Neural network information extraction method and system fused with external dictionary
By combining external entity dictionary with word participle dictionary, and using named entity annotation and feature vector matrix to connect input text with external knowledge features, the problem of difficulty in combining external knowledge in neural networks is solved, and the accuracy and recall rate of information extraction are improved.
Patent Information
- Application Number
- CN202411342506.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-25
- Publication Date
- 2024-12-13
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Existing neural networks are difficult to effectively combine external knowledge in information extraction tasks, require massive training data and limited learning effects, which limits their performance in information extraction.
By combining external entity dictionary with word participle dictionary, the entity information is marked using the named entity annotation method, and the external feature vector matrix is initialized, the input text is connected with the external knowledge feature vector, and the input is trained and sequence annotated as input to the neural network, using CRF and BiLSTM model for processing.
Without increasing the amount of training data, the accuracy and recall of the model are improved, the existing external dictionary is effectively utilized, and the information extraction effect is improved.
Smart Images
Figure CN119129579A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to information extraction technology, and in particular to a neural network information extraction method and system integrating an external dictionary. Background Art
[0002] With the development of electronic informatization, massive amounts of text data have been generated in business scenarios in multiple fields such as finance, insurance, and banking. These text data contain valuable business information, and it has become an urgent need to improve the effectiveness of existing information extraction methods.
[0003] In recent years, deep neural networks and related technologies have developed rapidly in the field of natural language processing. Compared with traditional artificial rules and machine learning algorithms, they have significant advantages such as high accuracy, strong generalization, and independence from feature engineering. They have been widely used in information extraction related tasks such as named entity recognition and entity relationship extraction, and have achieved remarkable results. They have gradually become the mainstream algorithm for current information extraction.
[0004] Studies have shown that if some external knowledge is combined during information extraction, such as combining existing map POI information during address extraction, or combining existing singer song names, address books, and other information during vehicle computer command recognition, the effect of information extraction algorithms can be effectively improved. However, if neural networks are required to learn external knowledge from data on their own, massive amounts of training data are required, and the learning effect is limited, which limits the performance of neural network models in information extraction tasks. Summary of the invention
[0005] The embodiments of the present invention provide a neural network information extraction method and system integrating an external dictionary, which can solve the problems in the prior art.
[0006] According to a first aspect of the embodiments of the present invention,
[0007] Provides a neural network information extraction method that integrates external dictionaries, including:
[0008] The external entity dictionary is added to the word segmentation dictionary, and the original input text is segmented into Chinese words to obtain an input text sequence, or the original input text is directly used as the input text sequence; according to the external entity dictionary, the words existing in the input text sequence are searched from the external entity dictionary, and the entity information in the external entity dictionary is marked in the input text sequence using a named entity tagging method as an external knowledge feature;
[0009] Initialize an external feature vector matrix, wherein the number of rows of the external feature vector matrix is the number of external knowledge labels, and the number of columns is the external knowledge feature dimension; vectorize the input text sequence to obtain an input text vector; use the external feature vector matrix to vectorize external knowledge features to obtain an external knowledge feature vector;
[0010] The input text vector is connected with the external knowledge feature vector to obtain a final input vector; the final input vector is used as the input of the neural network and input into the neural network sequence labeling model for training and sequence labeling to obtain a sequence labeling result; and the sequence labeling result is converted into a semantic understanding result.
[0011] In an optional embodiment,
[0012] The external entity dictionary is added to the word segmentation dictionary, and the original input text is segmented into Chinese words to obtain an input text sequence, or the original input text is directly used as the input text sequence; according to the external entity dictionary, the words existing in the input text sequence are searched from the external entity dictionary, and the entity information in the external entity dictionary is marked in the input text sequence using a named entity tagging method, and the external knowledge features include:
[0013] Searching the external entity dictionary for words present in the input text sequence according to the external entity dictionary in combination with a dual array tree data structure;
[0014] The named entity tagging method using the BIO tagging method marks the entity information in the external entity dictionary of the characters / words in the input text sequence as external knowledge features.
