual content length between the two cells in each cell group to be tested. The smaller the content length difference value of each cell group to be tested is, the more consistent the lengths of the content information of the two cells in the cell group to be tested are. Therefore, the accumulated values ​​of the content length difference values ​​of all cell groups to be tested can be negatively correlated to obtain the possibility of the review date of the target column. The greater the possibility of the review date is, the more consistent the lengths of the content information of each cell in the target column are, and thus the target column is more likely to be the column where the review date is located.

[0083] Using the elements in the preset first keyword set, keyword matching is performed on the content information in each cell of the target column, and the cells whose content information contains the elements in the preset first keyword set are used as the review date cells of the target column. The more review date cells in the target column, the more likely the target column is to be the column where the review date is located. Among them, keyword matching is a technical means well known to those skilled in the art and will not be elaborated here.

[0084] Then, the possibility of the review date of the target column and the number of review date cells can be integrated and normalized to obtain the confidence of the review date of the target column.

[0085] In the embodiment of the present invention, the integration of the review date possibility and the number of review date cells in the target column can be achieved by calculating the sum or product of the two, which is not limited here.

[0086] As an example, in one embodiment of the present invention, the expression of the review date confidence of the target column may be specifically, for example, as follows:

[0087]

[0088] Where D represents the confidence level of the review date of the target column; L m and L′ m Respectively represent the actual content length of the two cells in the mth cell group to be tested; |L m -L′ m | represents the content length difference value of the mth cell group to be tested; M represents the number of cell groups to be tested; Represents the possibility of the review date of the target column; Z represents the number of review date cells in the target column; norm() represents a normalization function, which is used to limit the value range of the review date confidence D to [0, 1]; ε represents a preset adjustment parameter, which is used to prevent the denominator from being 0, and the value range of ε is [0.001, 0.01]. In one embodiment of the present invention, ε is set to 0.01. The specific value of ε can also be set by the implementer according to the specific implementation scenario, which is not limited here.

[0089] It should be noted that in other embodiments of the present invention, negative correlation mapping may be achieved through other basic mathematical operations, which will not be elaborated herein.

[0090] The same method as above can be used to obtain the review date confidence of each column in the target review opinion table.

[0091] For the column where the review subject is located, the content information in the cells of this column reflects the object or project being reviewed. The keywords in its content information appear frequently in the entire review opinion table. Therefore, we can first extract keywords from the content information in each cell of the target column to obtain the keywords in each cell of the target column and the word frequency of each keyword, providing a data basis for subsequent calculation and analysis of the possibility that the target column is the column where the review subject is located.

[0092] The TF-IDF algorithm can extract keywords from a document and calculate the frequency of each keyword. Therefore, in one embodiment of the present invention, the TF-IDF algorithm can be used to process the content information in each cell of the target column to obtain the keywords in each cell of the target column and the frequency of each keyword. The TF-IDF algorithm is a technical means well known to those skilled in the art and will not be described in detail here.

[0093] The greater the frequency of the keywords in each cell of the target column, the higher the frequency of the keywords in the content information of each cell appearing in the target review opinion table, which further indicates that the target column is likely to be the column where the review subject is located. Therefore, based on the frequency of the keywords in each cell of the target column, the review subject confidence of the target column can be obtained, and subsequently, based on the review subject confidence, the review subject column of the target review opinion table can be determined.

[0094] Preferably, in one embodiment of the present invention, the method for obtaining the confidence of the review subject of the target column specifically includes:

[0095] The average value of the word frequency of all keywords in each cell of the target column is taken as the overall word frequency of each cell of the target column; the accumulated value of the overall word frequency of all cells of the target column is normalized to obtain the review subject confidence of the target column.

[0096] As an example, in one embodiment of the present invention, the expression of the review subject confidence of the target column may be specifically, for example, as follows:

[0097]

[0098] Among them, E represents the review subject confidence of the target column; represents the overall word frequency of the h-th cell in the target column; H represents the number of cells in the target column; f (h，t) represents the frequency of the tth keyword in the hth cell of the target column; T h represents the number of keywords in the hth cell of the target column; norm() represents a normalization function, which is used to limit the value range of the review subject confidence E to [0, 1].

[0099] The same method as above can be used to obtain the review subject confidence in each column of the target review opinion table.

