Information extraction method and device

Abstract

The present application provides an information extraction method and device, wherein the information extraction method comprises: performing sentence processing on the received text to be processed to obtain an initial sentence set; obtaining classification information corresponding to each sentence in the initial sentence set; determining a target sentence set in the initial sentence set according to the classification information corresponding to each sentence, and determining a target paragraph set according to the target sentence set; inputting the target paragraph set into a named entity recognition model for processing to obtain a target entity set output by the named entity recognition model; and determining the information extraction result of the text to be processed according to the target entity set. Through this method, the efficiency and accuracy of information extraction are improved.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to an information extraction method and apparatus, a computing device, and a computer-readable storage medium. Background Art

[0002] In the field of text processing, it is usually necessary to extract information from the text. The current processing method is to identify and extract the entire text, but the information extraction from the text belongs to the chapter-level information extraction. The current information model cannot process the text, or the extraction effect is poor. The information to be extracted in some texts usually only appears in some areas and does not appear in the entire text. For example, in a contract document, the information to be extracted appears in a specific area, and the clause information usually does not contain entity information. When information is extracted from the entire text, the extraction efficiency is very low. The current information extraction method for the entire text is rule-based information extraction, that is, using artificial summary rules to construct some rules, and finding entity information from the text based on the rules. However, due to the many types of texts, it is difficult for the rule-based information extraction method to achieve very good generalization. Different rules must be configured for different types of texts, and the workload is also very huge, consuming manpower and material resources. Summary of the invention

[0003] In view of this, the embodiments of the present application provide an information extraction method and apparatus, a computing device, and a computer-readable storage medium to solve the technical defects existing in the prior art.

[0004] According to a first aspect of an embodiment of the present application, there is provided an information extraction method, comprising:

[0005] Sentence processing is performed on the received text to be processed to obtain an initial sentence set;

[0006] Obtaining classification information corresponding to each sentence in the initial sentence set;

[0007] Determine a target sentence set in the initial sentence set according to the classification information corresponding to each sentence, and determine a target paragraph set according to the target sentence set;

[0008] Inputting the target paragraph set into a named entity recognition model for processing to obtain a target entity set output by the named entity recognition model;

[0009] The information extraction result of the text to be processed is determined according to the target entity set.

[0010] According to a second aspect of an embodiment of the present application, there is provided an information extraction device, including:

[0011] A sentence segmentation module is configured to perform sentence segmentation processing on the received text to be processed to obtain an initial sentence set;

[0012] A sentence classification module is configured to obtain classification information corresponding to each sentence in the initial sentence set;

[0013] A determination module, configured to determine a target sentence set in the initial sentence set according to classification information corresponding to each sentence, and determine a target paragraph set according to the target sentence set;

[0014] An entity recognition module is configured to input the target paragraph set into a named entity recognition model for processing, and obtain a target entity set output by the named entity recognition model;

[0015] The information extraction module is configured to determine the information extraction result of the text to be processed according to the target entity set.

[0016] According to a third aspect of an embodiment of the present application, a computing device is provided, comprising a memory, a processor, and computer instructions stored in the memory and executable on the processor, wherein the processor implements the steps of the information extraction method when executing the computer instructions.

[0017] According to a fourth aspect of an embodiment of the present application, a computer-readable storage medium is provided, which stores computer instructions, and when the computer instructions are executed by a processor, the steps of the information extraction method are implemented.

[0018] According to a fifth aspect of an embodiment of the present application, a chip is provided, which stores computer instructions, and the computer instructions implement the steps of the information extraction method when executed by the chip.

[0019] The information extraction method provided by the embodiment of the present application includes: performing sentence processing on the received text to be processed to obtain an initial sentence set; obtaining classification information corresponding to each sentence in the initial sentence set; determining a target sentence set in the initial sentence set according to the classification information corresponding to each sentence, and determining a target paragraph set according to the target sentence set; inputting the target paragraph set into a named entity recognition model for processing to obtain a target entity set output by the named entity recognition model; and determining the information extraction result of the text to be processed according to the target entity set. Through this method, sentences that do not contain entity information in the text to be processed are first filtered out through sentence classification, and then entity recognition is performed on sentences containing entity information, thereby avoiding the problem of low extraction efficiency caused by entity extraction of the entire document, and also avoiding the problem of low information extraction accuracy caused by extracting wrong entities in paragraphs that do not contain entities. The efficiency and accuracy of information extraction are improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1is a structural block diagram of a computing device provided in an embodiment of the present application;

[0021] Figure 2 is a flow chart of the information extraction method provided in an embodiment of the present application;

[0022] Figure 3 It is a schematic diagram of extracting information from a contract text using the information extraction method provided in an embodiment of the present application;

[0023] Figure 4 It is a structural diagram of the information extraction device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0024] Many specific details are described in the following description to facilitate a full understanding of the present application. However, the present application can be implemented in many other ways than those described herein, and those skilled in the art can make similar generalizations without violating the connotation of the present application, so the present application is not limited by the specific implementation disclosed below.

[0025] The terms used in one or more embodiments of the present application are only for the purpose of describing specific embodiments, and are not intended to limit one or more embodiments of the present application. The singular forms of "a", "said" and "the" used in one or more embodiments of the present application and the appended claims are also intended to include plural forms, unless the context clearly indicates other meanings. It should also be understood that the term "and / or" used in one or more embodiments of the present application refers to and includes any or all possible combinations of one or more associated listed items.

[0026] It should be understood that, although the terms first, second, etc. may be used to describe various information in one or more embodiments of the present application, these information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of one or more embodiments of the present application, the first may also be referred to as the second, and similarly, the second may also be referred to as the first. Depending on the context, the word "if" as used herein may be interpreted as "in response to determination".

[0027] First, the terms involved in one or more embodiments of the present invention are explained.

[0028] Entity: refers to an object or thing that exists objectively in the real world and can be distinguished from each other, such as a person's name, an address name, a contract name, a date, etc.

[0029] Bert (Bidirectional Encoder Representations from Transformers): A bidirectional attention neural network model, a pre-training technology for natural language processing.

