Sequence labeling method and system based on dynamic mapping knowledge domain enhancement

Abstract

The invention discloses a sequence labeling method and system based on dynamic mapping knowledge domain enhancement, and the method comprises the steps: S1, mining basic attribute information and related entity information from text information with sparse basic attributes and attribute missing, and constructing a mapping knowledge domain; s2, mapping each Token of the input sequence to an entity node of the knowledge graph, and extracting a local sub-graph; s3, performing feature extraction through a graph convolutional network to obtain an entity embedding feature of each Token; s4, fusing the original feature of each Token with the entity embedded feature; s5, inputting the fusion feature of each Token into the BiLSTM model to obtain an output feature of the input sequence; and S6, inputting the output features of the input sequence into the CRF model to obtain the final comprehensive features of the input sequence. According to the method, on the basis of BiLSTM and CRF, the GCN model is introduced to generate the characteristics based on the knowledge graph, the original input characteristics and the characteristics generated by the GCN are fused, sequence labeling is carried out in the text with sparse basic attributes and missing attributes, and good performance is achieved.

Description

Technical Field

[0001] The present invention relates to technical fields such as knowledge graphs, named entity recognition, deep learning, sequence labeling, and data analysis, and specifically to a sequence labeling method and system based on dynamic knowledge graph enhancement. Background Art

[0002] There is a large amount of text information on the Internet, but many text information often has characteristics such as high noise, non-standard format, and variable aliases. This makes the high-precision recognition and annotation of entity attributes in texts face severe challenges. The main purpose of sequence annotation is to accurately annotate these text information and obtain certain attribute values ​​and features corresponding to the text information for subsequent use. Sequence annotation was originally done manually one by one, which is very time-consuming and labor-intensive. Although some semi-automatic tools have been used as auxiliary tools in recent years, a lot of manual intervention is still required. At present, research on sequence annotation methods for text information mainly revolves around the following aspects: 1. The initial work of sequence labeling is mainly completed through manual labeling or hidden Markov model HMM, but there is currently a lack of methods to label the attributes that may exist in some texts with sparse basic attributes or scarce attributes.

[0003] 2. Although the method of combining BiLSTM+CRF for sequence labeling can extract key attribute information to a certain extent, it is still insufficient for sequence labeling when attributes in some texts are scarce.

[0004] A series of existing solutions provide theoretical basis and methodological support for the research on sequence labeling of text information in the clear network. However, these traditional methods are weak in modeling long-distance contexts and cannot understand the dependencies between sentences in complex text semantics such as those in the dark network. In particular, when there is a large distance between entities in long texts or sentences, the accuracy of sequence labeling will deteriorate. Although LSTM models can process time series data well, they often cannot effectively capture the bidirectional dependencies in sentences, and are prone to gradient vanishing or exploding problems during training, resulting in unsatisfactory results of the model when processing long text data. Summary of the invention

[0005] In view of the shortcomings in the prior art, the present invention provides a sequence labeling method and system based on dynamic knowledge graph enhancement, which utilizes the characteristics of sparse basic attributes and missing attributes of these text information, introduces a dynamically updateable knowledge graph, constructs enhanced features based on the input feature subgraph through the GCN model, and then uses the known BiLSTM+CRF model to achieve high-precision sequence labeling of the input text.

[0006] The present invention provides a sequence annotation method based on dynamic knowledge graph enhancement, the method comprising: S1. Mining basic attribute information and related entity information from text information with sparse basic attributes or missing attributes to construct a knowledge graph, wherein the knowledge graph includes a node set and an edge set, the node set includes attribute nodes corresponding to the basic attribute information and entity nodes corresponding to the related entity information, and the edge set includes entity-entity edges and entity-attribute edges; S2. Map each token of the input sequence to an entity node of the knowledge graph through named entity recognition, and extract the local subgraph related to each entity node; S3. Based on the node feature matrix and adjacency matrix of the local subgraph corresponding to each Token, the local subgraph corresponding to each Token is extracted through the graph convolutional network to obtain the entity embedding feature of each Token. S4, fusing the original features of each Token with the entity embedding features to obtain the fused features of each Token, wherein the original features are obtained by extracting features from the original Tokens of the input sequence; S5. Input the fusion feature of each Token into the BiLSTM model to generate the context feature of each Token, and obtain the output feature of the input sequence; S6. Input the output features of the input sequence into the CRF model, calculate the score of each possible label sequence, find the optimal label sequence with the highest score from the scores of each possible label sequence, and obtain the attribute embedding matrix based on the optimal label sequence through embedding matrix transformation. After fusing the entity embedding features of each Token with the corresponding attribute embedding matrix, they are concatenated with the corresponding original features to obtain the final comprehensive features of the input sequence.

