A method and device for extracting document numerical indicators based on machine learning
Through the machine learning-based document numerical index extraction method, natural language processing and deep learning technology are used to solve the problem of low efficiency of document numerical index extraction, achieving more efficient and accurate unstructured data processing, and improving data in the financial industry Analytical ability.
Patent Information
- Application Number
- CN202111487097.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-12-06
- Publication Date
- 2024-11-19
- Estimated Expiration
- 2041-12-06
AI Technical Summary
The extraction efficiency of document numerical indicators in the prior art is low and unstructured data processing is difficult, resulting in the financial industry being inefficient when processing large amounts of PDF and WORD type document data.
Using machine learning-based methods, the index extraction model is constructed through natural language processing and deep learning, the document is divided into key paragraphs and non-key paragraphs, and the index feature scoring model is used for screening, combining historical sequence data to calculate confidence, improve the accuracy of extraction rate and stability.
It significantly improves the accuracy and efficiency of document numerical indicator extraction, reduces the workload of business personnel, improves the efficiency of unstructured data utilization, and enhances the focus ability and stability of indicator extraction.
Smart Images

Figure CN114398853B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of data processing technology, and in particular to a method and device for extracting document numerical indicators based on machine learning. Background Art
[0002] With the advent of the data age, the financial industry needs to process a large amount of data that is growing exponentially. In addition to structured data that is easy to use directly, these data also contain alternative data that cannot be used directly, such as a huge amount of PDF and WORD document data. These alternative data are huge in quantity and unstructured, and key information is hidden in lengthy documents. The business has a strong demand for intelligent processing of this type of unstructured data, but due to the difficulty in processing, it is currently mainly processed manually, which is inefficient.
[0003] Unstructured data cannot be directly applied to data analysis. It needs to be structured and extracted to be used effectively. The extraction of unstructured data has always been difficult due to its varied expressions, diverse forms, and complex structures. There are two existing technologies. One is to manually search for keywords directly from the document through a document reading tool, locate a series of hit document locations, and process the text with the naked eye to determine whether it is an indicator that needs to be extracted. If so, continue to extract the numerical value and unit and normalize it. Otherwise, continue to check other locations of the searched document, and even change multiple keywords for searching. The other is to first extract all the text paragraphs from the document, then perform regular or keyword matching, and finally manually process the recalled information. Summary of the invention
[0004] The purpose of the present invention is to provide a method and device for extracting document numerical indicators based on machine learning, so as to solve the problem of low efficiency in extracting document numerical indicators in the prior art.
[0005] To achieve the above object, the present invention provides a method for extracting document numerical indicators based on machine learning, comprising:
[0006] Divide the document to be processed into key paragraphs and non-key paragraphs according to preset rules;
[0007] An index extraction model is constructed by using natural language processing and deep learning, and the key paragraphs and the non-key paragraphs are respectively input into the index extraction model, and the index extraction model outputs key paragraph indexes and non-key paragraph indexes respectively;
[0008] The key paragraph index and the non-key paragraph index are integrated, and a preset index feature scoring model is input to screen the key paragraph index and the non-key paragraph index, and an index extraction result is output.
[0009] Preferably, the method for extracting document numerical indicators based on machine learning further includes:
[0010] The historical sequence data is input into the preset prediction model to obtain the prediction result, and the average of the prediction result and the screened indicator result is calculated as the confidence for sorting, including the indicator name, indicator value, indicator unit, indicator confidence, short sentences and paragraph texts that hit the indicator.
[0011] Preferably, dividing the document to be processed into key paragraphs and non-key paragraphs according to a preset rule includes:
[0012] Determine the format of the document to be input, convert the PDF document format into word format, and generate the input document;
[0013] Parsing the input document to obtain a set of text and table elements;
[0014] A rule matching method is adopted to locate the key paragraph according to the text and the table element set, and the key paragraph is output.
[0015] Preferably, the key paragraph index and the non-key paragraph index are integrated, and a preset index feature scoring model is input to screen the key paragraph index and the non-key paragraph index, and the index extraction result is output, including:
[0016] Mark the indicators in the key paragraph indicators and the non-key paragraph indicators, sort and normalize the indicator features, and input them into the logistic regression model for training to obtain the preset indicator feature scoring model;
[0017] The key paragraph index and the non-key paragraph index are integrated, input into the preset index feature scoring model, the index below the first threshold is filtered, and the index extraction result is output.