[0030] Text to be processed: the carrier from which information needs to be extracted, such as text information in contracts, text information in legal documents, etc.

[0031] Files to be processed: files corresponding to the texts to be processed, such as contracts, legal documents, etc.

[0032] Initial sentence set: the set of initial sentences obtained after the text to be processed is processed by sentence segmentation, and the classification of sentences is not distinguished in the initial sentence set.

[0033] Target sentence set: After screening in the initial sentence set, the classification information of the sentences is determined to be a set consisting of sentences including entity information, that is, the target sentences in the target sentence set all include entity information.

[0034] Target paragraph set: A paragraph consisting of a target sentence is a target paragraph, and multiple target paragraphs constitute a target paragraph set.

[0035] Target entity set: a set of entity information extracted from the target paragraph, including the target entity, the target entity type, and the sentence number corresponding to the target entity.

[0036] Preset sentence identifier: an identifier used to divide the initial sentence, such as ".", "?", "!", etc.

[0037] Classification information: information used to indicate sentence classification, which is divided into first-category sentences and second-category sentences. First-category sentences represent sentences that include entity information, and second-category sentences represent sentences that do not include entity information.

[0038] In the field of text processing, it is usually necessary to extract information from the text. The current processing method is to identify and extract the entire text. The information extraction of the text belongs to the chapter-level information extraction. The current information model cannot process the text, or the extraction effect is poor. For example, for the text corresponding to the contract, information extraction can be performed through the information extraction model, but the information extraction model is sentence-level information extraction, the processed text length is short, and the training data generally contains relatively rich contextual information, which makes it easier to infer the category of entities and the actual relationship. The information extraction of the entire contract can also be carried out based on rules, that is, some rules are constructed using artificial summary rules to find the text information in the contract. However, due to the large number of contract writing types, rule-based methods are difficult to generalize well.

[0039] Based on this, in the present application, an information extraction method and apparatus, a computing device and a computer-readable storage medium are provided, which are described in detail one by one in the following embodiments.

[0040] Figure 1The structure block diagram of a computing device 100 according to an embodiment of the present application is shown. The components of the computing device 100 include but are not limited to a memory 110 and a processor 120. The processor 120 is connected to the memory 110 via a bus 130, and the database 150 is used to store data.

[0041] The computing device 100 also includes an access device 140 that enables the computing device 100 to communicate via one or more networks 160. Examples of these networks include a public switched telephone network (PSTN), a local area network (LAN), a wide area network (WAN), a personal area network (PAN), or a combination of communication networks such as the Internet. The access device 140 may include one or more of any type of network interface (e.g., a network interface card (NIC)) whether wired or wireless, such as an IEEE 802.11 wireless local area network (WLAN) wireless interface, a World Wide Interoperability for Microwave Access (Wi-MAX) interface, an Ethernet interface, a universal serial bus (USB) interface, a cellular network interface, a Bluetooth interface, a near field communication (NFC) interface, and the like.

[0042] In one embodiment of the present application, the above components of the computing device 100 and Figure 1 Other components not shown in the figure may also be connected to each other, for example, via a bus. It should be understood that Figure 1 The computing device structure block diagram shown is only for the purpose of illustration, and is not intended to limit the scope of the present application. Those skilled in the art may add or replace other components as needed.

[0043] The computing device 100 may be any type of stationary or mobile computing device, including a mobile computer or mobile computing device (e.g., a tablet computer, a personal digital assistant, a laptop computer, a notebook computer, a netbook, etc.), a mobile phone (e.g., a smart phone), a wearable computing device (e.g., a smart watch, smart glasses, etc.), or other types of mobile devices, or a stationary computing device such as a desktop computer or PC. The computing device 100 may also be a mobile or stationary server.

[0044] The processor 120 may execute Figure 2 Steps in the information extraction method shown. Figure 2 A flowchart of an information extraction method according to an embodiment of the present application is shown, including steps 202 to 210.

[0045] Step 202: Sentence processing is performed on the received text to be processed to obtain an initial sentence set.

[0046] Among them, the text to be processed is the text that needs to be extracted information, such as text information in a contract, text information in a legal document, etc. The information extraction method provided by the present application provides a method for extracting relevant entity information from a text. The information extraction method provided by the present application can process both long texts and short texts. Specifically, long text refers to texts in which the number of characters in the text to be processed exceeds a threshold, and short text specifically refers to texts in which the number of characters in the text to be processed does not exceed a threshold. Preferably, the information extraction method provided by the present application has a more obvious improvement in the information extraction rate and extraction accuracy of long texts.

[0047] After obtaining the text to be processed, it is necessary to segment the text to be processed, that is, to divide the text to be processed into multiple initial sentences to form an initial sentence set, wherein the initial sentence set is a set of initial sentences obtained after the text to be processed is segmented.

[0048] In practical applications, the text to be processed is composed of multiple sentences. The method of segmenting the text to be processed is usually to divide it by punctuation marks. Therefore, the received text to be processed is segmented to obtain an initial sentence set, including S2022-S2024:

[0049] S2022, obtaining a preset sentence identifier;

[0050] S2024. Segment the text to be processed according to the preset sentence identifier to obtain an initial sentence set.

[0051] Among them, the preset sentence identifier specifically refers to the identifier used to divide the initial sentence, such as ".", "?", "!", etc. In actual applications, when performing sentence processing, any punctuation mark can be used as the preset sentence identifier, or specific punctuation marks can be used as the preset sentence identifier. The specific content of the preset sentence identifier is not limited and is subject to actual application.

[0052] The text to be processed can be divided into sentences according to preset sentence identifiers, so that the text to be processed can be divided into multiple initial sentences to form an initial sentence set.

[0053] In a specific implementation provided in the present application, the text to be processed is "Sentence 1, sentence 2, sentence 3. Sentence 4; sentence 5; sentence 6! Sentence 7. ...", and preset sentence identifiers ",",".",";,"?","!","!", etc. are obtained. The text to be processed is sentence processed according to the preset sentence identifiers to obtain multiple initial sentences and form an initial sentence set {sentence 1, sentence 2, sentence 3...sentence n}.