[0015] In an optional embodiment,
[0016] The method further comprises:
[0017] Based on the pre-collected external entity information, collect texts containing vocabulary content and general texts to train domain word vectors;
[0018] Based on the domain word vector, candidate vectors are obtained from the user input text, and the candidate vectors are merged to build a vector representation list. The target candidate vector is obtained through a layer of linear neural network encoding.
[0019] In an optional embodiment,
[0020] Initialize an external feature vector matrix, wherein the number of rows of the external feature vector matrix is the number of external knowledge labels, and the number of columns is the external knowledge feature dimension; vectorize the input text sequence to obtain an input text vector; use the external feature vector matrix to vectorize the external knowledge features, and obtain an external knowledge feature vector including:
[0021] Initialize an external feature vector matrix, where the number of rows of the external feature vector matrix is the number of external knowledge labels, and the number of columns is the external knowledge feature dimension;
[0022] One-hot encoding and vectorization are performed on the input text sequence to obtain an input text vector;
[0023] According to the external entity information marked in the input text sequence, the corresponding external knowledge label is searched in the external feature vector matrix, and the vector of the corresponding row is extracted as the external knowledge feature vector of the external entity;
[0024] The external knowledge feature vectors of all external entities in the input text sequence are concatenated to obtain a complete external knowledge feature vector.
[0025] In an optional embodiment,
[0026] The input text vector is connected with the external knowledge feature vector to obtain a final input vector; the final input vector is used as the input of the neural network and input into the neural network sequence labeling model for training and sequence labeling, and the sequence labeling results obtained include:
[0027] The sequence annotation results are determined according to the following formula:
[0028] prediction=CRF(BiLSTM(Final_input_embedding));
[0029] Among them, prediction represents the sequence labeling result, CRF() represents the processing result of conditional random field, BiLSTM() represents the processing result of bidirectional long short-term memory network, and Final_input_embedding represents the final input vector.
[0030] In an optional embodiment,
[0031] The method further includes training the neural network sequence labeling model, and the training method is as follows:
[0032] The CRF loss function is used as the training objective, and the training objective is to maximize the proportion of the true path score in all paths:
[0033]
[0034] Among them, Loss represents the loss value of the CRF loss function, p real_path represents the fraction of the true path of the sample, P i represents the score of the i-th path, and n represents the number of all paths in the sample.
[0035] In an optional embodiment,
[0036] The final input vector is used as the input of the neural network and input into the neural network sequence labeling model for training and sequence labeling, and the sequence labeling results obtained include:
[0037] The Viterbi decoding algorithm is used to decode the optimal sequence as the sequence labeling result.
[0038] In the second aspect of the embodiments of the present application,
[0039] Provides a neural network information extraction system that integrates external dictionaries, including:
[0040] The first unit is used to add an external entity dictionary to a word segmentation dictionary, perform Chinese word segmentation on an original input text to obtain an input text sequence, or directly use the original input text as an input text sequence; according to the external entity dictionary, search for words existing in the input text sequence from the external entity dictionary, and use a named entity tagging method to mark entity information in the external entity dictionary in the input text sequence as an external knowledge feature;
[0041] The second unit is used to initialize an external feature vector matrix, wherein the number of rows of the external feature vector matrix is the number of external knowledge labels, and the number of columns is the external knowledge feature dimension; vectorize the input text sequence to obtain an input text vector; and use the external feature vector matrix to vectorize external knowledge features to obtain an external knowledge feature vector;
[0042] The third unit is used to connect the input text vector with the external knowledge feature vector to obtain a final input vector; use the final input vector as the input of the neural network, input it into the neural network sequence labeling model for training and sequence labeling, and obtain a sequence labeling result; and convert the sequence labeling result into a semantic understanding result.
[0043] According to a third aspect of the embodiments of the present invention,
[0044] An electronic device is provided, comprising:
[0045] processor;
[0046] a memory for storing processor-executable instructions;
[0047] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.
[0048] A fourth aspect of the embodiments of the present invention is:
[0049] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.