[0018] The present invention also provides a device for extracting document numerical indicators based on machine learning, comprising:
[0019] A division module is used to divide the document to be processed into key paragraphs and non-key paragraphs according to preset rules;
[0020] An extraction module, used for constructing an index extraction model by using natural language processing and deep learning, and inputting the key paragraphs and the non-key paragraphs into the index extraction model respectively, and the index extraction model outputs the key paragraph index and the non-key paragraph index respectively;
[0021] The screening module is used to integrate the key paragraph index and the non-key paragraph index, input a preset index feature scoring model to screen the key paragraph index and the non-key paragraph index, and output an index extraction result.
[0022] Preferably, it also includes a calculation module, wherein the calculation module is used for:
[0023] The historical sequence data is input into the preset prediction model to obtain the prediction result, and the average of the prediction result and the screened indicator result is calculated as the confidence for sorting, including the indicator name, indicator value, indicator unit, indicator confidence, short sentences and paragraph texts that hit the indicator.
[0024] Preferably, the extraction module is further used for:
[0025] Determine the format of the document to be input, convert the PDF document format into word format, and generate the input document;
[0026] Parsing the input document to obtain a set of text and table elements;
[0027] A rule matching method is adopted to locate the key paragraph according to the text and the table element set, and the key paragraph is output.
[0028] Preferably, the screening module is further used for:
[0029] Mark the indicators in the key paragraph indicators and the non-key paragraph indicators, sort and normalize the indicator features, and input them into the logistic regression model for training to obtain the preset indicator feature scoring model;
[0030] The key paragraph index and the non-key paragraph index are integrated, input into the preset index feature scoring model, the index below the first threshold is filtered, and the index extraction result is output.
[0031] The present invention also provides a terminal device, comprising:
[0032] one or more processors;
[0033] A memory, coupled to the processor, for storing one or more programs;
[0034] When the one or more programs are executed by the one or more processors, the one or more processors implement the method for extracting document numerical indicators based on machine learning as described in any of the above items.
[0035] The present invention also provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the method for extracting document numerical indicators based on machine learning as described in any of the above items is implemented.
[0036] Compared with the prior art, the beneficial effects of the present invention are:
[0037] According to preset rules, the document to be processed is divided into key paragraphs and non-key paragraphs; an index extraction model is constructed by natural language processing and deep learning, and the key paragraphs and the non-key paragraphs are respectively input into the index extraction model, and the index extraction model outputs key paragraph indexes and non-key paragraph indexes respectively; the key paragraph indexes and the non-key paragraph indexes are integrated, and a preset index feature scoring model is input to screen the key paragraph indexes and the non-key paragraph indexes, and the index extraction results are output. The present invention greatly improves the focusing ability of the extraction through the hierarchical extraction method of key paragraph separation, reduces the workload of traditional extraction indicators, and effectively improves the accuracy of traditional indicator extraction by building an index extraction model based on natural language processing and deep learning.
[0038] Furthermore, the historical sequence data is input into the preset prediction model to obtain the prediction results, and the average of the prediction results and the filtered indicator results is calculated as the confidence level for sorting, including the indicator name, indicator value, indicator unit, indicator confidence level, short sentences and paragraph texts that hit the indicator. By building a preset prediction model, the error indicators can be further screened by calculating the prediction results and the filtered indicator results, thereby improving the stability of the indicator extraction effect. BRIEF DESCRIPTION OF THE DRAWINGS
[0039] In order to more clearly illustrate the technical solution of the present invention, the drawings required for use in the implementation mode will be briefly introduced below. Obviously, the drawings described below are only some implementation modes of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0040] Figure 1 It is a flowchart of a method for extracting document numerical indicators based on machine learning provided by a certain embodiment of the present invention;
[0041] Figure 2 It is a schematic diagram of a process of extracting key paragraphs provided by a certain embodiment of the present invention;
[0042] Figure 3 It is a schematic diagram of a flow chart of index extraction provided by a certain embodiment of the present invention;
[0043] Figure 4is a flowchart of an example of an indicator extraction method provided by another embodiment of the present invention;
[0044] Figure 5 It is a flowchart of a method for extracting document numerical indicators based on machine learning provided by another embodiment of the present invention;
[0045] Figure 6 It is a structural schematic diagram of a device for extracting document numerical indicators based on machine learning provided in a certain embodiment of the present invention;
[0046] Figure 7 It is a schematic diagram of the structure of a terminal device provided by a certain embodiment of the present invention. DETAILED DESCRIPTION
[0047] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0048] It should be understood that the step numbers used in this document are only for convenience of description and are not intended to limit the order in which the steps are executed.