[0054] In practical applications, the text to be processed usually appears in the form of a file. Based on this, the method further includes:

[0055] Receive files to be processed;

[0056] Parse the file to be processed to obtain the text to be processed corresponding to the file to be processed.

[0057] The file to be processed is the file corresponding to the text to be processed. For example, if the text to be processed is a contract text, then the file to be processed is the contract document; for another example, if the text to be processed is a legal document, then the file to be processed is the legal document. Specifically, the file to be processed can be a file in txt format, a file in pdf format, a file in docx format, etc. In this application, the format of the file to be processed is not limited. The file to be processed can be uploaded by the user through the client, or it can be pre-saved in the client and selected by the user.

[0058] Receive a file to be processed, parse the file through text recognition technology, if the file to be processed is of type DOC, DOCX, WPS or TXT, process the file to be processed as DOCX type, if the file to be processed is not of DOCX type, convert the file to DOCX type through tools such as LibreOffice (which has powerful data import and export functions, can directly import PDF documents, Microsoft Word (.doc files), LotusWord, and supports major OpenXML formats), and extract text from the DOCX format; if the file to be processed is of type PDF, PNG, JPG, JPEG or TIFF, process the file to be processed as PDF type, if the file to be processed is not of PDF type, convert the file to PDF type, perform target recognition and OCR recognition to extract text for the PDF type file to be processed.

[0059] The text content in the to-be-processed file is extracted to obtain the corresponding to-be-processed text, wherein the to-be-processed text only retains the text content in the to-be-processed file and clears the formatting style and other settings in the to-be-processed file.

[0060] Step 204: Obtain classification information corresponding to each sentence in the initial sentence set.

[0061] After obtaining the initial sentence set, each sentence needs to be classified to determine the classification information of each sentence.

[0062] In practical applications, the classification information of each sentence can be determined through manual labeling, that is, each sentence is displayed, and the staff classifies and labels each sentence to determine the classification information of each sentence. However, the above method has high requirements for the staff and consumes more energy. When the amount of data is large, misjudgment may occur. Therefore, obtaining the classification information corresponding to each sentence in the initial sentence set can also be achieved through a sentence classification model.

[0063] Specifically, obtaining classification information corresponding to each sentence in the initial sentence set includes:

[0064] The initial sentence set is input into a sentence classification model for processing, and classification information corresponding to each sentence in the initial sentence set output by the sentence classification model is obtained.

[0065] The sentence classification model specifically refers to an artificial intelligence model used to classify the initial sentences and output the classification information of the initial sentences. In practical applications, the sentence classification model classifies sentences based on whether the initial sentences contain entity information. The classification information is the classification corresponding to each initial sentence.

[0066] Specifically, the initial sentence set is input into a sentence classification model for processing, and classification information corresponding to each sentence in the initial sentence set output by the sentence classification model is obtained, including S2042-S2046:

[0067] S2042, determining a target initial sentence in the initial sentence set, and inputting the target initial sentence into a sentence classification model for sentence classification, wherein the sentence classification model is trained to determine whether a sentence contains entity information;

[0068] S2044, when the target initial sentence contains entity information, obtaining classification information corresponding to the target initial sentence output by the sentence classification model as a first category sentence;

[0069] S2046. When the target initial sentence does not contain entity information, the classification information corresponding to the target initial sentence output by the sentence classification model is obtained as a second-category sentence.

[0070] In practical applications, the sentence classification model can process only one sentence at a time, or can process multiple sentences at a time. The target initial sentence is the initial sentence that needs to be input into the sentence classification model for processing.

[0071] In practical applications, each initial sentence in the initial sentence set can be used as a target initial sentence in turn according to the actual situation; or some initial sentences with special meanings in the initial sentence set can be determined as target initial sentences, such as the initial sentence set includes paragraph information corresponding to each initial sentence, and only the first initial sentence in each paragraph is selected as the target initial sentence. In order to improve the accuracy of information extraction, it is preferred in this solution that each initial sentence in the initial sentence set is processed as a target initial sentence.

[0072] When the sentence classification model processes only one sentence, a target initial sentence can be determined in the initial sentence set, and the target initial sentence can be input into the sentence classification model for processing. At this time, the selection can be made according to the order of each initial sentence in the initial sentence set, or the target initial sentence can be randomly selected from the initial sentence set. The specific selection method is not limited. It should be noted that each initial sentence is only selected and input into the classification model for processing once, and some or all of the initial sentences in the initial sentence set are input into the sentence classification model for processing; when the sentence classification model processes multiple sentences, multiple target initial sentences can be determined in the initial sentence set, the multiple target initial sentences are spliced, and the spliced ​​target initial sentences are input into the sentence classification model for processing.

[0073] The sentence classification model is trained to determine whether the target initial sentence contains entity information. When the target initial sentence includes entity information, the classification information corresponding to the target initial sentence is output as a first-category sentence; when the target initial sentence does not include entity information, the classification information corresponding to the target initial sentence is output as a second-category sentence, where the first-category sentence represents a sentence that includes entity information, and the second-category sentence represents a sentence that does not include entity information.

[0074] In a specific implementation provided in the present application, taking each initial sentence in the initial sentence set as a target initial sentence as an example, following the above example, the initial sentence set is {sentence 1, sentence 2, sentence 3...sentence n}. When the sentence classification model processes only one sentence at a time, sentence 1, sentence 2, sentence 3...sentence n are selected from the initial sentence set in turn as target initial sentences, and input into the sentence classification model to obtain sentence classification information corresponding to each sentence, which can be specifically identified as "sentence 1, first category sentence", "sentence 2, first category sentence", "sentence 3, second category sentence"... "sentence n, second category sentence", that is, each sentence in the initial sentence set is input into the sentence classification model to complete classification, and sentence classification information corresponding to each sentence is obtained.