[0007] Preferably, the basic attribute information includes: application field, product value, information content, information date, and information size; the related entity information includes: related countries, related regions, related persons, and related institutions.

[0008] Preferably, in step S2, the local subgraph of the entity node corresponding to the Token is represented as: ;

[0009] in, For the t Token corresponding to the entity node The relevant local subgraph, V is the node set of the knowledge graph, E is the edge set of the knowledge graph, For the node setV The nodes in For the edge set E The edge of i, j, t Is a positive integer.

[0010] Preferably, in step S3, the propagation rule of the graph convolutional network is: ; in, For the The node feature representation of the layer, is the adjacency matrix, = + is the adjacency matrix with self-loops added, is the logarithmic matrix, For the Layer weight matrix, initial node features Usually represented by word embeddings or entity attribute vectors, l is a positive integer; The graph convolutional network finally generates high-order features of the local subgraph It is expressed as: ; in, For the t The entity embedding features corresponding to the Token, t、T Is a positive integer.

[0011] Preferably, in step S4, assuming that the input sequence is , and its corresponding attribute sequence is , the fusion feature of each Token is expressed as: ; in, For the t The fusion characteristics of the Tokens, For the t The original characteristics of a token, To be the first t The entity embedding features are output after the tokens are mapped to the knowledge graph and processed by the graph convolutional network.

[0012] Preferably, in step S5, after the fusion features of each Token are input into the BiLSTM model, the context features of each Token are generated by the following method to obtain the output features of the input sequence: Generate the above features of each Token through the forward LSTM of the BiLSTM model; Generate the following features of each Token through the backward LSTM of the BiLSTM model; The previous and following features of each Token are bidirectionally concatenated to generate the contextual features of each Token and obtain the output features of the input sequence. ; ; ; ; in, For the t The fusion characteristics of the Tokens, For the t The above features of a Token, For the t The above features of a Token, For the t The contextual features of a token, is the output feature of the input sequence.

[0013] Preferably, in step S6, the final comprehensive features of the input sequence are obtained by the following method: ; ; ; in, is the score of each possible label sequence output by the CRF model, For attributes The corresponding weight vector, For the t The contextual features of a token, For the t The entity embedding features of the token, is the transfer score between adjacent attributes, and the optimal label sequence is , is the attribute embedding matrix, To combine the final comprehensive features of attribute information and external knowledge graph information, t、T Is a positive integer.

[0014] Preferably, in step S6, the Viterbi algorithm is used for decoding to find the optimal tag sequence with the highest score from the scores of each possible tag sequence.

[0015] Based on the same inventive concept, the present invention also provides a sequence annotation system based on dynamic knowledge graph enhancement, the system comprising: A knowledge graph construction module is used to mine basic attribute information and related entity information from text information with sparse basic attributes or missing attributes, and construct a knowledge graph, wherein the knowledge graph includes a node set and an edge set, the node set includes attribute nodes corresponding to the basic attribute information and entity nodes corresponding to the related entity information, and the edge set includes entity-entity edges and entity-attribute edges; A local subgraph extraction module is used to map each token of the input sequence to an entity node of the knowledge graph through named entity recognition, and extract a local subgraph related to each entity node; The feature extraction module is used to extract features from the local subgraph corresponding to each Token through a graph convolutional network based on the node feature matrix and adjacency matrix of the local subgraph corresponding to each Token, and obtain the entity embedding features of each Token; A feature fusion module is used to fuse the original features of each Token with the entity embedding features to obtain the fused features of each Token, wherein the original features are obtained by extracting features from the original Tokens of the input sequence; A context feature generation module is used to input the fusion feature of each Token into the BiLSTM model to generate the context feature of each Token and obtain the output feature of the input sequence; The comprehensive feature output module is used to input the output features of the input sequence into the CRF model, calculate the score of each possible label sequence, find the optimal label sequence with the highest score from the scores of each possible label sequence, and obtain the attribute embedding matrix based on the optimal label sequence through embedding matrix transformation. After fusing the entity embedding features of each Token with the corresponding attribute embedding matrix, they are spliced ​​with the corresponding original features to obtain the final comprehensive features of the input sequence.