[0049] It should be understood that the terms used in the present specification are only for the purpose of describing specific embodiments and are not intended to limit the present invention. As used in the present specification and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.
[0050] The terms “include” and “comprising” indicate the presence of described features, integers, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or combinations thereof.
[0051] The term "and / or" means and includes any and all possible combinations of one or more of the associated listed items.
[0052] See also Figure 1 , a certain embodiment of the present invention provides a method for extracting document numerical indicators based on machine learning. Figure 1 As shown, the method for extracting document numerical indicators based on machine learning includes steps S10 to S30. The specific steps are as follows:
[0053] S10: Divide the document to be processed into key paragraphs and non-key paragraphs according to preset rules.
[0054] Please refer to Figure 2 , specifically, extract key paragraphs from the document, extract a few key paragraphs from the document with a large number of pages, and store them as a new document for subsequent manual processing.
[0055] Determine the format of the document to be input, convert the pdf document format to the word format to generate an input document, parse the input document to obtain a set of text and table elements, and use a rule matching method to locate the key paragraphs according to the set of text and table elements, and output the key paragraphs.
[0056] Input the document to be processed, which can be a word document or a pdf document. If it is a pdf document, it is converted to the word format for subsequent processing. The conversion method is to use ilovepdf (an existing tool) for transcribing. After transcribing, parse elements such as text, tables, and charts through the docx package, and then locate the key paragraphs. The location method is rule matching, and the rule is: paragraph feature regular expression + keyword. For example, {any digital title} credit business, {any digital title} is the paragraph feature regular expression designed by us. For example: r'^[(\[
(]*[\d一二三四五六七八九十第节章]{1,3}[)\]
[0057] S20: Build an index extraction model using natural language processing and deep learning, and input the key paragraphs and the non-key paragraphs into the index extraction model respectively. The index extraction model outputs key paragraph indexes and non-key paragraph indexes respectively.
[0058] Please refer to Figure 3 and Figure 4 , specifically, build an index extraction model through natural language processing and deep learning, create a regular intelligent construction method to enable the business to more conveniently build and maintain index regular expressions, and intelligently generate regular expressions. As Figure 4 shown in, the business input item (including index subject A and index dimension U), and examples of the generation method. Among them, N: that is, Neighbour, any character of any length, and the length can be configured. NUM: random_num: match any number, which can contain common digital characters such as,. '|': or character. Arrow: represents string concatenation.
[0059] According to such as Figure 4The regular expression constructed as shown matches the text set one by one, and records the hit text, numerical value, and mark hit type features as regular expression matches. For keyword recall in key paragraphs and non-key paragraphs, the text set is recalled through keywords, and multiple keywords (keyword relationships include and, or) are supported. For syntactic analysis in key paragraphs and non-key paragraphs, an indicator candidate set is established, and the syntactic structure analysis of the sentence for keyword recall is performed through Baidu's open source ddparser syntactic analysis package, and the SVO, ATT_N, and ADV_V relationships are extracted by programming, where ATT_N represents a modification relationship, ADV_V represents a vice-verb relationship, and SVO represents a subject-predicate-object relationship.
[0060] Evaluation and ranking of candidate indicator sets: Calculate the similarity between the indicator name X recalled by the above syntactic analysis and the indicator name I provided by the business. The calculation method is: jieba.cut_for_search(I)&jieba.cut_for_search(X), that is, use jieba's search word segmentation mode to segment the words, take the union of the two name segmentation results, and calculate the proportion of the union words in I. The higher the proportion, the higher the similarity, and the more likely it is the target indicator to be found.
[0061] The TextCNN model scores the semantic granularity of sentences: A TextCNN model is established to output a score between 0 and 1, where 0 means that the sentence is not a target indicator sentence from a semantic point of view, and 1 means that the sentence is a target indicator sentence.
[0062] Finally, the indicator names, indicator values, hit types, evaluation scores, etc. of the above two are integrated (the same sentence features are integrated together) to provide them to the next step for processing.
[0063] S30: integrating the key paragraph index and the non-key paragraph index, inputting a preset index feature scoring model to screen the key paragraph index and the non-key paragraph index, and outputting an index extraction result.
[0064] The indicators in the key paragraph indicators and the non-key paragraph indicators are marked, and the indicator features are sorted and normalized and input into the logistic regression model for training to obtain the preset indicator feature scoring model, the key paragraph indicators and the non-key paragraph indicators are integrated and input into the preset indicator feature scoring model, the indicators below the first threshold are filtered, and the indicator extraction result is output.
