Information extraction method, device, computer equipment and storage medium

Abstract

The present application relates to an information extraction method, device, computer equipment and storage medium. The method comprises: obtaining a target question and a target content; inputting the target question and the target content into a pre-trained information extraction model for information extraction, and obtaining target information that answers the target question output by the information extraction model; wherein the information extraction model is used to determine the target probability of extracting information from a preset vocabulary according to the target question and the target content, and extract the target information from the preset vocabulary or the target content according to the target probability. The use of this method can reduce the difficulty of information extraction and improve the accuracy of information extraction.

Description

Technical Field

[0001] The present application relates to the technical field of information extraction, and in particular to an information extraction method, apparatus, computer equipment and storage medium. Background Art

[0002] Acquiring user information is an essential step in conducting business in the insurance, banking and other industries. Extracting this important information quickly and accurately is very important for improving corporate development, service quality and business indicators.

[0003] In the related art, a model is usually trained based on training samples and manual annotations corresponding to the training samples to obtain an information extraction model. Then, the document from which information is to be extracted is input into the information extraction model, and the information extraction model extracts information from the document.

[0004] However, the above information extraction method has the following problems: it is difficult for the information extraction model to extract information from the document, or the information extracted from the document by the information extraction model is inaccurate. Summary of the invention

[0005] Based on this, it is necessary to provide an information extraction method, device, computer equipment and storage medium that can reduce the difficulty of information extraction and improve the accuracy of information extraction in response to the above technical problems.

[0006] An information extraction method, the method comprising:

[0007] Get the target questions and target content;

[0008] Input the target question and target content into the pre-trained information extraction model to extract information, and obtain the target information that answers the target question output by the information extraction model;

[0009] Among them, the information extraction model is used to determine the target probability of extracting information from a preset vocabulary according to the target question and the target content, and extract the target information from the preset vocabulary or the target content according to the target probability.

[0010] In one embodiment, the information extraction model includes an encoding sub-model and a decoding sub-model, and the target probability of extracting information from a preset vocabulary according to the target question and the target content is determined, and the target information is extracted from the preset vocabulary or the target content according to the target probability, including:

[0011] The target question and target content are encoded using the encoding sub-model to obtain a feature vector;

[0012] The feature vector is subjected to multiple rounds of decoding processing using the decoding submodel to obtain the target probability corresponding to each round of decoding, the first probability distribution corresponding to the preset vocabulary, and the second probability distribution corresponding to the feature vector; wherein the first probability distribution is used to represent the first extraction probability corresponding to each word in the preset vocabulary, and the second probability distribution is used to represent the second extraction probability corresponding to each word in the target content;

[0013] Determine a third probability distribution corresponding to each round of decoding according to the target probability corresponding to each round of decoding, the first probability distribution, and the second probability distribution; the third probability distribution is used to characterize the third extraction probability corresponding to each word in the preset word list and the target content respectively;

[0014] Information extraction is performed according to the third probability distribution corresponding to each round of decoding to obtain multiple candidate information, and target information is screened out from the candidate information.

[0015] In one embodiment, the feature vector includes a classification feature vector, a question feature vector, and a content feature vector, and the decoding sub-model includes a bidirectional long short-term memory network, a first fully connected layer, and a second fully connected layer. The decoding sub-model is used to perform multiple rounds of decoding processing on the feature vector to obtain a target probability corresponding to each round of decoding, a first probability distribution corresponding to a preset vocabulary, and a second probability distribution corresponding to the feature vector, including:

[0016] For each round of decoding, the classification feature vector is concatenated with the network feature vector corresponding to the current round of the bidirectional long short-term memory network to obtain a concatenated vector, and the concatenated vector is input into the first fully connected layer to obtain the target probability;

[0017] Inputting the classification feature vector into the second fully connected layer to obtain a first probability distribution;

[0018] Regularization is performed on the question feature vector and the content feature vector, and a second probability distribution is obtained according to the regularization result.

[0019] In one embodiment, determining the third probability distribution corresponding to each round of decoding according to the target probability corresponding to each round of decoding, the first probability distribution, and the second probability distribution includes:

[0020] For each round of decoding, determining a first weight corresponding to the first probability distribution and a second weight corresponding to the second probability distribution according to the target probability;

[0021] The first probability distribution and the second probability distribution are weightedly summed according to the first weight and the second weight to obtain a third probability distribution.

[0022] In one embodiment, the information extraction according to the third probability distribution corresponding to each round of decoding obtains multiple candidate information, including:

[0023] For each round of decoding, the words with the third highest probability of being extracted in the preset word list and the target content are determined according to the third probability distribution, and the words with the third highest probability of being extracted are determined as candidate information.

[0024] In one embodiment, the encoding sub-model is used to encode the target question and the target content to obtain a feature vector, including:

[0025] Perform splicing processing on the target problem and the target content to obtain splicing information;

[0026] Performing word segmentation processing on the spliced ​​information to obtain multiple information segments; wherein the information segments include classification identification segments;

[0027] Multiple information segments are input into the encoding sub-model for encoding processing to obtain feature vectors corresponding to each information segment output by the encoding sub-model; wherein the classification identification segment corresponds to the classification feature vector.

[0028] In one embodiment, before inputting the target question and the target content into the pre-trained information extraction model for information extraction, the method further includes:

[0029] Use crawler tools to obtain general corpus from the Internet; general corpus consists of questions, content and answers;

[0030] Perform data processing on general corpus to obtain training questions, training content and training labels;

[0031] The model is trained according to the training questions, training content and training labels to obtain the information extraction model.

[0032] An information extraction device, comprising:

[0033] A question content acquisition module is used to acquire target questions and target content;

[0034] An information extraction module is used to input the target question and target content into a pre-trained information extraction model to extract information, and obtain the target information that answers the target question output by the information extraction model;

[0035] Among them, the information extraction model is used to determine the target probability of extracting information from a preset vocabulary according to the target question and the target content, and extract the target information from the preset vocabulary or the target content according to the target probability.