[0075] In another specific implementation manner provided by the present application, taking some initial sentences in the initial sentence set as target initial sentences as an example, following the above example, the initial sentence set is {sentence 1, sentence 2, sentence 3 ... sentence n}, wherein the target initial sentences are sequentially or randomly determined to be sentence 1, sentence 2, sentence 3 ... sentence m, where m is less than or equal to n, the target initial sentences are spliced ​​together with specific symbols, such as using "#" for splicing, and the spliced ​​target initial sentences are obtained as "sentence 1# sentence 2# sentence 3# ... sentence m", and the spliced ​​target initial sentences are input into the sentence classification model, and the sentence classification model outputs each target initial sentence The corresponding classification information is "first category sentence #first category sentence #second category sentence...second category sentence", where, when m is less than n, that is, some of the initial sentences in the initial sentence set are used as target initial sentences, which are spliced ​​and input into the sentence classification model. It can be understood that when all the initial sentences in the initial sentence set need to be input into the sentence classification model, and m is less than n, the target initial sentences are determined from the remaining nm sentences, and spliced ​​in the same way. The spliced ​​target initial sentences are input into the sentence classification model to obtain the classification information corresponding to each target initial sentence, and so on, until all sentences in the initial sentence set are input into the sentence classification model to complete the classification.

[0076] The sentence classification model is a trained artificial intelligence model. The sentence classification model can be a Bert model, a Transformer model, etc. In the embodiment of the present application, the sentence classification model is explained by taking the Bert model as an example. The sentence classification model includes an embedding layer, 12 sequentially connected encoding layers and a fully connected layer. The target initial sentence is input into the embedding layer to generate a target initial sentence vector, and then the target initial sentence vector is sequentially input into the 12 encoding layers for processing to obtain the target initial sentence encoding vector, and then the target initial sentence encoding vector is input into the fully connected layer for processing, and the classification information corresponding to the target initial sentence encoding vector is output. The sentence classification model is pre-trained and has the ability to output the classification information corresponding to the sentence according to the sentence. Specifically, the sentence classification model is trained by the following steps:

[0077] Obtaining sample sentences and sample classification information corresponding to the sample sentences;

[0078] Inputting the sample sentence into a sentence classification model to be trained, so that the sentence classification model to be trained determines whether the sample sentence contains entity information;

[0079] Obtaining predicted classification information corresponding to the sample sentence output by the sentence classification model to be trained;

[0080] Calculate the classification model loss value according to the predicted classification information and the sample classification information;

[0081] The model parameters of the sentence classification model to be trained are adjusted according to the classification model loss value, and the sentence classification model to be trained is continued to be trained until the training stop condition is reached to obtain a trained sentence classification model.

[0082] Among them, the sample sentences and the sample classification information corresponding to the sample sentences are training samples of the sentence classification model, wherein the sample sentences are sentences used to input into the sentence classification model, and the sample classification information is the classification information corresponding to the sample sentences, such as entity information and non-entity information.

[0083] The sample sentence is input into the sentence classification model for processing. The sentence classification model identifies whether the sample sentence contains entity information, and outputs predicted classification information based on whether the entity information is identified in the sample sentence. The loss value is calculated by combining the obtained predicted classification information with the sample classification information. The loss function for calculating the loss value can be a 0-1 loss function, an absolute value loss function, a square loss function, a cross entropy loss function, etc. in actual applications. In this application, the choice of loss function is not limited and is subject to actual application. Here, the cross entropy loss function is used as an example for explanation, see the following formula 1:

[0084]

[0085] Among them, C represents the loss value Loss, x represents the sample sentence, y represents the sample classification information, a represents the predicted output, and n represents the total number of samples. After calculating the Loss value, the learning rate parameter, Batch_size parameter, etc. in the sentence classification model are adjusted according to the Loss value back propagation, and the sentence classification model is continued to be trained with the next batch of sample data until the loss value is less than the preset threshold and / or the training rounds reach the preset rounds.

[0086] In a specific implementation provided in the present application, taking the example of stopping the training of the sentence classification model when the loss value is less than a preset threshold, the preset threshold is 0.3. When the calculated Loss value is less than 0.3, it is determined that the sentence classification model training is completed.

[0087] In another specific implementation provided in the present application, taking the preset training rounds to stop training the sentence classification model as an example, the preset training rounds are 10 rounds. When the training rounds of the sample data reach 10 rounds, it is determined that the sentence classification model has been trained.

[0088] In another specific implementation provided in the present application, two training stop conditions, namely a preset threshold of the loss value and a preset number of training rounds, are set, and the loss value and the training rounds are monitored at the same time. When either the loss value or the number of training rounds meets the training stop condition, it is determined that the sentence classification model has been trained.

[0089] Step 206: Determine a target sentence set in the initial sentence set according to the classification information corresponding to each sentence, and determine a target paragraph set according to the target sentence set.

[0090] After obtaining the classification information corresponding to each sentence, the initial sentences can be screened in the initial sentence set according to the classification information corresponding to each sentence, the target sentences can be determined, and the target sentence set can be formed. Then, the target sentences can be spliced ​​into a target paragraph combination according to the position information of each target sentence in the target sentence set. The target sentence set is a set of sentences including entity information after screening in the initial sentence set and determining the classification information of the sentences. The target paragraph is a paragraph composed of target sentences. The target paragraph set is a set composed of at least one target paragraph.

[0091] Specifically, determining a target sentence set in the initial sentence set according to the classification information corresponding to each sentence, and determining a target paragraph set according to the target sentence set, including S2062-S2064:

[0092] S2062, determining an initial sentence whose classification information is a first-category sentence as a target sentence, and obtaining a target sentence set;

[0093] S2064: splice each target sentence according to the position information of each target sentence in the text to be processed to generate a target paragraph set.

[0094] Specifically, each target sentence is concatenated according to the position information of each target sentence in the text to be processed to generate a target paragraph set, including:

[0095] Concatenate the target sentences corresponding to the consecutive adjacent position information to generate a target paragraph;

[0096] A target paragraph set is generated according to each target paragraph.

[0097] Among them, the initial sentences whose classification information is the first category of sentences in the initial sentence set are determined as target sentences, that is, the initial sentences containing entity information are determined as target sentences, and they form a target sentence set. In practical applications, when the text to be processed is segmented, each initial sentence is numbered to identify the position information of each sentence in the text to be processed, and then the target sentences with connected numbers are spliced ​​to form a target paragraph, and the generated target paragraphs are formed into a target paragraph set, that is, the target sentences with adjacent positions are spliced ​​to form a target paragraph, and the target paragraphs are then formed into a target paragraph set.