[0016] Preferably, the basic attribute information includes: application field, product value, information content, information date, and information size; the related entity information includes: related countries, related regions, related persons, and related institutions.

[0017] Compared with the prior art, the present invention has the following beneficial effects: 1. This paper takes a large amount of illegally sold text information in the dark web as the research object. After data cleaning and text segmentation, on the basis of traditional sequence labeling using BiLSTM model and CRF model, it introduces dynamically updated knowledge graph and graph convolutional network (GCN) to enhance feature representation, so as to effectively deal with the problems of missing attributes, obscure expression, entity ambiguity and dynamic changes in the field of dark web text.

[0018] 2. The present invention constructs a dynamically updated knowledge graph based on the basic attribute information and related entity information in text information with sparse basic attributes or missing attributes, and obtains the knowledge subgraph corresponding to the input by mapping the original input to the knowledge graph, thereby identifying and generating certain attributes that may exist in the input text.

[0019] 3. Based on the BiLSTM model and the CRF model, the present invention introduces the GCN model to generate features based on the knowledge graph, and fuses the original input features with the features generated by GCN, achieving better performance in sequence labeling in texts with sparse basic attributes and missing attributes. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 A schematic diagram of a flow chart of a sequence annotation method based on dynamic knowledge graph enhancement provided by the present invention; Figure 2 The overall architecture diagram of a sequence annotation method based on dynamic knowledge graph enhancement provided by the present invention; Figure 3 A schematic diagram of the structure of a sequence annotation system based on dynamic knowledge graph enhancement provided by the present invention. DETAILED DESCRIPTION

[0021] 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 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.

[0022] The present invention is further described in detail below in conjunction with the accompanying drawings.

[0023] like Figure 1 As shown, an embodiment of the present invention provides a sequence labeling method based on dynamic knowledge graph enhancement, the method comprising: S1. From text information with sparse basic attributes or missing attributes, basic attribute information and related entity information are mined to construct a knowledge graph, wherein the knowledge graph includes a node set and an edge set. The node set includes attribute nodes corresponding to the basic attribute information and entity nodes corresponding to the related entity information. The edge set includes entity-entity edges and entity-attribute edges. In an embodiment of the present invention, basic attribute information includes: application field, product value, information content, information date, information size, etc. For example, the text may mention the scale of a specific data leak (information size), the field of traded items (such as network security), and related dates. Related entity information includes: related countries, related regions, related people, related institutions, etc. Dark web texts often involve international or cross-regional criminal gangs, financial institutions, or specific personal information. The present invention focuses on identifying these attributes to support subsequent sequence labeling tasks.

[0024] In the dynamically updated knowledge graph, each relevant entity information / basic attribute information is a network node of this knowledge graph. The edges of the knowledge graph are also divided into two categories: entity-entity edges and entity-attribute edges. If there is a relationship between entities, it means that there is an interactive behavior / event between the two entity nodes. If there is a relevant relationship between entities and basic attributes, it means that this entity has the attribute feature.

[0025] S2, through named entity recognition, map each token of the input sequence to the entity node of the knowledge graph, and extract the local subgraph related to each entity node; In the embodiment of the present invention, in step S2, the local subgraph of the entity node corresponding to the Token is represented as: ; in, For the t Token corresponding to the entity node The relevant local subgraph, V is the node set of the knowledge graph, E is the edge set of the knowledge graph, For a node collection V The nodes in Edge Set E The edge of i, j, t Is a positive integer.

