Target positioning named entity recognition and classification system based on multilayer feature fusion

Abstract

The invention relates to the technical field of natural language processing, and discloses a target positioning named entity recognition and classification system based on multilayer feature fusion. The system comprises a full-text feature extraction module used for extracting named entity features with context features and providing semantic information for subsequent processing; the high-level feature extraction module extracts multi-level semantic features from a text through local convolution operation, and fuses the features to obtain text information of different granularities. Then, the contextual features and the high-level features are fused together through the fusion module, a comprehensive feature set is generated, and powerful support is provided for the recognition module; the recognition module locates the named entity based on the features and calculates the initial position and length of the named entity. And finally, a classification module classifies the named entities by using the positioning information, and feeds back final results such as the position information, the length and the category to the user through an output module. And the position, the length and the category of each named entity in the target text can be accurately determined.

Description

Technical Field

[0001] The present invention relates to the technical field of natural language processing, and in particular to a target positioning named entity recognition and classification system based on multi-layer feature fusion. Background Art

[0002] Named Entity Recognition (NER) is a basic and key task in Natural Language Processing (NLP). Its main goal is to automatically identify entities with specific semantics from text, such as names of people, places, organizations, time, date, quantity, currency units, etc. NER technology is widely used in information extraction, search engines, question-answering systems, machine translation, sentiment analysis and other fields, especially when processing large amounts of unstructured text data. By identifying entities in text, NER lays the foundation for subsequent tasks such as semantic understanding, relationship extraction and knowledge graph construction.

[0003] At present, traditional named entity recognition technology is gradually replaced by neural network-based models, which can capture richer contextual information through training and learning, thereby improving the accuracy of recognition. However, although existing deep learning methods have achieved certain success in processing standard texts, there are still some challenges when facing complex texts (such as texts with a large number of synonyms, contextual ambiguity or irregular structure). For example, insufficient contextual information is a major problem faced by traditional named entity recognition technology. Named entities often rely on contextual information to determine their specific meaning and category. Secondly, traditional named entity recognition technologies mostly rely on manual feature extraction and simple dictionary matching, which limits their ability to express text. In deep learning methods, although word vectors are used to improve the feature representation ability of the model, these methods still have problems, that is, they cannot effectively utilize multiple levels of feature information (such as character-level, word-level, sentence-level features, etc.) at the same time, resulting in limited performance when processing complex texts. In addition, in many practical application scenarios, named entities do not always appear in the form of simple words, but may be nested in longer sentences or interfered with by other entity information, resulting in difficulty in positioning. For these complex texts, traditional named entity recognition technology usually cannot obtain effective location information at an early stage, thus affecting the overall recognition accuracy.

[0004] Therefore, there is an urgent need to invent a named entity recognition and classification technology to solve the problems of context ambiguity, insufficient feature fusion and difficulty in accurately locating named entities in the existing technology in processing complex texts. Summary of the invention

[0005] In view of this, the present invention proposes a target localization named entity recognition and classification system based on multi-layer feature fusion, aiming to solve the problems of context ambiguity, insufficient feature fusion and difficulty in accurately locating named entities in current technology in processing complex texts.

[0006] The present invention proposes a target localization named entity recognition and classification system based on multi-layer feature fusion, comprising:

[0007] A full-text feature extraction module, configured to extract named entity features with contextual features from an input text;

[0008] A high-level feature extraction module is configured to extract multi-level features in the input text based on local convolution, and extract text features of the granularity corresponding to the features of each level, and the high-level feature extraction module is further configured to fuse the text features of the granularity corresponding to the features of each level to generate a comprehensive feature set;

[0009] A fusion module, electrically connected to the full-text feature extraction module and the high-level feature extraction module, respectively, and configured to perform splicing and fusion between the context feature and the comprehensive feature set to generate recognition features;

[0010] A recognition module, electrically connected to the fusion module, the recognition module being configured to locate and recognize named entities in the input text based on the recognition features, and determine a starting position offset and a length of each named entity;

[0011] A classification module, electrically connected to the recognition module, the classification module being configured to classify the named entities according to the positioning result of the recognition module and determine the category of each named entity;

[0012] An output module is electrically connected to the classification module, and is configured to output the location information, length and category of each named entity in the input text.