[0036] In one embodiment, the information extraction model includes an encoding sub-model and a decoding sub-model, and the information extraction module includes:

[0037] The encoding submodule is used to encode the target question and target content using the encoding submodel to obtain a feature vector;

[0038] A decoding submodule, used to perform multiple rounds of decoding processing on the feature vector using the decoding submodel, to obtain the target probability corresponding to each round of decoding, the first probability distribution corresponding to the preset vocabulary, and the second probability distribution corresponding to the feature vector; wherein the first probability distribution is used to represent the first extraction probability corresponding to each word in the preset vocabulary, and the second probability distribution is used to represent the second extraction probability corresponding to each word in the target content;

[0039] A probability determination submodule, which determines a third probability distribution corresponding to each round of decoding according to the target probability corresponding to each round of decoding, the first probability distribution, and the second probability distribution; the third probability distribution is used to characterize the third extraction probability corresponding to each word in the preset word list and the target content respectively;

[0040] The information extraction submodule is used to extract information according to the third probability distribution corresponding to each round of decoding to obtain multiple candidate information, and filter out target information from the candidate information.

[0041] In one embodiment, the feature vector includes a classification feature vector, a question feature vector and a content feature vector, and the decoding sub-model includes a bidirectional long short-term memory network, a first fully connected layer and a second fully connected layer. The above-mentioned decoding sub-module is specifically used to, for each round of decoding, concatenate the classification feature vector with the network feature vector corresponding to the current round of the bidirectional long short-term memory network to obtain a concatenated vector, and input the concatenated vector into the first fully connected layer to obtain a target probability; input the classification feature vector into the second fully connected layer to obtain a first probability distribution; regularize the question feature vector and the content feature vector, and obtain a second probability distribution based on the regularization result.

[0042] In one of the embodiments, the above-mentioned probability determination submodule is specifically used to determine, for each round of decoding, a first weight corresponding to the first probability distribution and a second weight corresponding to the second probability distribution according to the target probability; and perform weighted summation processing on the first probability distribution and the second probability distribution according to the first weight and the second weight to obtain a third probability distribution.

[0043] In one embodiment, the information extraction submodule is specifically used to determine the words with the third highest probability of being extracted in the preset word list and the target content according to the third probability distribution for each round of decoding, and determine the words with the third highest probability of being extracted as candidate information.

[0044] In one of the embodiments, the above-mentioned encoding submodule is specifically used to splice the target question and the target content to obtain splicing information; perform word segmentation on the splicing information to obtain multiple information segments; wherein the information segments include classification identification segments; input the multiple information segments into the encoding submodel for encoding processing to obtain feature vectors corresponding to each information segment output by the encoding submodel; wherein the classification identification segment corresponds to the classification feature vector.

[0045] In one embodiment, the device further comprises:

[0046] The corpus acquisition module is used to acquire general corpus from the Internet using crawler tools; general corpus consists of questions, content and answers;

[0047] Corpus processing module, used to process the general corpus to obtain training questions, training content and training labels;

[0048] The training module is used to train the model according to the training questions, training content and training labels to obtain the information extraction model.

[0049] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:

[0050] Get the target questions and target content;

[0051] Input the target question and target content into the pre-trained information extraction model to extract information, and obtain the target information that answers the target question output by the information extraction model;

[0052] Among them, the information extraction model is used to determine the target probability of extracting information from a preset vocabulary according to the target question and the target content, and extract the target information from the preset vocabulary or the target content according to the target probability.

[0053] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the following steps:

[0054] Get the target questions and target content;

[0055] Input the target question and target content into the pre-trained information extraction model to extract information, and obtain the target information that answers the target question output by the information extraction model;

[0056] Among them, the information extraction model is used to determine the target probability of extracting information from a preset vocabulary according to the target question and the target content, and extract the target information from the preset vocabulary or the target content according to the target probability.

[0057] The above-mentioned information extraction method, device, computer equipment and storage medium obtain the target question and target content; input the target question and target content into the pre-trained information extraction model for information extraction, and obtain the target information that answers the target question output by the information extraction model. Since the information extraction model first determines the target probability of extracting information from the preset vocabulary according to the target question and target content, and then extracts the target information from the preset vocabulary or target content according to the target probability, it can avoid the problem of difficulty in extracting information from the target content, reducing the difficulty of information extraction; and the preset vocabulary provides information other than the target content, so it can also improve the accuracy of information extraction. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 is a flow chart of an information extraction method in one embodiment;

[0059] Figure 2 A schematic diagram of a process of extracting target information in one embodiment;

[0060] Figure 3 A schematic diagram of a flow chart of encoding processing steps in one embodiment;

[0061] Figure 4 A schematic diagram of a decoding process in one embodiment;

[0062] Figure 5 A schematic diagram of a flow chart of a step of determining a third probability distribution in one embodiment;

[0063] Figure 6 A schematic diagram of a flow chart of a model training step in an embodiment;

[0064] Figure 7 is a structural block diagram of an information extraction device in an embodiment;

[0065] Figure 8 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0066] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0067] This application provides an information extraction method, such as Figure 1As shown, the method is applied to a terminal for example. It is understandable that the method can also be applied to a server, and can also be applied to a system including a terminal and a server, and is implemented through the interaction between the terminal and the server. Among them, the terminal can be but not limited to various personal computers, laptops, smart phones and tablets, and the server can be implemented as an independent server or a server cluster composed of multiple servers. In the embodiment of the present disclosure, the following steps can be included:

[0068] Step 101, obtaining target questions and target content.

[0069] Among them, the target question is used to indicate the target of information extraction, and the target content is used to indicate the basis for information extraction. For example, the target question is "What color clothes make people look whiter", and the target content includes "black is a versatile color... bright red can make people look white..."