[0050] The present invention mainly aims at the problem that the existing neural network is difficult to combine external knowledge, and proposes a neural network information extraction method integrating external dictionaries. Without increasing the amount of data required for training, the existing external dictionaries are effectively utilized, thereby improving the accuracy and recall rate of the model. BRIEF DESCRIPTION OF THE DRAWINGS
[0051] Figure 1 A schematic diagram of a flow chart of a neural network information extraction method integrating an external dictionary according to an embodiment of the present invention;
[0052] Figure 2 Schematic diagram of the structure of a neural network information extraction system integrating an external dictionary according to an embodiment of the present invention. DETAILED DESCRIPTION
[0053] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0054] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.
[0055] Figure 1 FIG. 1 is a flow chart of a neural network information extraction method integrating an external dictionary according to an embodiment of the present invention. Figure 1 As shown, the method includes:
[0056] S101. Add an external entity dictionary to the word segmentation dictionary, perform Chinese word segmentation on the original input text to obtain an input text sequence, or directly use the original input text as the input text sequence; according to the external entity dictionary, search for words existing in the input text sequence from the external entity dictionary, and use a named entity tagging method to mark entity information in the external entity dictionary in the input text sequence as an external knowledge feature;
[0057] S102. Initialize an external feature vector matrix, wherein the number of rows of the external feature vector matrix is the number of external knowledge labels, and the number of columns is the external knowledge feature dimension; vectorize the input text sequence to obtain an input text vector; use the external feature vector matrix to vectorize external knowledge features to obtain an external knowledge feature vector;
[0058] S103. Connect the input text vector with the external knowledge feature vector to obtain a final input vector; use the final input vector as the input of the neural network, input it into the neural network sequence labeling model for training and sequence labeling, and obtain a sequence labeling result; convert the sequence labeling result into a semantic understanding result.
[0059] In an optional embodiment,
[0060] The external entity dictionary is added to the word segmentation dictionary, and the original input text is segmented into Chinese words to obtain an input text sequence, or the original input text is directly used as the input text sequence; according to the external entity dictionary, the words existing in the input text sequence are searched from the external entity dictionary, and the entity information in the external entity dictionary is marked in the input text sequence using a named entity tagging method, and the external knowledge features include:
[0061] Searching the external entity dictionary for words present in the input text sequence according to the external entity dictionary in combination with a dual array tree data structure;
[0062] Using a named entity tagging method of the BIO tagging method to tag the characters / words in the input text sequence with entity information in the external entity dictionary as external knowledge features;
[0063] The method further comprises:
[0064] Based on the pre-collected external entity information, collect texts containing vocabulary content and general texts to train domain word vectors;
[0065] Based on the domain word vector, candidate vectors are obtained from the user input text, and the candidate vectors are merged to build a vector representation list. The target candidate vector is obtained through a layer of linear neural network encoding.
[0066] For example,
[0067] Take named entity recognition in the music field as an example: "Listen to Jay Chou's "The End of War" from "Seven Miles of Fragrance", we need to extract the singer: Jay Chou, the album: Seven Miles of Fragrance, and the song: "The End of War".
[0068] 1. Collect external entity information in advance, including singer table (Jay Chou, Zhou Jie, JJ Lin, JJ...), song table (The End of War, A Thousand Years Later, Qilixiang, No. 89757, Daoxiang...), album table (Qilixiang, No. 89757...), and collect texts containing the above vocabulary and general texts to train them into domain word vectors.
[0069] 2. Match the singer list, song list, and album list from the input text "Listen to Jay Chou's "The End of War" in "Seven Miles of Fragrance" to get the matching results: possible singers [Jay Chou, Jay Chou], possible songs [Seven Miles of Fragrance, The End of War], possible albums [Seven Miles of Fragrance].
[0070] 3. Get the vector representation of possible singers from the domain word vector [V 周杰伦 , V 周杰 ], the vector representation of the song [V 七里香 , V 止战之殇 ], the vector representation of the album [V 七里香 ]. And merge them respectively. The merging methods include but are not limited to summing, averaging, using neural network encoding, etc. Taking the singer as an example:
[0071] V song =Relu(W(concat(Sum(V 周杰伦 , V 周杰 ); Avg(V 周杰伦 , V 周杰 )))+b);
[0072] Among them, Relu represents the activation function, W represents the weight value, concat() represents the concatenation function, Sum() represents the summation function, Avg() represents the averaging function, and b represents the bias.