[0026] The local subgraph includes The subgraph is the first-order neighbor node (that is, the attributes and entities directly associated with this node), and this subgraph can continue to be expanded to the second-order neighbors to capture a wider range of contextual relationships.

[0027] S3. Based on the node feature matrix and adjacency matrix of the local subgraph corresponding to each Token, the local subgraph corresponding to each Token is extracted through the graph convolutional network to obtain the entity embedding feature of each Token. In the embodiment of the present invention, in step S3, the propagation rule of the graph convolutional network is: ; in, For the The node feature representation of the layer, is the adjacency matrix, = + is the adjacency matrix with self-loops added, is the logarithmic matrix, For the Layer weight matrix, initial node features Usually represented by word embeddings or entity attribute vectors, l is a positive integer; High-level features of the local subgraph finally generated by the graph convolutional network It is expressed as: ; in, For the t The entity embedding features corresponding to the Token, t、T Is a positive integer.

[0028] S4, fuse the original features of each Token with the entity embedding features to obtain the fused features of each Token, wherein the original features are obtained by extracting features from the original Token of the input sequence; In the embodiment of the present invention, in step S4, it is assumed that the input sequence is , and its corresponding attribute sequence is , the fusion feature of each Token is expressed as: ; in, For the t The fusion characteristics of the Tokens, For the t The original characteristics of a token, To be the first t After each token is mapped to the knowledge graph, it is processed by the graph convolutional network and output as entity embedding features.

[0029] S5. Input the fusion features of each Token into the BiLSTM model to generate the context features of each Token and obtain the output features of the input sequence. In the embodiment of the present invention, in step S5, after the fusion feature of each Token is input into the BiLSTM model, the context feature of each Token is generated by the following method to obtain the output feature of the input sequence: Generate the above features of each token through the forward LSTM of the BiLSTM model; Generate the context features of each token through the backward LSTM of the BiLSTM model; The previous and following features of each Token are bidirectionally concatenated to generate the contextual features of each Token and obtain the output features of the input sequence. ; ; ; ; in, For the t The fusion characteristics of the Tokens, For the t The above features of a Token, For the t The above features of a Token, For the t The contextual features of a token, is the output feature of the input sequence.

[0030] The entity embedding features output by GCN are integrated into the existing BiLSTM+CRF sequence labeling solution. The role of the bidirectional LSTM model (BiLSTM) is to capture the contextual information of each token in the input sequence. For each input sequence, BiLSTM generates the contextual feature representation of each token through the combination of forward LSTM and backward LSTM, while CRF is responsible for modeling the global dependency between labels, thereby achieving high-precision sequence labeling.

[0031] S6. Input the output features of the input sequence into the CRF model, calculate the score of each possible label sequence, find the optimal label sequence with the highest score from the scores of each possible label sequence, and obtain the attribute embedding matrix based on the optimal label sequence through embedding matrix transformation. After fusing the entity embedding features of each Token with the corresponding attribute embedding matrix, they are concatenated with the corresponding original features to obtain the final comprehensive features of the input sequence.

[0032] In the embodiment of the present invention, in step S6, the final comprehensive features of the input sequence are obtained by the following method: ; ; ; in, is the score of each possible label sequence output by the CRF model, For attributes The corresponding weight vector, For the t The contextual features of a token, For the t The entity embedding features of the token, is the transfer score between adjacent attributes, and the optimal label sequence is , is the attribute embedding matrix, To combine the final comprehensive features of attribute information and external knowledge graph information, t、T Is a positive integer.

[0033] In the embodiment of the present invention, in step S6, the Viterbi algorithm is used to decode and find the optimal tag sequence with the highest score from the scores of each possible tag sequence. The combination of attribute information and external knowledge graph information provides richer and more accurate input features for subsequent tasks, which can be used as input for subsequent model training and classification.