[0070] The terminal can obtain the target question and target content input by the user, or can obtain the target question and target content pre-stored in the terminal; or can obtain the target question input by the user and the pre-stored target content. The disclosed embodiment does not limit the method of obtaining the target question and target content.

[0071] Step 102, input the target question and target content into a pre-trained information extraction model to extract information, and obtain the target information that answers the target question output by the information extraction model.

[0072] The information extraction model is used to determine the target probability of extracting information from the preset vocabulary according to the target question and the target content, and extract the target information from the preset vocabulary or the target content according to the target probability. The preset vocabulary includes multiple words, and the preset vocabulary can be constructed according to a specific technical field (such as the financial field, the insurance field), or according to a general scenario, which is not limited in the embodiments of the present disclosure.

[0073] The terminal is provided with a pre-trained information extraction model and a preset vocabulary. After obtaining the target question and target content, the target question and target content are input into the information extraction model, and the information extraction model determines the target probability of extracting information from the preset vocabulary according to the target question and target content. If the target probability is greater than the preset probability threshold, the target information is extracted from the preset vocabulary; if the target probability is less than or equal to the preset probability threshold, the target information is extracted from the target content.

[0074] For example, the preset probability threshold is 70%. If the information extraction model determines that the target probability is 80% based on the target question and the target content, the target probability is determined to be greater than the preset probability threshold, and the target information is extracted from the preset vocabulary; if the information extraction model determines that the target probability is 60% based on the target question and the target content, the target probability is determined to be less than the preset probability threshold, and the target information is extracted from the target content.

[0075] In the above information extraction method, a target question and a target content are obtained; the target question and the target content are input into a pre-trained information extraction model for information extraction, and the target information that answers the target question is output by the information extraction model. Since the information extraction model first determines the target probability of extracting information from a preset vocabulary based on the target question and the target content, and then extracts the target information from the preset vocabulary or the target content based on the target probability, the problem of difficulty in extracting information from the target content can be avoided, and the difficulty of information extraction is reduced; and the preset vocabulary provides information other than the target content, so the accuracy of information extraction can also be improved.

[0076] In one embodiment, the information extraction model includes an encoding sub-model and a decoding sub-model, such as Figure 2 As shown, the above step of determining the target probability of extracting information from the preset vocabulary according to the target question and the target content, and extracting the target information from the preset vocabulary or the target content according to the target probability may include:

[0077] Step 201: Encode the target question and target content using the encoding sub-model to obtain a feature vector.

[0078] The feature vector includes a question feature vector and a content feature vector.

[0079] The target question and target content are input into the encoding sub-model in the information extraction model. The encoding sub-model encodes the target question and target content to obtain a question feature vector corresponding to the target question and a content feature vector corresponding to the target content.

[0080] Step 202, using the decoding sub-model to perform multiple rounds of decoding processing on the feature vector, to obtain the target probability corresponding to each round of decoding, the first probability distribution corresponding to the preset vocabulary and the second probability distribution corresponding to the feature vector.

[0081] Among them, the first probability distribution is used to represent the first extraction probability corresponding to each word in the preset vocabulary, and the second probability distribution is used to represent the second extraction probability corresponding to each word in the target content; the first extraction probability and the second extraction probability are both used to represent the probability value of the word being extracted.

[0082] The decoding sub-model can perform multiple rounds of decoding processing on the feature vector in a variety of ways, one of which includes: the decoding sub-model can perform N rounds of decoding processing on the feature vector, N is the maximum decoding rounds preset by the user, and N is a positive integer. In each round of decoding, the decoding sub-model can obtain the target probability, the first probability distribution corresponding to the preset vocabulary, and the second probability distribution corresponding to the feature vector.

[0083] For example, in the first round of decoding, the decoding sub-model obtains a target probability of 80%, the first extraction probability of word 1 in the preset vocabulary is a1, the first extraction probability of word 2 is a2..., the second extraction probability of word 1 in the target content is b1, the second extraction probability of word 2 is b2...

[0084] Another method includes: the decoding sub-model performs decoding processing on the feature vector, and if an end identifier is generated, the decoding processing is terminated. The embodiment of the present disclosure does not limit the number of rounds of decoding processing.

[0085] Step 203: Determine a third probability distribution corresponding to each round of decoding according to the target probability corresponding to each round of decoding, the first probability distribution, and the second probability distribution.

[0086] The third probability distribution is used to represent the third extraction probability corresponding to each word in the preset word list and the target content. The third extraction probability is used to represent the probability value of the word being extracted.

[0087] In each round of decoding, the first probability distribution and the second probability distribution are normalized according to the target probability to obtain the third extraction probability corresponding to each word in the preset vocabulary and the third extraction probability corresponding to each word in the target content.

[0088] Step 204: extract information according to the third probability distribution corresponding to each round of decoding to obtain multiple candidate information, and select target information from the candidate information.

[0089] After each round of decoding, information extraction can be performed according to the third extraction probabilities corresponding to each word in the preset vocabulary and the third extraction probabilities corresponding to each word in the target content to obtain at least one candidate information.

[0090] In one embodiment, for each round of decoding, the words with the third highest probability of being extracted in the preset word list and the target content are determined according to the third probability distribution, and the words with the third highest probability of being extracted are determined as candidate information.

[0091] For example, the words in the preset word list and the words in the target content are sorted from large to small according to the probability according to the third extraction probability, and then the word with the largest third extraction probability is determined as the candidate information.

[0092] In practical applications, words whose third extraction probability is greater than a preset extraction threshold value may also be determined as candidate information. The embodiment of the present disclosure does not limit the method for extracting candidate information.

[0093] After each round of decoding, at least one candidate information can be obtained. After multiple rounds of decoding, multiple candidate information can be obtained. After that, the terminal screens out the target information from the multiple candidate information. Specifically, the candidate information with the most occurrences can be determined as the target information, or the candidate information with the third highest probability of being extracted can be determined as the target information. The present disclosure embodiment does not limit the screening method.