[0100] For the column where the review opinions are located, some keywords about problems and suggestions often appear in the content information of the cells in this column, such as "defect", "problem", "optimization" and "improvement", etc. Therefore, the differences in the elements between the set composed of all keywords in each cell of the target column and the preset second keyword set can be analyzed, and the obtained review opinion confidence can reflect the possibility that the target column is the column where the review opinions are located. Subsequently, based on the review opinion confidence, the column where the review opinions of the target review opinion table are located can be determined, where the preset second keyword set is {"defect", "problem", "suggestion", "optimization", "improvement", "adjustment", "unreasonable"}. The preset second keyword set can also be set by the implementer according to the specific implementation scenario, and is not limited here.

[0101] Preferably, in one embodiment of the present invention, the method for obtaining the confidence level of the review opinion of the target column specifically includes:

[0102] The set consisting of all the keywords in each cell of the target column is used as the keyword set of each cell of the target column. The more similar the elements in the keyword set of each cell are to the elements in the preset second keyword set, the more likely the content information in each cell is related to the review opinions. The Jaccard coefficient can measure the similarity between the two sets. Therefore, the Jaccard coefficient between the keyword set of each cell of the target column and the preset second keyword set can be used as the possibility of the review opinions of each cell of the target column, and then the accumulated values ​​of the possibility of the review opinions of all cells in the target column are normalized to obtain the confidence of the review opinions of the target column. Among them, the Jaccard coefficient is a technical means well known to those skilled in the art and will not be elaborated here.

[0103] As an example, in one embodiment of the present invention, the expression of the review opinion confidence of the target column may be specifically, for example, as follows:

[0104]

[0105] Among them, F represents the confidence of the review opinions of the target column; Jh represents the Jaccard coefficient between the keyword set of the h-th cell of the target column and the preset second keyword set, that is, the possibility of the review opinions of the h-th cell; H represents the number of cells in the target column; norm() represents the normalization function, which is used to limit the value range of the review opinion confidence F to [0, 1].

[0106] The same method as above can be used to obtain the confidence level of the review opinions in each column of the target review opinion table.

[0107] Since the review date column, review subject column and review opinion column are the three most important columns in the target review opinion table, the content information they contain is highly correlated with the review information and the content information they contain is relatively critical. Therefore, based on the review date confidence, review subject confidence and review opinion confidence, key information of the target review opinion table can be extracted to obtain the key review information of the target review opinion table, thereby improving the accuracy of key information extraction from the target review opinion table.

[0108] Preferably, in one embodiment of the present invention, the method for obtaining key review information of the target review opinion form specifically includes:

[0109] In the target review opinion table, the column corresponding to the maximum value of the review date confidence is used as the review date column, the column corresponding to the maximum value of the review subject confidence is used as the review subject column, and the column corresponding to the maximum value of the review opinion confidence is used as the review opinion column. The content information of all cells in the review date column, review subject column, and review opinion column in the target review opinion table is used as the key review information of the target review opinion table.

[0110] The same method as above can be used to extract key information from each review opinion form, thereby obtaining the key review information of each review opinion form.

[0111] Step S4: Based on the AI ​​model, the key review information of each review opinion form is processed to obtain a review summary.

[0112] After obtaining the key review information of each review opinion form, the AI ​​model can be used to process the key review information of each review opinion form to obtain a review summary, thereby further refining the core content of the key review information and improving the effect of extracting review data information for substation projects.

[0113] Preferably, in one embodiment of the present invention, the method for obtaining the review summary specifically includes:

[0114] The key review information of all review opinion forms is input into the trained AI language model, and the AI ​​language model generates a review summary. The AI ​​language model can use the existing GPT or BERT model, which is not limited or elaborated here.

[0115] It should be noted that the sequence of the above embodiments of the present invention is only for description and does not represent the advantages and disadvantages of the embodiments. The processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0116] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from other embodiments.

Claims

1. A method for extracting information from substation engineering review materials based on AI, characterized in that: The method comprises: Acquire a plurality of review forms for the substation project, wherein each cell of the review form contains content information; Taking any review form as the target review form, extracting characters from the content information in each cell of the target review form to obtain characters in each cell, wherein the characters include numeric characters and text characters; obtaining the importance of the target review form according to the difference in the number of characters between cells in the target review form and the difference in the number of numeric characters and text characters in all cells; and screening out the review opinion form from all review forms based on the importance; Take any review opinion table as the target review opinion table, take any column of cells in the target review opinion table as the target column, and obtain the review date confidence of the target column according to the difference in the number of characters between the cells of the target column and the number of cells whose content information in the target column contains elements in a preset first keyword set; perform keyword extraction on the content information in each cell of the target column to obtain the keywords in each cell of the target column and the word frequency of each keyword; obtain the review subject confidence of the target column according to the word frequency of the keywords in each cell of the target column; obtain the review opinion confidence of the target column according to the difference in elements between the set composed of all the keywords in each cell of the target column and a preset second keyword set; extract key information from the target review opinion table based on the review date confidence, the review subject confidence and the review opinion confidence to obtain key review information of the target review opinion table; Based on the AI ​​model, the key review information of each review opinion form is processed to obtain a review summary.