[0065] Specifically, the indicator names, indicator values, hit types, and evaluation scores of the key paragraphs and non-key paragraphs are integrated according to step S20. The recalled indicator names, recalled indicator values, and features are labeled as data sets, with the correct indicator label marked as 1 and the incorrect indicator label marked as 0. The features are sorted and normalized, and input into the logistic regression model for training to obtain a preset indicator feature scoring model. The integrated indicators are input into the preset indicator feature scoring model for screening, and the newly recalled indicators and features are predicted by the model, and the indicators below the first threshold of 0.5 points are filtered out (at the same time, at least one indicator with the maximum prediction score is retained. If it is below the first threshold of 0.5 points, it is marked to facilitate manual correction by the business.). Output the indicator extraction results.
[0066] The use of unstructured data has always been a difficulty in the field of artificial intelligence. Current data analysis technology mainly targets structured data. Unstructured data has difficulties such as many expression changes, diverse forms, and complex structures. Current technology has problems of low efficiency and low accuracy in the use of unstructured data. What the present invention aims to solve is the problem of intelligent extraction of unstructured data for indicators in the financial field.
[0067] The present invention uses a hierarchical extraction method that separates key paragraphs, which better combines expert experience and model capabilities, greatly improving the focus of extraction. Compared with the huge number of pages in the original document, the extracted key paragraphs also greatly reduce the workload of business personnel. By building an indicator extraction model based on natural language processing and deep learning, the accuracy of traditional indicator extraction is effectively improved, the extraction is more comprehensive and effective, and at the same time, invalid recalls are reduced, the workload of business personnel is reduced, and work efficiency is improved.
[0068] See also Figure 5 In a specific embodiment, after step S30 outputs the index extraction result, the following steps are also included:
[0069] S40: Input the historical sequence data into a preset prediction model to obtain a prediction result, calculate the average of the prediction result and the screened indicator result as the confidence level for sorting, including the indicator name, indicator value, indicator unit, indicator confidence level, short sentences that hit the indicator, and paragraph text.
[0070] Adopt ARIMA model, input the historical series of each indicator, train the model and predict the value P of the next series of the value (P is extremely unlikely to be zero, but for logical consistency, if it is zero, it is taken as 0.01), and calculate it with the indicator R returned by the indicator extraction model. The calculation method is: 1-(RP)^2 / P^2. If this value is less than 0, it is taken as 0.
[0071] Finally, the mean results of the comprehensive indicator feature scoring model and the indicator historical sequence prediction model are sorted as confidence levels, and the indicator extraction results are displayed to the business. The indicator name, indicator value, indicator unit, indicator confidence level, short sentences and paragraph texts of the hit indicators are also displayed to facilitate quick browsing by the business.
[0072] The present invention further builds a sequence prediction model based on the ARIMA model, so that the entire invention has common sense and rational judgment ability, filters out erroneous indicators, and improves the stability of the indicator extraction effect. Current data prediction is performed based on historical data, and the difference calculation is performed with the actual extracted values. The prior art does not have inference based on historical data after the indicator is extracted. The present invention is based on ARIMA and develops an indicator historical sequence prediction model, which can deduce the extracted values and improve the accuracy of indicator extraction.
[0073] See also Figure 6 Another embodiment of the present invention provides a device for extracting document numerical indicators based on machine learning, comprising:
[0074] The division module 11 is used to divide the document to be processed into key paragraphs and non-key paragraphs according to preset rules.
[0075] The extraction module 12 is used to construct an index extraction model using natural language processing and deep learning, and input the key paragraphs and the non-key paragraphs into the index extraction model respectively, and the index extraction model outputs the key paragraph index and the non-key paragraph index respectively.
[0076] The screening module 13 is used to integrate the key paragraph index and the non-key paragraph index, input a preset index feature scoring model to screen the key paragraph index and the non-key paragraph index, and output an index extraction result.
[0077] Preferably, it also includes a calculation module, which is used to: input historical sequence data into a preset prediction model to obtain prediction results, calculate the average of the prediction results and the screened indicator results as confidence for sorting, including indicator name, indicator value, indicator unit, indicator confidence, short sentences and paragraph text that hit the indicator.
[0078] Preferably, the extraction module is further used to: determine the format of the document to be input, convert the PDF document format into a word format, and generate an input document;
[0079] Parsing the input document to obtain a set of text and table elements;
[0080] A rule matching method is adopted to locate the key paragraph according to the text and the table element set, and the key paragraph is output.