[0094] In the above embodiment, the terminal uses the encoding sub-model to encode the target question and the target content to obtain a feature vector; uses the decoding sub-model to perform multiple rounds of decoding on the feature vector to obtain the target probability corresponding to each round of decoding, the first probability distribution corresponding to the preset vocabulary, and the second probability distribution corresponding to the feature vector; determines the third probability distribution corresponding to each round of decoding according to the target probability, the first probability distribution, and the second probability distribution corresponding to each round of decoding; extracts information according to the third probability distribution corresponding to each round of decoding to obtain multiple candidate information, and selects the target information from the candidate information. Through the embodiment of the present disclosure, the encoding sub-model and the decoding sub-model of the information extraction model perform encoding and decoding processing to determine whether to extract information from the preset vocabulary or the target content. Since the preset vocabulary provides information other than the target content, the difficulty of information extraction can be reduced and the accuracy of information extraction can be improved.

[0095] In one embodiment, Figure 3 As shown, the step of encoding the target question and the target content using the encoding sub-model to obtain the feature vector may include:

[0096] Step 2011, splicing the target question and the target content to obtain splicing information.

[0097] After the target question and the target content are input into the information extraction model, the information extraction model performs concatenation processing on the target question and the target content. Specifically, a classification mark [cls] is added before the target question, and a sentence mark [sep] is added between the target question and the target content.

[0098] For example, the spliced ​​information includes: [cls] What color clothes make you look whiter [sep] Black is a versatile color... Bright red can make people look fairer...

[0099] In practical applications, other splicing methods may also be used, which are not limited in the embodiments of the present disclosure.

[0100] Step 2012, performing word segmentation processing on the spliced ​​information to obtain multiple information fragments.

[0101] The information extraction model performs word segmentation processing on the spliced ​​information according to preset rules to obtain multiple information fragments, for example, a classification identifier fragment corresponding to the classification identifier, a question information fragment corresponding to the target question, and a content information fragment corresponding to the target content.

[0102] If the text of the target question is long, multiple question information segments can be obtained; if the text of the target content is long, multiple content information segments can also be obtained.

[0103] The above-mentioned preset rules may include word length, Euclidean distance between words, etc., and the present disclosure embodiment does not limit the preset rules. It can be understood that the number of information fragments obtained by word segmentation processing varies with the actual word segmentation situation, and the present disclosure embodiment does not limit this.

[0104] Step 2013: Input multiple information segments into the encoding sub-model for encoding processing, and obtain the feature vector corresponding to each information segment output by the encoding sub-model.

[0105] Among them, the encoding sub-model can be a BERT (Bidirectional Encoder Representation from Transformers) model.

[0106] After the information extraction model obtains multiple information fragments through word segmentation, the multiple information fragments are input into the encoding sub-model. The encoding sub-model encodes each information fragment and outputs a feature vector corresponding to each information fragment.

[0107] For example, the encoding sub-model outputs a classification feature vector H_cls corresponding to the classification identification segment, a question feature vector H_q corresponding to the question information segment, and a content feature vector H_p corresponding to the content information segment.

[0108] In the above embodiment, the terminal performs splicing processing on the target question and the target content to obtain splicing information; performs word segmentation processing on the splicing information to obtain multiple information fragments; and inputs the multiple information fragments into the encoding sub-model for encoding processing to obtain feature vectors corresponding to each information fragment output by the encoding sub-model. Through the disclosed embodiment, the information extraction model performs splicing processing, word segmentation processing, and encoding processing on the target question and the target content to obtain feature vectors corresponding to the target question and the target content, respectively, so that the target probability of extracting information from the preset word list can be determined later, and then the target information can be determined.

[0109] In one embodiment, the feature vector includes a classification feature vector, a question feature vector, and a content feature vector; the decoding sub-model includes a bidirectional long short-term memory network (Bi Long Short-Term Memory, BiLSTM), a first fully connected layer and a second fully connected layer, such as Figure 4 As shown, the above-mentioned step of performing multiple rounds of decoding processing on the feature vector using the decoding sub-model to obtain the target probability corresponding to each round of decoding, the first probability distribution corresponding to the preset vocabulary, and the second probability distribution corresponding to the feature vector may include:

[0110] Step 2021, for each round of decoding, the classification feature vector is concatenated with the network feature vector corresponding to the current round of the bidirectional long short-term memory network to obtain a concatenated vector, and the concatenated vector is input into the first fully connected layer to obtain the target probability.

[0111] In each round of decoding, the decoding sub-model concatenates the classification feature vector with the network feature vector corresponding to the current round of BiLSTM to obtain a concatenated vector, and inputs the concatenated vector into the first fully connected layer; then, the output result of the first fully connected layer is activated using the first preset activation function to obtain the target probability.

[0112] For example, in the first round of decoding, the decoding sub-model concatenates the classification feature vector H_cls1 with the network feature vector Ht1 corresponding to the first round of BiLSTM to obtain the concatenated vector H1', ​​inputs the concatenated vector H1' into the first fully connected layer, and uses the first preset activation function to activate the first round output result of the first fully connected layer to obtain the target probability Pt1. Then, the classification feature vector H_cls1 is input into BiLSTM to obtain the classification feature vector H_cls2 output by BiLSTM, and the BiLSTM is updated to obtain the network feature vector Ht2 corresponding to the second round of BiLSTM. Then, in the second round of decoding, the decoding sub-model concatenates the classification feature vector H_cls2 with the network feature vector Ht2 corresponding to the second round of BiLSTM to obtain the concatenated vector H2', inputs the concatenated vector H2' into the first fully connected layer, and uses the first preset activation function to activate the second round output result of the first fully connected layer to obtain the target probability Pt2. Similarly, when the maximum decoding rounds set by the user are reached or a terminator is generated, the decoding process ends.

[0113] The above-mentioned first preset activation function can be a sigmoid function or other functions, which is not limited in the embodiment of the present disclosure.