[0098] Step 208: Input the target paragraph set into a named entity recognition model for processing to obtain a target entity set output by the named entity recognition model.

[0099] The named entity recognition model is also a trained artificial intelligence model. The named entity recognition model can be a Bert model, a Transformer model, etc. In the embodiment of the present application, the named entity recognition model is still explained by taking the Bert model as an example. The named entity recognition model includes an embedding layer, 12 sequentially connected encoding layers and a fully connected layer. The target paragraph is input into the embedding layer to generate a target paragraph vector, and then the target paragraph vector is input into the 12 encoding layers in sequence for processing to obtain a target paragraph encoding vector, and then the target paragraph encoding vector is input into the fully connected layer for processing, and the target entity corresponding to the target paragraph encoding vector is output to finally obtain a target entity set.

[0100] The named entity recognition model is pre-trained and has the ability to output entities corresponding to paragraphs according to paragraphs, wherein the named entity recognition model is trained through the following steps:

[0101] Obtaining a sample paragraph and an annotation entity corresponding to the sample paragraph;

[0102] Inputting the sample paragraph into a named entity recognition model to be trained, so that the named entity recognition model to be trained recognizes entities in the sample paragraph;

[0103] Obtaining a predicted entity corresponding to the sample paragraph output by the named entity recognition model to be trained;

[0104] Calculate a recognition model loss value according to the predicted entity and the labeled entity;

[0105] The model parameters of the named entity recognition model to be trained are adjusted according to the recognition model loss value, and the named entity recognition model to be trained is continued to be trained until the training stop condition is reached to obtain a trained named entity recognition model.

[0106] The sample paragraph is a paragraph consisting of at least one sentence, and the sample paragraph is annotated with an annotation entity. The specific annotation method can be BIO annotation, BIEO annotation, BIOSE annotation, etc. Taking BIO annotation as an example, each character in the paragraph is annotated as "BX", "IX" or "O", where "BX" means that the element belongs to type X in the text and the element is at the beginning of the text, "IX" means that the text where the element belongs to type X and the element is in the middle of the text, and "O" means that it does not belong to any type. For example, the sample paragraph is "The range of A missile is 500 kilometers", which is annotated with BIO as (A: B-arm, guide: I-arm, missile: I-arm, of: O, shot: O, range: O, length: O, 500: O, kilometers: O, miles: O), that is, in the sample paragraph, "A missile" is the annotation entity corresponding to the sample paragraph. "arm" is the entity category of the annotated entity.

[0107] The sample paragraph is input into the named entity recognition model to be trained. The named entity recognition model is used to identify the entity information in the sample paragraph and output the predicted entity corresponding to the sample paragraph. The loss value of the predicted entity and the labeled entity is then calculated. The model parameters of the named entity recognition model are adjusted according to the loss value until the training stop condition is reached to obtain a trained named entity recognition model. The specific contents of the method of calculating the loss value, the model training stop condition, and the back-propagation adjustment of the model parameters can be found in the description of the calculation of the loss value of the sentence classification model above, which will not be repeated here.

[0108] Step 210: Determine the information extraction result of the text to be processed according to the target entity set.

[0109] After obtaining the target entity set, the information extraction result of the text to be processed can be determined according to the target entity set.

[0110] In practical applications, there may be multiple identical entities in the target entity set, for example, the same entity appears in multiple places in the text to be processed. Therefore, the target entity set needs to be screened to obtain the final information extraction result. Specifically, the target entity set includes the target entity, entity type, and sentence number corresponding to the entity.

[0111] Determining the information extraction result of the to-be-processed text according to the target entity set includes:

[0112] The target entity is associated according to the sentence sequence number, entity type and preset rules corresponding to the entity to obtain the information extraction result of the text to be processed.

[0113] The target entity set includes the target entity, the entity type corresponding to the target entity, and the sentence number corresponding to the target entity. The target entity is associated according to the preset rules to obtain the information extraction result of the text to be processed. The preset rules include but are not limited to:

[0114] 1. Entities that appear in the same sentence have a corresponding relationship. For example, in a specific implementation provided by this application, if the two entities "Party A" and "Zhang San" appear in the same sentence, it can be determined that Zhang San is Party A.

[0115] 2. Entities appearing in adjacent sentences have a corresponding relationship. For example, in a specific implementation provided by the present application, the two entities "Party A" and "Zhang San" both appear in the first sentence, "Address A" appears in the second sentence, the two entities "Party B" and "Li Si" both appear in the third sentence, and "Address B" appears in the fourth sentence. It can be determined that Address A is Zhang San's address and Address B is Li Si's address.

[0116] According to the entity type corresponding to each entity, the sentence number corresponding to the entity and the preset rules, the target entity can be associated to form the information extraction result corresponding to the text to be processed.

[0117] The information extraction method provided by the embodiment of the present application includes: performing sentence processing on the received text to be processed to obtain an initial sentence set; inputting the initial sentence set into a sentence classification model for processing to obtain classification information corresponding to each sentence in the initial sentence set output by the sentence classification model; determining a target sentence set in the initial sentence set according to the classification information corresponding to each sentence, and determining a target paragraph set according to the target sentence set; inputting the target paragraph set into a named entity recognition model for processing to obtain a target entity set output by the named entity recognition model; and determining the information extraction result of the text to be processed according to the target entity set. Through this method, sentences that do not contain entity information in the text to be processed are first filtered out through sentence classification, and then entity recognition is performed on sentences containing entity information, thereby avoiding the problem of low extraction efficiency caused by entity extraction of the entire document, and also avoiding the problem of low information extraction accuracy caused by extracting wrong entities in paragraphs that do not contain entities. The information extraction efficiency and accuracy are improved.

[0118] Figure 3 An information extraction method according to an embodiment of the present application is shown. The information extraction method is described by taking information extraction from a contract text as an example, and includes steps 302 to 318.

[0119] Step 302: Receive the contract document to be processed.

[0120] In a specific embodiment provided by the present application, the extraction of information in a contract is used as an example for explanation, and a contract file to be processed is received, and the contract file to be processed is in PDF format. The contract file to be processed is a "free deposit contract for movable property".