[0034] like Figure 3 As shown, the present invention also provides a sequence annotation system based on dynamic knowledge graph enhancement, the system comprising: The knowledge graph construction module 100 is used to mine basic attribute information and related entity information from text information with sparse basic attributes or missing attributes, and construct a knowledge graph, wherein the knowledge graph includes a node set and an edge set, the node set includes attribute nodes corresponding to the basic attribute information and entity nodes corresponding to the related entity information, and the edge set includes entity-entity edges and entity-attribute edges; The local subgraph extraction module 200 is used to map each token of the input sequence to an entity node of the knowledge graph through named entity recognition, and extract the local subgraph related to each entity node; A feature extraction module 300 is used to extract features of the local subgraph corresponding to each Token through a graph convolutional network based on the node feature matrix and adjacency matrix of the local subgraph corresponding to each Token, so as to obtain the entity embedding features of each Token; A feature fusion module 400 is used to fuse the original features of each Token with the entity embedding features to obtain the fused features of each Token, wherein the original features are obtained by extracting features from the original Tokens of the input sequence; The context feature generation module 500 is used to input the fusion feature of each Token into the BiLSTM model to generate the context feature of each Token and obtain the output feature of the input sequence; The comprehensive feature output module 600 is used to input the output features of the input sequence into the CRF model, calculate the score of each possible label sequence, find the optimal label sequence with the highest score from the scores of each possible label sequence, and obtain the attribute embedding matrix based on the optimal label sequence through embedding matrix transformation. After fusing the entity embedding features of each Token with the corresponding attribute embedding matrix, they are spliced ​​with the corresponding original features to obtain the final comprehensive features of the input sequence.

[0035] Compared with the prior art, the advantages of the present invention are: At present, there is no effective attribute annotation system for text information with sparse basic attributes or missing attributes. Therefore, the present invention constructs a dynamically updated knowledge graph based on the basic attribute information and related entity information in these texts, and obtains the knowledge subgraph corresponding to the input by mapping the original input to the knowledge graph, thereby identifying and generating certain attributes that may exist in the input text.

[0036] Traditionally, sequence labeling mainly relies on manual labeling or hidden Markov model HMM for auxiliary labeling. Even with some semi-automatic tools such as HMM model, it still requires a lot of manual intervention. The labeling process is time-consuming and labor-intensive, and it is difficult to expand to large-scale data sets, especially in complex scenarios such as dark web and telegram groups. The labeling efficiency of text data is extremely low. This solution introduces the GCN model based on the BiLSTM model and CRF model to generate features based on the knowledge graph, and fuses the original input features and the features generated by GCN. It achieves good performance in sequence labeling in texts with sparse basic attributes and missing attributes.

[0037] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A sequence labeling method based on dynamic knowledge graph enhancement, characterized in that: The method comprises: S1. Mining basic attribute information and related entity information from text information with sparse basic attributes or missing attributes to construct a knowledge graph, wherein the knowledge graph includes a node set and an edge set, the node set includes attribute nodes corresponding to the basic attribute information and entity nodes corresponding to the related entity information, and the edge set includes entity-entity edges and entity-attribute edges; S2. Map each token of the input sequence to an entity node of the knowledge graph through named entity recognition, and extract the local subgraph related to each entity node; S3. Based on the node feature matrix and adjacency matrix of the local subgraph corresponding to each Token, the local subgraph corresponding to each Token is extracted through the graph convolutional network to obtain the entity embedding feature of each Token. S4, fusing the original features of each Token with the entity embedding features to obtain the fused features of each Token, wherein the original features are obtained by extracting features from the original Tokens of the input sequence; S5. Input the fusion feature of each Token into the BiLSTM model to generate the context feature of each Token, and obtain the output feature of the input sequence; S6. Input the output features of the input sequence into the CRF model, calculate the score of each possible label sequence, find the optimal label sequence with the highest score from the scores of each possible label sequence, and obtain the attribute embedding matrix based on the optimal label sequence through embedding matrix transformation. After fusing the entity embedding features of each Token with the corresponding attribute embedding matrix, they are concatenated with the corresponding original features to obtain the final comprehensive features of the input sequence.

2. The method according to claim 1, characterized in that The basic attribute information includes: application field, product value, information content, information date, and information size; the related entity information includes: related countries, related regions, related people, and related institutions.