[0114] Step 2022: Input the classification feature vector into the second fully connected layer to obtain a first probability distribution.

[0115] In each round of decoding, the classification feature vector is input into the second fully connected layer, and the output result of the second fully connected layer is activated using a second preset activation function to obtain a first probability distribution corresponding to the preset vocabulary.

[0116] For example, in the first round of decoding, the classification feature vector H_cls1 is input into the second fully connected layer, and the first round output result of the second fully connected layer is activated by the second preset activation function to obtain the first probability distribution Pwt1 corresponding to the preset vocabulary. Then, the classification feature vector H_cls2 is input into the second fully connected layer, and the second round output result of the second fully connected layer is activated by the second preset activation function to obtain the first probability distribution Pwt2 corresponding to the preset vocabulary. And so on, the first probability distribution of each round is obtained.

[0117] The second preset activation function may be a softMax function or other functions, which is not limited in the embodiment of the present disclosure.

[0118] It can be understood that the first preset activation function may be the same as or different from the second preset activation function.

[0119] Step 2023: Regularize the question feature vector and the content feature vector, and obtain a second probability distribution based on the regularization result.

[0120] The decoding sub-model performs regularization processing on the question feature vector and the content feature vector to obtain a regularization result, and then calculates a second probability distribution based on the regularization result.

[0121] For example, the problem feature vector H_q and the content feature vector H_p are distributed by L2 regularization to obtain the regularized results Lq and Lp, and then the second probability distribution Put is calculated according to the formula Put = softMax (Lp^T*Lq). Wherein, Lp^T is the transposed matrix of Lp. In practical applications, other regularization processing methods can also be used, and the embodiments of the present disclosure do not limit this.

[0122] It can be understood that when the texts of the target question and the target content are long, the question feature vector can be a vector matrix composed of feature vectors corresponding to multiple question information fragments, and the content feature vector can be a vector matrix composed of feature vectors corresponding to multiple content information fragments.

[0123] In the above embodiment, for each round of decoding, the classification feature vector is concatenated with the network feature vector corresponding to the current round of the bidirectional long short-term memory network to obtain a concatenated vector, and the concatenated vector is input into the first fully connected layer to obtain the target probability; the classification feature vector is input into the second fully connected layer to obtain the first probability distribution; the question feature vector and the content feature vector are regularized, and the second probability distribution is obtained according to the regularization result. Through the embodiment of the present disclosure, the target probability, the first probability distribution and the second probability distribution of each round of decoding can be determined, so that the candidate information can be screened out according to the target probability, the first probability distribution and the second probability distribution, and then the target information can be determined.

[0124] In one embodiment, Figure 5 As shown, the step of determining the third probability distribution corresponding to each round of decoding according to the target probability corresponding to each round of decoding, the first probability distribution, and the second probability distribution may include:

[0125] Step 2031: For each round of decoding, determine a first weight corresponding to the first probability distribution and a second weight corresponding to the second probability distribution according to the target probability.

[0126] In each round of decoding, the target probability is determined as the first weight corresponding to the first probability distribution, and the difference between 1 and the target probability is determined as the second weight corresponding to the second probability.

[0127] For example, Pt is determined as the first weight corresponding to the first probability distribution Pwt, and 1-Pt is determined as the second weight corresponding to the second probability distribution Put.

[0128] Step 2032: Perform weighted sum processing on the first probability distribution and the second probability distribution according to the first weight and the second weight to obtain a third probability distribution.

[0129] The third probability distribution P is calculated according to the formula P=Pt*Pwt+(1-Pt)*Put.

[0130] In the above embodiment, for each round of decoding, the first weight corresponding to the first probability distribution and the second weight corresponding to the second probability distribution are determined according to the target probability; the first probability distribution and the second probability distribution are weighted and summed according to the first weight and the second weight to obtain a third probability distribution. Through the embodiment of the present disclosure, the first probability distribution and the second probability distribution can be weighted according to the target probability, and then the extraction probability of each word in the predicted word list and the target content can be obtained, so that information can be extracted according to the extraction probability of each word in the future.

[0131] In one embodiment, Figure 6 As shown, based on the above embodiment, a training process of the information extraction model may also be included, such as the following steps:

[0132] Step 301: Use a crawler tool to obtain general corpus from the Internet.

[0133] Among them, the general corpus consists of questions, content and answers.

[0134] There is a large amount of question and answer data on the Internet. By using crawler tools to crawl this data from the Internet, we can obtain general corpus.

[0135] For example, we crawled data from an encyclopedia website on the Internet, where the question “What color clothes make you look whiter” and the answer was “Black is a versatile color… bright red can make people look fair…”, and obtained general corpus with the question “What color clothes make you look whiter”, the content was “Black is a versatile color… bright red can make people look fair…”, and the answer was “black, bright red”.

[0136] Step 302, perform data processing on the general corpus to obtain training questions, training content and training labels.

[0137] After acquiring the general corpus, the terminal performs data processing on the general corpus to obtain training questions, training content and training labels for the training information extraction model.

[0138] For example, the general corpus includes: the question "What color clothes can make you look fairer", the content "Black is a versatile color... Bright red can make people look fairer...", and the answer "black, bright red". Based on the general corpus, the training question "What color clothes can make you look fairer", the training content "Black is a versatile color... Bright red can make people look fairer...", and the training label "black, bright red" can be obtained. By analogy, a large number of training questions, training content, and training labels can be obtained to form a training set.

[0139] Step 303: Perform model training according to the training questions, training content and training labels to obtain an information extraction model.

[0140] Input the training problem and training content into the initial model, obtain the training results output by the initial model, and judge whether the model meets the preset convergence conditions based on the training labels and training results. If it does not meet the preset convergence conditions, adjust the adjustable parameters in the model to continue training; if it meets the preset convergence conditions, end the training and determine the model at the end of the training as the information extraction model.