[0121] Step 304: Parse the pending contract file to obtain the pending contract text corresponding to the pending contract file.

[0122] In a specific embodiment provided by the present application, after receiving the contract file to be processed, the contract file to be processed is parsed through OCR recognition and document conversion, and the contract file to be processed is converted into the contract text to be processed, and the contract text to be processed is "Contract for Gratuitous Entrustment of Movable Property\nDepositor Zhang San (hereinafter referred to as Party A), Trustee Li Si (hereinafter referred to as Party B). The conditions for the consignment of movable furniture are agreed upon by both parties as follows:\nArticle 1...\nParty A: Zhang San\nAddress: A\nParty B: Li Si\nAddress: B\n****year**month**day", where "\n" indicates a line break in the contract text to be processed.

[0123] Step 306: Obtain a preset sentence identifier, perform sentence processing on the contract text to be processed according to the preset sentence identifier, and obtain an initial sentence set.

[0124] In a specific embodiment provided by the present application, "\n" is used as a preset sentence identifier, and the contract text to be processed is processed according to "\n" to obtain multiple initial sentences to form an initial sentence set. The initial sentence set is shown in Table 1 below.

[0125] Table 1

[0126]

[0127]

[0128] Step 308: Determine a target initial sentence in the initial sentence set, and input the target initial sentence into a sentence classification model for sentence classification.

[0129] In a specific embodiment provided in the present application, each initial sentence in the initial sentence set is used as a target initial sentence and is spliced ​​with "#", and each spliced ​​target initial sentence is input into a pre-trained sentence classification model for sentence classification. The sentence classification model is trained to identify whether the sentence contains entity information based on the sentence.

[0130] Step 310: Obtain sentence classification information corresponding to the target initial sentence output by the sentence classification model, wherein the sentence classification information includes first-category sentences or second-category sentences.

[0131] In a specific embodiment provided in the present application, the sentence classification model outputs sentence classification information corresponding to each target initial sentence after splicing, see Table 2 below, where 1 represents the first category of sentences, indicating that the sentence includes entity information; 0 represents the second category of sentences, indicating that the sentence does not include entity information.

[0132] Table 2

[0133]

[0134]

[0135] Step 312: Determine the initial sentences whose sentence classification information is the first category of sentences as target sentences, and obtain a set of target sentences.

[0136] In a specific embodiment provided by the present application, according to the sentence classification information in Table 2 above, the initial sentence with the sentence classification information of 1 is determined as the target sentence, the initial sentence with the sentence classification information of 0 is filtered, and only the initial sentence with the sentence classification information of 1 is retained to form the target sentence set. The target sentence set is shown in Table 3 below.

[0137] Table 3

[0138]

[0139]

[0140] Step 314: splicing each target sentence according to the position information of each target sentence set to generate a target paragraph set.

[0141] In a specific embodiment provided in the present application, according to the sentence serial numbers in the target sentence set, the target sentences corresponding to the consecutive serial numbers are concatenated to obtain a target paragraph set. The target paragraph set is shown in Table 4 below.

[0142] Table 4

[0143]

[0144] Step 316: Input the target paragraph set into the named entity recognition model for processing to obtain a target entity set output by the named entity recognition model, wherein the target entity set includes a target entity, an entity type, and a sentence number corresponding to the entity.

[0145] In a specific embodiment provided in the present application, a target paragraph in a target paragraph set is input into a named entity recognition model, and the named entity recognition model outputs a corresponding target entity set according to each target paragraph in the target paragraph set, see Table 5 below.

[0146] Table 5

[0147]

[0148]

[0149] Step 318: Associating the target entity according to the sentence number, entity type and preset rules corresponding to the entity, and obtaining the information extraction result of the contract text to be processed.

[0150] In a specific embodiment provided in the present application, it can be determined through the target entity set that the two entities "Party A" and "Zhang San" both appear in the n-4th sentence, and the two entities "Party B" and "Li Si" both appear in the n-2th sentence. Therefore, it can be determined that Zhang San is Party A and Li Si is Party B.

[0151] At the same time, A in the n-3th sentence is an address, B in the n-1th sentence is an address, the n-3th sentence is after Party A's information, and the n-2th sentence is after Party B's information. Therefore, it can be determined that address A is Party A's address and address B is Party B's address.

[0152] Based on this, it is determined that the information extracted from the text to be processed is "Contract for Gratuitous Entrustment of Movable Property (Contract Name), ** Month ** Day **** (Signing Date), Zhang San (Party A), A (Party A's Address), Li Si (Party B), B (Party B's Address)".

[0153] The information extraction method provided in the embodiment of the present application is applied to the information extraction of the contract text. Through this method, the sentences that do not contain entity information in the text to be processed are first filtered out through sentence classification, and then the sentences containing entity information are entity recognized, thereby avoiding the problem of low extraction efficiency caused by entity extraction of the entire document, and also avoiding the problem of low information extraction accuracy caused by extracting wrong entities in paragraphs that do not contain entities. The information extraction efficiency and accuracy are improved.

[0154] Corresponding to the above method embodiment, the present application also provides an information extraction device embodiment, Figure 4 FIG. 1 is a schematic diagram showing the structure of an information extraction device according to an embodiment of the present application. Figure 4 As shown, the device comprises:

[0155] The sentence segmentation module 402 is configured to perform sentence segmentation processing on the received text to be processed to obtain an initial sentence set;

[0156] The sentence classification module 404 is configured to obtain classification information corresponding to each sentence in the initial sentence set;

[0157] A determination module 406 is configured to determine a target sentence set in the initial sentence set according to the classification information corresponding to each sentence, and determine a target paragraph set according to the target sentence set;

[0158] The entity recognition module 408 is configured to input the target paragraph set into a named entity recognition model for processing, and obtain a target entity set output by the named entity recognition model;

[0159] The information extraction module 410 is configured to determine the information extraction result of the text to be processed according to the target entity set.

[0160] Optionally, the sentence segmentation module 402 is further configured to:

[0161] Get the preset sentence identifier;

[0162] The text to be processed is processed by sentence segmentation according to the preset sentence segmentation identifier to obtain an initial sentence set.