[0073] The above formula means that the vectors in the vector representation list are summed and averaged respectively, and then connected. Then, the candidate song vectors are obtained through a layer of linear neural network encoding.
[0074] In an optional embodiment,
[0075] Initialize an external feature vector matrix, wherein the number of rows of the external feature vector matrix is the number of external knowledge labels, and the number of columns is the external knowledge feature dimension; vectorize the input text sequence to obtain an input text vector; use the external feature vector matrix to vectorize the external knowledge features, and obtain an external knowledge feature vector including:
[0076] Initialize an external feature vector matrix, where the number of rows of the external feature vector matrix is the number of external knowledge labels, and the number of columns is the external knowledge feature dimension;
[0077] One-hot encoding and vectorization are performed on the input text sequence to obtain an input text vector;
[0078] According to the external entity information marked in the input text sequence, the corresponding external knowledge label is searched in the external feature vector matrix, and the vector of the corresponding row is extracted as the external knowledge feature vector of the external entity;
[0079] The external knowledge feature vectors of all external entities in the input text sequence are concatenated to obtain a complete external knowledge feature vector.
[0080] In an optional embodiment,
[0081] The input text vector is connected with the external knowledge feature vector to obtain a final input vector; the final input vector is used as the input of the neural network and input into the neural network sequence labeling model for training and sequence labeling, and the sequence labeling results obtained include:
[0082] The sequence annotation results are determined according to the following formula:
[0083] prediction=CRF(BiLSTM(Final_input_embedding));
[0084] Among them, prediction represents the sequence labeling result, CRF() represents the processing result of conditional random field, BiLSTM() represents the processing result of bidirectional long short-term memory network, and Final_input_embedding represents the final input vector.
[0085] In an optional embodiment,
[0086] The method further includes training the neural network sequence labeling model, and the training method is as follows:
[0087] The CRF loss function is used as the training objective, and the training objective is to maximize the proportion of the true path score in all paths:
[0088]
[0089] Among them, Loss represents the loss value of the CRF loss function, p real_path represents the fraction of the true path of the sample, P i represents the score of the i-th path, and n represents the number of all paths in the sample.
[0090] In an optional embodiment,
[0091] The final input vector is used as the input of the neural network and input into the neural network sequence labeling model for training and sequence labeling, and the sequence labeling results obtained include:
[0092] The Viterbi decoding algorithm is used to decode the optimal sequence as the sequence labeling result.
[0093] Another specific embodiment of the present application is as follows:
[0094] Neural network: Artificial neural network (ANN), also known as neural network (NN), is a mathematical model or computational model that imitates the structure and function of biological neural networks. A neural network is a computational model that consists of a large number of nodes (or "neurons") and the connections between them. Each node represents a specific output function, called an activation function. The connection between each two nodes represents a weighted value for the signal passing through the connection, called a weight, which is equivalent to the memory of the artificial neural network. The output of the network varies depending on the network connection method, weight value and activation function.
[0095] Sequence Tagging: Sequence Tagging is a relatively simple NLP task, but it can also be called the most basic task. Its input is a text sequence, and its output is the labeling of each token (character / word) in the text. Sequence tagging covers a wide range of problems and can be used to solve a series of character classification problems, such as word segmentation, part-of-speech tagging, named entity recognition, relationship extraction, etc.
[0096] Information Extraction: The goal of information extraction (IE) is to convert text information into structured information. It was originally used to locate specific information in natural language documents and is a subfield of natural language processing.
[0097] The text fed into a neural network is usually input in the form of words or phrases;
[0098] If the input is by word, the external entity dictionary needs to be added to the word segmentation dictionary, and then the original input text is segmented into Chinese words to obtain the input text sequence. For example, if the original input text is: "I work at a certain bank's credit card branch center", and the word "a certain bank's credit card branch center" is added to the external dictionary, it will be segmented into "I work at a certain bank's credit card branch center". In contrast, if the word does not exist in the external dictionary, the segmentation result is "I work at a certain bank's credit card branch center".
[0099] If the input is by word, the original input text is directly used as the input text sequence.