[0141] In the above embodiment, a crawler tool is used to obtain general corpus from the Internet; data processing is performed on the general corpus to obtain training questions, training content and training labels; model training is performed based on the training questions, training content and training labels to obtain an information extraction model. Through the disclosed embodiment, the terminal obtains the training questions, training content and training labels used for model training based on the general corpus. Since the general corpus is not limited to a certain field, the information extraction model trained based on the training questions, training content and training labels has strong versatility and transferability. Furthermore, since training labels can be obtained without manual annotation, it can not only reduce labor costs, save label acquisition time, improve label acquisition efficiency, and improve model training efficiency, but also avoid the risk of user information leakage and improve information security.

[0142] It should be understood that although Figures 1 to 6 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, Figures 1 to 6 At least part of the steps may include multiple steps or multiple stages. These steps or stages are not necessarily performed at the same time, but can be performed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed in turn or alternately with other steps or at least part of the steps or stages in other steps.

[0143] In one embodiment, Figure 7 As shown, an information extraction device is provided, comprising:

[0144] The question content acquisition module 401 is used to acquire the target question and target content;

[0145] The information extraction module 402 is used to input the target question and target content into a pre-trained information extraction model to extract information, and obtain the target information that answers the target question output by the information extraction model;

[0146] Among them, the information extraction model is used to determine the target probability of extracting information from a preset vocabulary according to the target question and the target content, and extract the target information from the preset vocabulary or the target content according to the target probability.

[0147] In one embodiment, the information extraction model includes an encoding sub-model and a decoding sub-model, and the information extraction module 402 includes:

[0148] The encoding submodule is used to encode the target question and target content using the encoding submodel to obtain a feature vector;

[0149] A decoding submodule, used to perform multiple rounds of decoding processing on the feature vector using the decoding submodel, to obtain the target probability corresponding to each round of decoding, the first probability distribution corresponding to the preset vocabulary, and the second probability distribution corresponding to the feature vector; wherein the first probability distribution is used to represent the first extraction probability corresponding to each word in the preset vocabulary, and the second probability distribution is used to represent the second extraction probability corresponding to each word in the target content;

[0150] A probability determination submodule, which determines a third probability distribution corresponding to each round of decoding according to the target probability corresponding to each round of decoding, the first probability distribution, and the second probability distribution; the third probability distribution is used to characterize the third extraction probability corresponding to each word in the preset word list and the target content respectively;

[0151] The information extraction submodule is used to extract information according to the third probability distribution corresponding to each round of decoding to obtain multiple candidate information, and filter out target information from the candidate information.

[0152] In one embodiment, the feature vector includes a classification feature vector, a question feature vector and a content feature vector, and the decoding sub-model includes a bidirectional long short-term memory network, a first fully connected layer and a second fully connected layer. The above-mentioned decoding sub-module is specifically used to, for each round of decoding, concatenate the classification feature vector with the network feature vector corresponding to the current round of the bidirectional long short-term memory network to obtain a concatenated vector, and input the concatenated vector into the first fully connected layer to obtain a target probability; input the classification feature vector into the second fully connected layer to obtain a first probability distribution; regularize the question feature vector and the content feature vector, and obtain a second probability distribution based on the regularization result.

[0153] In one of the embodiments, the above-mentioned probability determination submodule is specifically used to determine, for each round of decoding, a first weight corresponding to the first probability distribution and a second weight corresponding to the second probability distribution according to the target probability; and perform weighted summation processing on the first probability distribution and the second probability distribution according to the first weight and the second weight to obtain a third probability distribution.

[0154] In one embodiment, the information extraction submodule is specifically used to determine the words with the third highest probability of being extracted in the preset word list and the target content according to the third probability distribution for each round of decoding, and determine the words with the third highest probability of being extracted as candidate information.

[0155] In one of the embodiments, the above-mentioned encoding submodule is specifically used to splice the target question and the target content to obtain splicing information; perform word segmentation on the splicing information to obtain multiple information segments; wherein the information segments include classification identification segments; input the multiple information segments into the encoding submodel for encoding processing to obtain feature vectors corresponding to each information segment output by the encoding submodel; wherein the classification identification segment corresponds to the classification feature vector.

[0156] In one embodiment, the device further comprises:

[0157] The corpus acquisition module is used to acquire general corpus from the Internet using crawler tools; general corpus consists of questions, content and answers;

[0158] Corpus processing module, used to process the general corpus to obtain training questions, training content and training labels;

[0159] The training module is used to train the model according to the training questions, training content and training labels to obtain the information extraction model.

[0160] For the specific definition of the information extraction device, please refer to the definition of the information extraction method above, which will not be repeated here. Each module in the above-mentioned information extraction device 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.

[0161] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Figure 8 As shown. The computer device includes a processor, a memory, a communication interface, a display screen and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be achieved through WIFI, an operator network, NFC (near field communication) or other technologies. When the computer program is executed by the processor, an information extraction method is implemented. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch sub-model covered on the display screen, or a button, trackball or touchpad set on the computer device housing, or an external keyboard, touchpad or mouse, etc.

[0162] Those skilled in the art will understand that Figure 8 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0163] In one embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the following steps are implemented:

[0164] Get the target questions and target content;

[0165] Input the target question and target content into the pre-trained information extraction model to extract information, and obtain the target information that answers the target question output by the information extraction model;

[0166] Among them, the information extraction model is used to determine the target probability of extracting information from a preset vocabulary according to the target question and the target content, and extract the target information from the preset vocabulary or the target content according to the target probability.