[0081] Preferably, the screening module is further used to: mark the indicators in the key paragraph indicators and the non-key paragraph indicators, and sort and normalize the indicator features and input them into the logistic regression model for training to obtain the preset indicator feature scoring model;
[0082] The key paragraph index and the non-key paragraph index are integrated, input into the preset index feature scoring model, the index below the first threshold is filtered, and the index extraction result is output.
[0083] For the specific definition of the device for extracting document numerical indicators based on machine learning, please refer to the definition of the method for extracting document numerical indicators based on machine learning above, which will not be repeated here. Each module in the above-mentioned device for extracting document numerical indicators based on machine learning can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.
[0084] See also Figure 7 , a certain embodiment of the present invention provides a terminal device, including:
[0085] one or more processors;
[0086] A memory, coupled to the processor, for storing one or more programs;
[0087] When the one or more programs are executed by the one or more processors, the one or more processors implement the method for extracting document numerical indicators based on machine learning as described above.
[0088] The processor is used to control the overall operation of the terminal device to complete all or part of the steps of the above-mentioned method for extracting document numerical indicators based on machine learning. The memory is used to store various types of data to support the operation of the terminal device, and these data may include, for example, instructions for any application or method used to operate on the terminal device, and application-related data. The memory can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (Static Random Access Memory, referred to as SRAM), electrically erasable programmable read-only memory (Electrically Erasable Programmable Read-Only Memory, referred to as EEPROM), erasable programmable read-only memory (Erasable Programmable Read-Only Memory, referred to as EPROM), programmable read-only memory (Programmable Read-Only Memory, referred to as PROM), read-only memory (Read-Only Memory, referred to as ROM), magnetic memory, flash memory, disk or optical disk.
[0089] In an exemplary embodiment, the terminal device can be implemented by one or more application specific integrated circuits (Application Specific Integrated Circuit, ASC), digital signal processors (Digital Signal Processor, DSP), digital signal processing devices (Digital Signal Processing Device, DSPD), programmable logic devices (Programmable Logic Device, PLD), field programmable gate arrays (Field Programmable Gate Array, FPGA), controllers, microcontrollers, microprocessors or other electronic components to execute the method for extracting document numerical indicators based on machine learning as described in any of the above embodiments, and achieve the same technical effect as the above method.
[0090] In another exemplary embodiment, a computer-readable storage medium including a computer program is also provided, and when the computer program is executed by a processor, the steps of the method for extracting document numerical indicators based on machine learning as described in any of the above embodiments are implemented. For example, the computer-readable storage medium can be the above-mentioned memory including the computer program, and the above-mentioned computer program can be executed by the processor of the terminal device to complete the method for extracting document numerical indicators based on machine learning as described in any of the above embodiments, and achieve the technical effect consistent with the above method.
[0091] The above is a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several improvements and modifications without departing from the principle of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.
Claims
1. A method for extracting document numerical indicators based on machine learning, characterized in that: include: Divide the document to be processed into key paragraphs and non-key paragraphs according to preset rules; An index extraction model is constructed by using natural language processing and deep learning, and the key paragraphs and the non-key paragraphs are respectively input into the index extraction model, and the index extraction model outputs key paragraph indexes and non-key paragraph indexes respectively; Among them, natural language processing and deep learning are used to construct an index extraction model, and the key paragraphs and the non-key paragraphs are respectively input into the index extraction model, and the index extraction model outputs key paragraph indicators and non-key paragraph indicators respectively, including: Match the text set one by one according to the constructed regular expression, and record the hit text, value and mark hit type features as regular expression matching during the matching process; For keyword recall in key paragraphs and non-key paragraphs, the text set is recalled through keywords; for syntactic analysis in key paragraphs and non-key paragraphs, an indicator candidate set is established, and the sentence of keyword recall is analyzed through Baidu's open source ddparser syntax analysis package, and the subject-verb-object relationship, modification relationship, and adverb relationship are extracted by programming to obtain the indicator name recalled by syntactic analysis; the indicator name recalled by syntactic analysis is calculated for similarity with the indicator name provided by the business; the calculation method is to perform word segmentation through the search word segmentation mode of jieba, take the union of the two name word segmentation results, calculate the proportion of the union words to the indicator name provided by the business, obtain the proportion result, and determine the target indicator according to the proportion result; A TextCNN model is built to output the scoring features of the sentence semantic granularity. The TextCNN model outputs a score between 0 and 1, where 0 means that the sentence is not a target index sentence from a semantic point of view, and 1 means that the sentence is a target index sentence. Marking the indicators in the key paragraph indicators and the non-key paragraph indicators, sorting and normalizing the indicator features, and inputting them into the logistic regression model for training to obtain the preset indicator feature scoring model; The key paragraph index and the non-key paragraph index are integrated, input into the preset index feature scoring model, the index below the first threshold is filtered, and the index extraction result is output.