[0100] According to the external entity dictionary, with the help of the double array Trie tree data structure, the words existing in the input text sequence are quickly searched from the dictionary. The named entity annotation method including but not limited to the BIO annotation method is used to mark the entity information (including boundaries and types) in the external entity dictionary of the words / terms in the input text sequence as external knowledge features.
[0101] Initialize the external vector matrix. Initialize an external feature vector matrix M-extra, which will then be used to vectorize external knowledge features and train with the neural network model. The size of the matrix is M×N, where M is the number of external knowledge labels and N is the dimension of the external knowledge features. In the example above, there are 5 types of external knowledge labels: B-ORG, I-ORG, B-CAR, I-CAR, O, then M=5; if the external knowledge feature dimension is 128, the matrix size is 5×128.
[0102] Input text vectorization. Here you can use but not limited to static word2vec methods (such as GloVe, word2vec, etc.), or dynamic methods such as ELMo, BERT, etc. to vectorize the input text to obtain the input text vector Input_embedding. Use M-extra to vectorize external knowledge features to obtain external knowledge feature vectors;
[0103] Connect the two vectors to get Final_input_embedding. This can effectively utilize the boundaries and types in external knowledge features. For example, in the above example, the organization name "Credit Card Center of a certain bank" in the external knowledge can effectively guide the model to learn the boundaries of the organization name, while the occupation name "Agent" can guide the model to learn some non-organization names to prevent misidentification.
[0104] Final_input_embedding is used as the input of the neural network and input into the neural network sequence labeling model, such as BiLSTM+CRF, to obtain the sequence labeling result. During training, the CRF loss function is used as the training target, and the gradient descent method is used for model training. During prediction, the Viterbi decoding method is used to decode the optimal sequence as the sequence labeling result.
[0105] Figure 2 FIG. 1 is a schematic diagram of the structure of a neural network information extraction system integrating an external dictionary according to an embodiment of the present invention. Figure 2 As shown, the system comprises:
[0106] The first unit is used to add an external entity dictionary to a word segmentation dictionary, perform Chinese word segmentation on an original input text to obtain an input text sequence, or directly use the original input text as an input text sequence; according to the external entity dictionary, search for words existing in the input text sequence from the external entity dictionary, and use a named entity tagging method to mark entity information in the external entity dictionary in the input text sequence as an external knowledge feature;
[0107] The second unit is used to initialize an external feature vector matrix, wherein the number of rows of the external feature vector matrix is the number of external knowledge labels, and the number of columns is the external knowledge feature dimension; vectorize the input text sequence to obtain an input text vector; and use the external feature vector matrix to vectorize external knowledge features to obtain an external knowledge feature vector;
[0108] The third unit is used to connect the input text vector with the external knowledge feature vector to obtain a final input vector; use the final input vector as the input of the neural network, input it into the neural network sequence labeling model for training and sequence labeling, and obtain a sequence labeling result; and convert the sequence labeling result into a semantic understanding result.
[0109] According to a third aspect of the embodiments of the present invention,
[0110] An electronic device is provided, comprising:
[0111] processor;
[0112] a memory for storing processor-executable instructions;
[0113] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.
[0114] A fourth aspect of the embodiments of the present invention is:
[0115] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.
[0116] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.
[0117] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A neural network information extraction method integrating an external dictionary, characterized in that: include: The external entity dictionary is added to the word segmentation dictionary, and the original input text is segmented into Chinese words to obtain an input text sequence, or the original input text is directly used as the input text sequence; according to the external entity dictionary, the words existing in the input text sequence are searched from the external entity dictionary, and the entity information in the external entity dictionary is marked in the input text sequence using a named entity tagging method as an external knowledge feature; Initialize an external feature vector matrix, where the number of rows of the external feature vector matrix is the number of external knowledge labels, and the number of columns is the external knowledge feature dimension; Vectorizing the input text sequence to obtain an input text vector; using the external feature vector matrix to vectorize external knowledge features to obtain an external knowledge feature vector; The input text vector is connected with the external knowledge feature vector to obtain a final input vector; the final input vector is used as an input of a neural network and input into a neural network sequence labeling model for training and sequence labeling to obtain a sequence labeling result; The sequence labeling result is converted into a semantic understanding result.