2. According to the AI-based method for extracting information from substation engineering review materials as described in claim 1, it is characterized in that: The importance of obtaining the target review form includes: In the target review table, the product value of the number of the text characters in each cell and the preset first bit number is used as the text content length of each cell, the product value of the number of the digital characters in each cell and the preset second bit number is used as the digital content length of each cell, and the sum of the text content length and the digital content length is used as the actual content length of each cell; Take any two cells in the target review table as a reference cell group, and obtain the content length difference value of each reference cell group according to the difference in the actual content length between the two cells in each reference cell group; take the average value of the content length difference values ​​of all reference cell groups as the first important parameter of the target review table; The sum of the number of text characters in all cells is used as the numerator, the sum of the number of characters in all cells is used as the denominator, and the ratio is used as the second important parameter of the target review table; The first important parameter and the second important parameter are integrated and normalized to obtain the importance of the target review form.

3. According to the AI-based method for extracting information from substation engineering review materials as described in claim 1, it is characterized in that: Based on the importance, the review opinion forms are selected from all review forms, including: The review forms whose importance is greater than the preset importance threshold are used as review opinion forms.

4. According to the AI-based method for extracting information from substation engineering review materials as described in claim 2, it is characterized in that: The step of obtaining the review date confidence of the target column includes: Take any two cells in the target column as a cell group to be tested, and based on the calculation method of the content length difference value of each reference cell group, obtain the content length difference value of each cell group to be tested according to the difference of the real content length between the two cells in each cell group to be tested; The cumulative values ​​of the content length differences of all the tested cell groups are negatively correlated to obtain the possibility of the review date of the target column; Using the elements in the preset first keyword set, keyword matching is performed on the content information in each cell of the target column, and the cell whose content information contains the elements in the preset first keyword set is used as the review date cell of the target column; The review date possibility of the target column and the number of the review date cells are integrated and normalized to obtain the review date confidence of the target column.

5. According to the AI-based method for extracting information from substation engineering review materials as described in claim 1, it is characterized in that: The step of obtaining the review subject confidence of the target column includes: Taking the average of the word frequencies of all the keywords in each cell of the target column as the overall word frequency of each cell of the target column; The accumulated values ​​of the overall word frequencies of all cells of the target column are normalized to obtain the review subject confidence of the target column.

6. The AI-based method for extracting information from substation engineering review materials according to claim 1 is characterized in that: The step of obtaining the confidence level of the review opinions of the target column includes: The set consisting of all the keywords in each cell of the target column is used as the keyword set of each cell of the target column; The Jaccard coefficient between the keyword set of each cell of the target column and the preset second keyword set is used as the possibility of the review opinion of each cell of the target column; The accumulated values ​​of the likelihood of the review opinions of all cells in the target column are normalized to obtain the confidence of the review opinions of the target column.

7. The AI-based method for extracting information from substation engineering review materials according to claim 1 is characterized in that: The key review information of the target review opinion form includes: In the target review opinion table, the column corresponding to the maximum value of the review date confidence is used as the review date column, the column corresponding to the maximum value of the review subject confidence is used as the review subject column, and the column corresponding to the maximum value of the review opinion confidence is used as the review opinion column; The content information of all cells in the review date column, the review subject column and the review opinion column in the target review opinion table is used as the key review information of the target review opinion table.

8. The AI-based method for extracting information from substation engineering review materials according to claim 1 is characterized in that: Obtaining review summaries includes: The key review information of all review opinion forms is input into the trained AI language model, and the AI ​​language model generates a review summary.

9. The AI-based method for extracting information from substation engineering review materials according to claim 1 is characterized in that: The obtaining of the characters in each cell comprises: The content information in each cell of the target review table is processed using a regular expression matching method to obtain the characters in each cell.

10. The AI-based method for extracting information from substation engineering review materials according to claim 1, characterized in that: The step of obtaining the keyword in each cell of the target column and the word frequency of each keyword includes: Use the TF-IDF algorithm to process the content information in each cell of the target column to obtain the keywords in each cell of the target column and the word frequency of each keyword.

Images