[0163] Optionally, the sentence classification module 404 is further configured to:

[0164] The initial sentence set is input into a sentence classification model for processing, and classification information corresponding to each sentence in the initial sentence set output by the sentence classification model is obtained.

[0165] Optionally, the sentence classification module 404 is further configured to:

[0166] Determining a target initial sentence from the initial sentence set, and inputting the target initial sentence into a sentence classification model for sentence classification, wherein the sentence classification model is trained to determine whether a sentence contains entity information;

[0167] In the case where the target initial sentence contains entity information, the classification information corresponding to the target initial sentence output by the sentence classification model is obtained as a first-category sentence;

[0168] In the case that the target initial sentence does not contain entity information, the classification information corresponding to the target initial sentence output by the sentence classification model is obtained as a second-category sentence.

[0169] Optionally, the determining module 406 is further configured to:

[0170] Determine the initial sentence whose classification information is the first category sentence as the target sentence, and obtain the target sentence set;

[0171] Each target sentence is concatenated according to the position information of each target sentence in the text to be processed to generate a target paragraph set.

[0172] Optionally, the determining module 406 is further configured to:

[0173] Concatenate the target sentences corresponding to the consecutive adjacent position information to generate a target paragraph;

[0174] A target paragraph set is generated according to each target paragraph.

[0175] Optionally, the target entity set includes a target entity, an entity type, and a sentence sequence number corresponding to the entity;

[0176] The information extraction module 410 is further configured to:

[0177] The target entity is associated according to the sentence sequence number, entity type and preset rules corresponding to the entity to obtain the information extraction result of the text to be processed.

[0178] Optionally, the device further comprises:

[0179] A receiving module, configured to receive a file to be processed;

[0180] The parsing module is configured to parse the file to be processed and obtain the text to be processed corresponding to the file to be processed.

[0181] Optionally, the device further comprises:

[0182] The sentence classification model training module is configured as follows:

[0183] Obtaining sample sentences and sample classification information corresponding to the sample sentences;

[0184] Inputting the sample sentence into a sentence classification model to be trained, so that the sentence classification model to be trained determines whether the sample sentence contains entity information;

[0185] Obtaining predicted classification information corresponding to the sample sentence output by the sentence classification model to be trained;

[0186] Calculate the classification model loss value according to the predicted classification information and the sample classification information;

[0187] The model parameters of the sentence classification model to be trained are adjusted according to the classification model loss value, and the sentence classification model to be trained is continued to be trained until the training stop condition is reached to obtain a trained sentence classification model.

[0188] Optionally, the device further comprises:

[0189] The named entity recognition model training module is configured as follows:

[0190] Obtaining a sample paragraph and an annotation entity corresponding to the sample paragraph;

[0191] Inputting the sample paragraph into a named entity recognition model to be trained, so that the named entity recognition model to be trained recognizes entities in the sample paragraph;

[0192] Obtaining a predicted entity corresponding to the sample paragraph output by the named entity recognition model to be trained;

[0193] Calculate a recognition model loss value according to the predicted entity and the labeled entity;

[0194] The model parameters of the named entity recognition model to be trained are adjusted according to the recognition model loss value, and the named entity recognition model to be trained is continued to be trained until the training stop condition is reached to obtain a trained named entity recognition model.

[0195] The information extraction device provided by the embodiment of the present application includes: performing sentence segmentation processing on the received text to be processed to obtain an initial sentence set; inputting the initial sentence set into a sentence classification model for processing to obtain classification information corresponding to each sentence in the initial sentence set output by the sentence classification model; determining a target sentence set in the initial sentence set according to the classification information corresponding to each sentence, and determining a target paragraph set according to the target sentence set; inputting the target paragraph set into a named entity recognition model for processing to obtain a target entity set output by the named entity recognition model; and determining the information extraction result of the text to be processed according to the target entity set. Through this device, sentences that do not contain entity information in the text to be processed are first filtered out through sentence classification, and then entity recognition is performed on sentences containing entity information, thereby avoiding the problem of low extraction efficiency caused by entity extraction of the entire document, and also avoiding the problem of low information extraction accuracy caused by extracting wrong entities in paragraphs that do not contain entities. The information extraction efficiency and accuracy are improved.

[0196] The above is a schematic scheme of an information extraction device of this embodiment. It should be noted that the technical scheme of the information extraction device and the technical scheme of the above information extraction method belong to the same concept, and the details not described in detail in the technical scheme of the information extraction device can be referred to the description of the technical scheme of the above information extraction method.

[0197] It should be noted that the components in the device claim should be understood as the functional modules that must be established to implement the steps of the program flow or the steps of the method, and the functional modules are not actually divided or separated. The device claim defined by such a group of functional modules should be understood as the functional module architecture of the solution mainly implemented by the computer program recorded in the specification, and should not be understood as the physical device that mainly implements the solution by hardware.

[0198] In one embodiment of the present application, a computing device is also provided, including a memory, a processor, and computer instructions stored in the memory and executable on the processor, wherein the processor implements the steps of the information extraction method when executing the computer instructions.

[0199] The above is a schematic scheme of a computing device of this embodiment. It should be noted that the technical scheme of the computing device and the technical scheme of the above information extraction method belong to the same concept, and the details not described in detail in the technical scheme of the computing device can be referred to the description of the technical scheme of the above information extraction method.

[0200] An embodiment of the present application further provides a computer-readable storage medium storing computer instructions, which implement the steps of the information extraction method as described above when executed by a processor.

[0201] The above is a schematic scheme of a computer-readable storage medium of this embodiment. It should be noted that the technical scheme of the storage medium and the technical scheme of the above information extraction method belong to the same concept, and the details not described in detail in the technical scheme of the storage medium can be referred to the description of the technical scheme of the above information extraction method.

[0202] An embodiment of the present application discloses a chip storing computer instructions, which implement the steps of the information extraction method as described above when executed by a processor.

[0203] The above describes specific embodiments of the present application. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the embodiments and still achieve the desired results. In addition, 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.