3. The method according to claim 1, characterized in that In step S2, the local subgraph of the entity node corresponding to the Token is represented as: ； in, For the t Token corresponding to the entity node The relevant local subgraph, V is the node set of the knowledge graph, E is the edge set of the knowledge graph, For the node set V The nodes in For the edge set E The edge of i, j, t Is a positive integer.

4. The method according to claim 1, characterized in that: In step S3, the propagation rule of the graph convolutional network is: ； in, For the The node feature representation of the layer, is the adjacency matrix, = + is the adjacency matrix with self-loops added, is the logarithmic matrix, For the Layer weight matrix, initial node features Usually represented by word embeddings or entity attribute vectors, l is a positive integer; The graph convolutional network finally generates high-order features of the local subgraph It is expressed as: ； in, For the t The entity embedding features corresponding to the Token, t、T Is a positive integer.

5. The method according to claim 1, characterized in that In step S4, assume that the input sequence is , and its corresponding attribute sequence is , the fusion feature of each Token is expressed as: ； in, For the t The fusion characteristics of the Tokens, For the t The original characteristics of a token, To be the first t The entity embedding features are output after the tokens are mapped to the knowledge graph and processed by the graph convolutional network.

6. The method according to claim 1, characterized in that In step S5, after the fusion features of each Token are input into the BiLSTM model, the context features of each Token are generated by the following method to obtain the output features of the input sequence: Generate the above features of each Token through the forward LSTM of the BiLSTM model; Generate the following features of each Token through the backward LSTM of the BiLSTM model; The previous and following features of each Token are bidirectionally concatenated to generate the contextual features of each Token and obtain the output features of the input sequence. ； ； ； ； in, For the t The fusion characteristics of the Tokens, For the t The above features of a Token, For the t The above features of a Token, For the t The contextual features of a token, is the output feature of the input sequence.

7. The method according to claim 1, characterized in that In step S6, the final comprehensive features of the input sequence are obtained by the following method: ； ； ； in, is the score of each possible label sequence output by the CRF model, For attributes The corresponding weight vector, For the t The contextual features of a token, For the t The entity embedding features of the token, is the transfer score between adjacent attributes, and the optimal label sequence is , is the attribute embedding matrix, To combine the final comprehensive features of attribute information and external knowledge graph information, t、T Is a positive integer.

8. The method according to claim 1, characterized in that In step S6, the Viterbi algorithm is used to decode and find the optimal label sequence with the highest score from the scores of each possible label sequence.

9. A sequence annotation system based on dynamic knowledge graph enhancement, characterized in that: The system comprises: A knowledge graph construction module is used to mine basic attribute information and related entity information from text information with sparse basic attributes or missing attributes, and construct a knowledge graph, wherein the knowledge graph includes a node set and an edge set, the node set includes attribute nodes corresponding to the basic attribute information and entity nodes corresponding to the related entity information, and the edge set includes entity-entity edges and entity-attribute edges; A local subgraph extraction module is used to map each token of the input sequence to an entity node of the knowledge graph through named entity recognition, and extract a local subgraph related to each entity node; The feature extraction module is used to extract features from the local subgraph corresponding to each Token through a graph convolutional network based on the node feature matrix and adjacency matrix of the local subgraph corresponding to each Token, and obtain the entity embedding features of each Token; A feature fusion module is used to fuse the original features of each Token with the entity embedding features to obtain the fused features of each Token, wherein the original features are obtained by extracting features from the original Tokens of the input sequence; A context feature generation module is used to input the fusion feature of each Token into the BiLSTM model to generate the context feature of each Token and obtain the output feature of the input sequence; The comprehensive feature output module is used to input the output features of the input sequence into the CRF model, calculate the score of each possible label sequence, find the optimal label sequence with the highest score from the scores of each possible label sequence, and obtain the attribute embedding matrix based on the optimal label sequence through embedding matrix transformation. After fusing the entity embedding features of each Token with the corresponding attribute embedding matrix, they are spliced ​​with the corresponding original features to obtain the final comprehensive features of the input sequence.

10. The system according to claim 9, characterized in that The basic attribute information includes: application field, product value, information content, information date, and information size; the related entity information includes: related countries, related regions, related people, and related institutions.

Images