[0167] In one embodiment, the information extraction model includes an encoding sub-model and a decoding sub-model, and the processor further implements the following steps when executing the computer program:

[0168] The target question and target content are encoded using the encoding sub-model to obtain a feature vector;

[0169] The feature vector is subjected to multiple rounds of decoding processing using the decoding submodel to obtain the target probability corresponding to each round of decoding, the first probability distribution corresponding to the preset vocabulary, and the second probability distribution corresponding to the feature vector; wherein the first probability distribution is used to represent the first extraction probability corresponding to each word in the preset vocabulary, and the second probability distribution is used to represent the second extraction probability corresponding to each word in the target content;

[0170] Determine a third probability distribution corresponding to each round of decoding according to the target probability corresponding to each round of decoding, the first probability distribution, and the second probability distribution; the third probability distribution is used to characterize the third extraction probability corresponding to each word in the preset word list and the target content respectively;

[0171] Information extraction is performed according to the third probability distribution corresponding to each round of decoding to obtain multiple candidate information, and target information is screened out from the candidate information.

[0172] In one embodiment, the feature vector includes a classification feature vector, a question feature vector, and a content feature vector, the decoding sub-model includes a bidirectional long short-term memory network, a first fully connected layer, and a second fully connected layer, and the processor further implements the following steps when executing the computer program:

[0173] For each round of decoding, the classification feature vector is concatenated with the network feature vector corresponding to the current round of the bidirectional long short-term memory network to obtain a concatenated vector, and the concatenated vector is input into the first fully connected layer to obtain the target probability;

[0174] Inputting the classification feature vector into the second fully connected layer to obtain a first probability distribution;

[0175] Regularization is performed on the question feature vector and the content feature vector, and a second probability distribution is obtained according to the regularization result.

[0176] In one embodiment, when the processor executes the computer program, the processor further implements the following steps:

[0177] For each round of decoding, determining a first weight corresponding to the first probability distribution and a second weight corresponding to the second probability distribution according to the target probability;

[0178] The first probability distribution and the second probability distribution are weightedly summed according to the first weight and the second weight to obtain a third probability distribution.

[0179] In one embodiment, when the processor executes the computer program, the processor further implements the following steps:

[0180] For each round of decoding, the words with the third highest probability of being extracted in the preset word list and the target content are determined according to the third probability distribution, and the words with the third highest probability of being extracted are determined as candidate information.

[0181] In one embodiment, when the processor executes the computer program, the processor further implements the following steps:

[0182] Perform splicing processing on the target problem and the target content to obtain splicing information;

[0183] Performing word segmentation processing on the spliced ​​information to obtain multiple information segments; wherein the information segments include classification identification segments;

[0184] Multiple information segments are input into the encoding sub-model for encoding processing to obtain feature vectors corresponding to each information segment output by the encoding sub-model; wherein the classification identification segment corresponds to the classification feature vector.

[0185] In one embodiment, when the processor executes the computer program, the processor further implements the following steps:

[0186] Use crawler tools to obtain general corpus from the Internet; general corpus consists of questions, content and answers;

[0187] Perform data processing on general corpus to obtain training questions, training content and training labels;

[0188] The model is trained according to the training questions, training content and training labels to obtain the information extraction model.

[0189] In one embodiment, a computer readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the following steps are implemented:

[0190] Get the target questions and target content;

[0191] Input the target question and target content into the pre-trained information extraction model to extract information, and obtain the target information that answers the target question output by the information extraction model;

[0192] Among them, the information extraction model is used to determine the target probability of extracting information from a preset vocabulary according to the target question and the target content, and extract the target information from the preset vocabulary or the target content according to the target probability.

[0193] In one embodiment, the information extraction model includes an encoding sub-model and a decoding sub-model, and the computer program, when executed by a processor, further implements the following steps:

[0194] The target question and target content are encoded using the encoding sub-model to obtain a feature vector;

[0195] The feature vector is subjected to multiple rounds of decoding processing using the decoding submodel to obtain the target probability corresponding to each round of decoding, the first probability distribution corresponding to the preset vocabulary, and the second probability distribution corresponding to the feature vector; wherein the first probability distribution is used to represent the first extraction probability corresponding to each word in the preset vocabulary, and the second probability distribution is used to represent the second extraction probability corresponding to each word in the target content;

[0196] Determine a third probability distribution corresponding to each round of decoding according to the target probability corresponding to each round of decoding, the first probability distribution, and the second probability distribution; the third probability distribution is used to characterize the third extraction probability corresponding to each word in the preset word list and the target content respectively;

[0197] Information extraction is performed according to the third probability distribution corresponding to each round of decoding to obtain multiple candidate information, and target information is screened out from the candidate information.

[0198] In one embodiment, the feature vector includes a classification feature vector, a question feature vector, and a content feature vector, the decoding sub-model includes a bidirectional long short-term memory network, a first fully connected layer, and a second fully connected layer, and the computer program, when executed by the processor, further implements the following steps:

[0199] For each round of decoding, the classification feature vector is concatenated with the network feature vector corresponding to the current round of the bidirectional long short-term memory network to obtain a concatenated vector, and the concatenated vector is input into the first fully connected layer to obtain the target probability;

[0200] Inputting the classification feature vector into the second fully connected layer to obtain a first probability distribution;

[0201] Regularization is performed on the question feature vector and the content feature vector, and a second probability distribution is obtained according to the regularization result.

[0202] In one embodiment, when the computer program is executed by a processor, the following steps are also implemented:

[0203] For each round of decoding, determining a first weight corresponding to the first probability distribution and a second weight corresponding to the second probability distribution according to the target probability;

[0204] The first probability distribution and the second probability distribution are weightedly summed according to the first weight and the second weight to obtain a third probability distribution.

[0205] In one embodiment, when the computer program is executed by a processor, the following steps are also implemented:

[0206] For each round of decoding, the words with the third highest probability of being extracted in the preset word list and the target content are determined according to the third probability distribution, and the words with the third highest probability of being extracted are determined as candidate information.

[0207] In one embodiment, when the computer program is executed by a processor, the following steps are also implemented:

[0208] Perform splicing processing on the target problem and the target content to obtain splicing information;

[0209] Performing word segmentation processing on the spliced ​​information to obtain multiple information segments; wherein the information segments include classification identification segments;

[0210] Multiple information segments are input into the encoding sub-model for encoding processing to obtain feature vectors corresponding to each information segment output by the encoding sub-model; wherein the classification identification segment corresponds to the classification feature vector.