2. The method according to claim 1, characterized in that The external entity dictionary is added to the word segmentation dictionary, and the original input text is segmented into Chinese words to obtain an input text sequence, or the original input text is directly used as the input text sequence; according to the external entity dictionary, the words existing in the input text sequence are searched from the external entity dictionary, and the entity information in the external entity dictionary is marked in the input text sequence using a named entity tagging method, and the external knowledge features include: Searching the external entity dictionary for words present in the input text sequence according to the external entity dictionary in combination with a dual array tree data structure; The named entity tagging method using the BIO tagging method marks the entity information in the external entity dictionary of the characters / words in the input text sequence as external knowledge features.
3. The method according to claim 2, characterized in that The method further comprises: Based on the pre-collected external entity information, collect texts containing vocabulary content and general texts to train domain word vectors; Based on the domain word vector, candidate vectors are obtained from the user input text, and the candidate vectors are merged to build a vector representation list. The target candidate vector is obtained through a layer of linear neural network encoding.
4. The method according to claim 1, characterized in that: Initialize an external feature vector matrix, where the number of rows of the external feature vector matrix is the number of external knowledge labels, and the number of columns is the external knowledge feature dimension; Vectorizing the input text sequence to obtain an input text vector; The external knowledge feature vector is quantized using the external feature vector matrix to obtain an external knowledge feature vector including: Initialize an external feature vector matrix, where the number of rows of the external feature vector matrix is the number of external knowledge labels, and the number of columns is the external knowledge feature dimension; Perform one-hot encoding vectorization on the input text sequence to obtain an input text vector; According to the external entity information marked in the input text sequence, the corresponding external knowledge label is searched in the external feature vector matrix, and the vector of the corresponding row is extracted as the external knowledge feature vector of the external entity; The external knowledge feature vectors of all external entities in the input text sequence are concatenated to obtain a complete external knowledge feature vector.
5. The method according to claim 1, characterized in that The input text vector is connected with the external knowledge feature vector to obtain a final input vector; the final input vector is used as the input of the neural network and input into the neural network sequence labeling model for training and sequence labeling, and the sequence labeling results obtained include: The sequence annotation results are determined according to the following formula: prediction=CRF(BiLSTM(Final_input_embedding)); Among them, prediction represents the sequence labeling result, CRF() represents the processing result of conditional random field, BiLSTM() represents the processing result of bidirectional long short-term memory network, and Final_input_embedding represents the final input vector.
6. The method according to claim 5, characterized in that The method further includes training the neural network sequence labeling model, and the training method is as follows: The CRF loss function is used as the training objective, and the training objective is to maximize the proportion of the true path score in all paths: Among them, Loss represents the loss value of the CRF loss function, p real_path represents the fraction of the true path of the sample, P i represents the score of the i-th path, and n represents the number of all paths in the sample.
7. The method according to claim 1, characterized in that The final input vector is used as the input of the neural network and input into the neural network sequence labeling model for training and sequence labeling, and the sequence labeling results obtained include: The Viterbi decoding algorithm is used to decode the optimal sequence as the sequence labeling result.
8. A neural network information extraction system integrated with an external dictionary, used to implement the method of any one of claims 1 to 7, characterized in that: include: The first unit is used to add an external entity dictionary to a word segmentation dictionary, perform Chinese word segmentation on an original input text to obtain an input text sequence, or directly use the original input text as an input text sequence; according to the external entity dictionary, search for words existing in the input text sequence from the external entity dictionary, and use a named entity tagging method to mark entity information in the external entity dictionary in the input text sequence as an external knowledge feature; The second unit is used to initialize the external feature vector matrix, where the number of rows of the external feature vector matrix is the number of external knowledge labels, and the number of columns is the external knowledge feature dimension; Vectorizing the input text sequence to obtain an input text vector; using the external feature vector matrix to vectorize external knowledge features to obtain an external knowledge feature vector; The third unit is used to connect the input text vector with the external knowledge feature vector to obtain a final input vector; the final input vector is used as an input of the neural network, input into the neural network sequence labeling model for training and sequence labeling, and obtain a sequence labeling result; The sequence labeling result is converted into a semantic understanding result.
9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.
10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.