[0204] The computer instructions include computer program codes, which may be in source code form, object code form, executable files or some intermediate forms, etc. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal and software distribution medium, etc. It should be noted that the content contained in the computer-readable medium may be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electric carrier signals and telecommunication signals.

[0205] It should be noted that, for the above-mentioned method embodiments, for the sake of simplicity of description, they are all expressed as a series of action combinations, but those skilled in the art should be aware that the present application is not limited by the described action sequence, because according to the present application, certain steps can be performed in other sequences or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the present application.

[0206] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0207] The preferred embodiments of the present application disclosed above are only used to help explain the present application. The optional embodiments do not describe all the details in detail, nor do they limit the invention to the specific implementation methods described. Obviously, many modifications and changes can be made according to the content of the present application. The present application selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present application, so that those skilled in the art can understand and use the present application well. The present application is only limited by the claims and their full scope and equivalents.

Claims

1. An information extraction method, characterized in that: include: Sentence processing is performed on the received text to be processed to obtain an initial sentence set; Obtaining classification information corresponding to each sentence in the initial sentence set, wherein the classification information includes first-category sentences and second-category sentences, the first-category sentences represent sentences that include entity information, and the second-category sentences represent sentences that do not include entity information; Determine a target sentence set in the initial sentence set according to the classification information corresponding to each sentence, and determine a target paragraph set according to the target sentence set; Inputting the target paragraph set into a named entity recognition model for processing to obtain a target entity set output by the named entity recognition model; The information extraction result of the text to be processed is determined according to the target entity set.

2. The information extraction method according to claim 1, characterized in that: The received text to be processed is processed into sentences to obtain an initial sentence set, including: Get the preset sentence identifier; The text to be processed is processed by sentence segmentation according to the preset sentence segmentation identifier to obtain an initial sentence set.

3. The information extraction method according to claim 1, characterized in that: Obtaining classification information corresponding to each sentence in the initial sentence set, including: The initial sentence set is input into a sentence classification model for processing, and classification information corresponding to each sentence in the initial sentence set output by the sentence classification model is obtained.

4. The information extraction method according to claim 3, characterized in that: Inputting the initial sentence set into a sentence classification model for processing, and obtaining classification information corresponding to each sentence in the initial sentence set output by the sentence classification model, including: Determining a target initial sentence from the initial sentence set, and inputting the target initial sentence into a sentence classification model for sentence classification, wherein the sentence classification model is trained to determine whether a sentence contains entity information; In the case where the target initial sentence contains entity information, the classification information corresponding to the target initial sentence output by the sentence classification model is obtained as a first-category sentence; In the case that the target initial sentence does not contain entity information, the classification information corresponding to the target initial sentence output by the sentence classification model is obtained as a second-category sentence.

5. The information extraction method according to claim 4, characterized in that: Determining a target sentence set in the initial sentence set according to classification information corresponding to each sentence, and determining a target paragraph set according to the target sentence set, including: Determine the initial sentence whose classification information is the first category sentence as the target sentence, and obtain the target sentence set; Each target sentence is concatenated according to the position information of each target sentence in the text to be processed to generate a target paragraph set.

6. The information extraction method according to claim 5, characterized in that: Each target sentence is concatenated according to the position information of each target sentence in the text to be processed to generate a target paragraph set, including: Concatenate the target sentences corresponding to the consecutive adjacent position information to generate a target paragraph; A target paragraph set is generated according to each target paragraph.

7. The information extraction method according to claim 1, characterized in that: The target entity set includes the target entity, entity type and sentence sequence number corresponding to the entity; Determining the information extraction result of the to-be-processed text according to the target entity set includes: The target entity is associated according to the sentence sequence number, entity type and preset rules corresponding to the entity to obtain the information extraction result of the text to be processed.

8. The information extraction method according to any one of claims 1 to 7, characterized in that: The method further comprises: Receive files to be processed; Parse the file to be processed to obtain the text to be processed corresponding to the file to be processed.

9. The information extraction method according to claim 3, characterized in that: The sentence classification model is trained by the following steps: Obtaining sample sentences and sample classification information corresponding to the sample sentences; Inputting the sample sentence into a sentence classification model to be trained, so that the sentence classification model to be trained determines whether the sample sentence contains entity information; Obtaining predicted classification information corresponding to the sample sentence output by the sentence classification model to be trained; Calculate the classification model loss value according to the predicted classification information and the sample classification information; The model parameters of the sentence classification model to be trained are adjusted according to the classification model loss value, and the sentence classification model to be trained is continued to be trained until the training stop condition is reached to obtain a trained sentence classification model.

10. The information extraction method according to claim 1, characterized in that: The named entity recognition model is trained by the following steps: Obtaining a sample paragraph and an annotation entity corresponding to the sample paragraph; Inputting the sample paragraph into a named entity recognition model to be trained, so that the named entity recognition model to be trained recognizes entities in the sample paragraph; Obtaining a predicted entity corresponding to the sample paragraph output by the named entity recognition model to be trained; Calculate a recognition model loss value according to the predicted entity and the labeled entity; The model parameters of the named entity recognition model to be trained are adjusted according to the recognition model loss value, and the named entity recognition model to be trained is continued to be trained until the training stop condition is reached to obtain a trained named entity recognition model.

11. An information extraction device, characterized in that: include: A sentence segmentation module is configured to perform sentence segmentation processing on the received text to be processed to obtain an initial sentence set; A sentence classification module is configured to obtain classification information corresponding to each sentence in the initial sentence set, wherein the classification information includes first-category sentences and second-category sentences, the first-category sentences represent sentences that include entity information, and the second-category sentences represent sentences that do not include entity information; A determination module, configured to determine a target sentence set in the initial sentence set according to classification information corresponding to each sentence, and determine a target paragraph set according to the target sentence set; An entity recognition module is configured to input the target paragraph set into a named entity recognition model for processing, and obtain a target entity set output by the named entity recognition model; The information extraction module is configured to determine the information extraction result of the text to be processed according to the target entity set.

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

13. A computer-readable storage medium storing computer instructions, characterized in that: When the computer instructions are executed by a processor, the steps of the method described in any one of claims 1 to 10 are implemented.

Images