2. The method for extracting document numerical indicators based on machine learning according to claim 1 is characterized in that: Also includes: The historical sequence data is input into the preset prediction model to obtain the prediction result, and the mean of the prediction result and the screened indicator extraction result is calculated as the confidence for sorting, including the indicator name, indicator value, indicator unit, indicator confidence, short sentences and paragraph texts that hit the indicator.
3. The method for extracting document numerical indicators based on machine learning according to claim 2 is characterized in that: The step of dividing the document to be processed into key paragraphs and non-key paragraphs according to preset rules includes: Determine the format of the document to be input, convert the PDF document format into word format, and generate the input document; Parsing the input document to obtain a set of text and table elements; A rule matching method is adopted to locate the key paragraph according to the text and the table element set, and the key paragraph is output.
4. A device for extracting document numerical indicators based on machine learning, characterized in that: include: A division module is used to divide the document to be processed into key paragraphs and non-key paragraphs according to preset rules; An extraction module is used to construct an index extraction model by using natural language processing and deep learning, and the key paragraphs and the non-key paragraphs are respectively input into the index extraction model, and the index extraction model outputs key paragraph indicators and non-key paragraph indicators respectively; wherein, the index extraction model is constructed by using natural language processing and deep learning, and the key paragraphs and the non-key paragraphs are respectively input into the index extraction model, and the index extraction model outputs key paragraph indicators and non-key paragraph indicators respectively, including: matching the text set one by one according to the constructed regular expression, and recording the hit text, numerical value and mark hit type features as regular expression matching during the matching process; for keyword recall in key paragraphs and non-key paragraphs, the text set is recalled by keywords; for syntactic analysis in key paragraphs and non-key paragraphs Analyze and establish the candidate set of indicators. For the sentences recalled by keywords, perform syntactic structure analysis through Baidu's open source ddparser syntactic analysis package, program to extract the subject-verb-object relationship, modification relationship, and adverb-verb relationship, and obtain the indicator name recalled by syntactic analysis; calculate the similarity between the indicator name recalled by syntactic analysis and the indicator name provided by the business; the calculation method is to perform word segmentation through the search word segmentation mode of jieba, take the union of the two name word segmentation results, calculate the proportion of the union words to the indicator name provided by the business, obtain the proportion result, and determine the target indicator according to the proportion result; establish a TextCNN model to output the scoring features of the semantic granularity of the sentence; the TextCNN model outputs a score between 0 and 1, where 0 means that the sentence is not a target indicator sentence from a semantic point of view, and 1 means that the sentence is a target indicator sentence; The screening module is used to mark the indicators in the key paragraph indicators and the non-key paragraph indicators, and to organize and normalize the indicator features and input them into the logistic regression model for training to obtain the preset indicator feature scoring model; integrate the key paragraph indicators and the non-key paragraph indicators, input them into the preset indicator feature scoring model, filter the indicators below the first threshold, and output the indicator extraction results.
5. The device for extracting document numerical indicators based on machine learning according to claim 4, characterized in that: Also included is a calculation module, the calculation module is used to: The historical sequence data is input into the preset prediction model to obtain the prediction result, and the prediction result and the mean of the filtered index extraction result are calculated as the confidence for sorting, including the index name, index value, index unit, index confidence, short sentences and paragraph texts that hit the index.
6. The device for extracting document numerical indicators based on machine learning according to claim 5, characterized in that: The extraction module is further used for: Determine the format of the document to be input, convert the PDF document format into word format, and generate the input document; Parsing the input document to obtain a set of text and table elements; A rule matching method is adopted to locate the key paragraph according to the text and the table element set, and the key paragraph is output.
7. A terminal device, characterized in that: include: one or more processors; A memory, coupled to the processor, for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method for extracting document numerical indicators based on machine learning as described in any one of claims 1-3.
8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for extracting document numerical indicators based on machine learning as described in any one of claims 1 to 3 is implemented.
Citation Information
Patent Citations
Extraction method for evaluation document index
CN106484920A
Determining item feature information from user content
US10410224B1