[0211] In one embodiment, when the computer program is executed by a processor, the following steps are also implemented:

[0212] Use crawler tools to obtain general corpus from the Internet; general corpus consists of questions, content and answers;

[0213] Perform data processing on general corpus to obtain training questions, training content and training labels;

[0214] The model is trained according to the training questions, training content and training labels to obtain the information extraction model.

[0215] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory or optical memory, etc. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0216] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0217] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.

Claims

1. An information extraction method, characterized in that: The method comprises: Get the target questions and target content; Inputting the target question and the target content into a pre-trained information extraction model to extract information, and obtaining target information that answers the target question output by the information extraction model; The information extraction model is used to determine a target probability of extracting information from a preset vocabulary according to the target question and the target content, and extract the target information from the preset vocabulary or the target content according to the target probability; The information extraction model includes an encoding sub-model and a decoding sub-model, the decoding sub-model includes a bidirectional long short-term memory network, a first fully connected layer and a second fully connected layer; the target probability of extracting information from a preset vocabulary according to the target problem and the target content, and extracting the target information from the preset vocabulary or the target content according to the target probability, includes: Using the encoding sub-model to encode the target question and the target content to obtain a feature vector; the feature vector includes a classification feature vector, a question feature vector and a content feature vector; Using the decoding sub-model to perform multiple rounds of decoding processing on the feature vector, for each round of decoding, concatenating the classification feature vector with the network feature vector corresponding to the current round of the bidirectional long short-term memory network to obtain a concatenated vector, and inputting the concatenated vector into the first fully connected layer to obtain the target probability; Inputting the classification feature vector into the second fully connected layer to obtain a first probability distribution; Regularization is performed on the question feature vector and the content feature vector, and a second probability distribution is obtained according to the regularization result; wherein the first probability distribution is used to characterize the first extraction probability corresponding to each word in the preset vocabulary, and the second probability distribution is used to characterize the second extraction probability corresponding to each word in the target content; Determine a third probability distribution corresponding to each round of decoding according to the target probability corresponding to each round of decoding, the first probability distribution and the second probability distribution; the third probability distribution is used to characterize the third extraction probability corresponding to each word in the preset word list and the target content respectively; Information extraction is performed according to the third probability distribution corresponding to each round of decoding to obtain a plurality of candidate information, and the target information is screened out from the candidate information.

2. The method according to claim 1, characterized in that The determining, according to the target probability corresponding to each round of decoding, the first probability distribution, and the second probability distribution, a third probability distribution corresponding to each round of decoding includes: For each round of decoding, determining a first weight corresponding to the first probability distribution and a second weight corresponding to the second probability distribution according to the target probability; The first probability distribution and the second probability distribution are weightedly summed according to the first weight and the second weight to obtain the third probability distribution.

3. The method according to claim 1, characterized in that The extracting information according to the third probability distribution corresponding to each round of decoding to obtain a plurality of candidate information includes: For each round of decoding, the words with the third highest probability of being extracted in the preset word list and the target content are determined according to the third probability distribution, and the words with the third highest probability of being extracted are determined as the candidate information.

4. The method according to claim 1, characterized in that The using the encoding sub-model to encode the target question and the target content to obtain a feature vector includes: Performing splicing processing on the target question and the target content to obtain splicing information; Performing word segmentation processing on the spliced ​​information to obtain multiple information segments; wherein the information segments include classification identification segments; The multiple information segments are input into the encoding sub-model for encoding processing to obtain feature vectors corresponding to the information segments output by the encoding sub-model; wherein the classification identification segment corresponds to the classification feature vector.

5. The method according to claim 1, characterized in that Before inputting the target question and the target content into a pre-trained information extraction model for information extraction, the method further includes: Using crawler tools to obtain general corpus from the Internet; the general corpus consists of questions, content and answers; Performing data processing on the general corpus to obtain training questions, training content, and training labels; Model training is performed according to the training problem, the training content and the training label to obtain the information extraction model.

6. An information extraction device, characterized in that: The device comprises: A question content acquisition module is used to acquire target questions and target content; An information extraction module, used for inputting the target question and the target content into a pre-trained information extraction model to extract information, and obtaining target information that answers the target question output by the information extraction model; The information extraction model is used to determine a target probability of extracting information from a preset vocabulary according to the target question and the target content, and extract the target information from the preset vocabulary or the target content according to the target probability; The information extraction model includes an encoding sub-model and a decoding sub-model, and the decoding sub-model includes a bidirectional long short-term memory network, a first fully connected layer and a second fully connected layer; the information extraction model is specifically used for: Using the encoding sub-model to encode the target question and the target content to obtain a feature vector; the feature vector includes a classification feature vector, a question feature vector and a content feature vector; Using the decoding sub-model to perform multiple rounds of decoding processing on the feature vector, for each round of decoding, concatenating the classification feature vector with the network feature vector corresponding to the current round of the bidirectional long short-term memory network to obtain a concatenated vector, and inputting the concatenated vector into the first fully connected layer to obtain the target probability; Inputting the classification feature vector into the second fully connected layer to obtain a first probability distribution; Regularization is performed on the question feature vector and the content feature vector, and a second probability distribution is obtained according to the regularization result; wherein the first probability distribution is used to characterize the first extraction probability corresponding to each word in the preset vocabulary, and the second probability distribution is used to characterize the second extraction probability corresponding to each word in the target content; Determine a third probability distribution corresponding to each round of decoding according to the target probability corresponding to each round of decoding, the first probability distribution and the second probability distribution; the third probability distribution is used to characterize the third extraction probability corresponding to each word in the preset word list and the target content respectively; Information extraction is performed according to the third probability distribution corresponding to each round of decoding to obtain a plurality of candidate information, and the target information is screened out from the candidate information.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 5 are implemented.

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 steps of the method according to any one of claims 1 to 5 are implemented.

Images