[0013] Furthermore, when the full-text feature extraction module is configured to extract named entity features with context features from the input text, it includes:

[0014] The full-text feature extraction module is further configured to perform word segmentation processing on the input text and generate a plurality of semantic unit sequences;

[0015] The full-text feature extraction module is further configured to encode a plurality of the semantic unit sequences and extract feature sequences containing context dependencies;

[0016] The full-text feature extraction module is also configured to obtain the dependency relationship between the plurality of semantic unit sequences based on a self-attention mechanism, and determine the named entity features with the context features in the input text based on the dependency relationship between the plurality of semantic unit sequences and the feature sequence containing the context dependency relationship.

[0017] Furthermore, the high-level feature extraction module includes:

[0018] Local feature extraction unit, which captures adjacent token features based on a one-dimensional convolutional network;

[0019] Multi-layer convolution module, which extracts semantic features of different scales by expanding the number of channels and downsampling layer by layer;

[0020] The normalization and activation module performs batch normalization and nonlinear activation after the convolution operation to enhance the feature expression capability.

[0021] Furthermore, the fusion module includes:

[0022] A high-level feature downsampling unit, configured to perform downsampling processing on the high-level features generated by the high-level feature extraction module, so that the dimensions of the comprehensive feature set and the context feature are consistent;

[0023] A feature concatenation unit is configured to concatenate the context feature with the downsampled high-level feature along a feature dimension;

[0024] The weighted fusion unit is configured to fuse the spliced ​​features in a weighted manner and generate the recognition feature.

[0025] Furthermore, when the recognition module is configured to locate and recognize named entities in the input text based on the recognition features and determine the starting position offset and length of each named entity, it includes:

[0026] The recognition module is further configured to calculate the starting position offset and length of the named entity through linear transformation;

[0027] y0=(x center -x token )+Δ offset ;

[0028] y1=Length;

[0029] Among them, y0 is the entity starting position offset, y1 is the entity length, and x center is the center point of the entity, x token is the center point of the text, Δ offset is the offset;

[0030] The recognition module is further configured to determine a feature distribution of the named entity based on a starting position offset and a length of the named entity.

[0031] Furthermore, when the classification module is configured to classify the named entity according to the positioning result of the recognition module, it includes:

[0032] The classification module is further configured to calculate the category probability distribution using a softmax function according to the named entity location result provided by the recognition module:

[0033] y2=p(ci|x)(Wx+b);

[0034] Among them, y2 is the category probability distribution, W is the weight matrix, b is the bias, and P(ci|x) is the mathematical function of the probability distribution, where:

[0035]

[0036] Among them, zi is the score of the named entity ci, It is expressed as the exponential sum of the scores for all categories, which is used for normalization so that the sum of the probabilities of all categories is 1, and n is the total number of categories;

[0037] The classification module is also configured to determine a category label of the named entity according to the category probability distribution, and transmit the classification result to the output module.

[0038] Furthermore, when the output module is configured to output the location information, length and category of each named entity in the input text, it includes:

[0039] The output module is also configured to obtain a downstream task input format, and convert the starting position offset and length of each named entity in the input text and the category information of each named entity into the downstream task input format for output.

[0040] Compared with the prior art, the beneficial effect of the present invention is that: by setting a full-text feature extraction module and a high-level feature extraction module, the named entity features of the input text are extracted from the context dependency and the multi-level semantic granularity respectively, and these features are spliced ​​and fused through the fusion module. This comprehensive processing method of multi-level features improves the accuracy of named entity recognition, so that the system can more accurately capture the entity features in a complex semantic environment. Secondly, the starting position offset and length of the named entity are accurately located by the recognition module, and the category of the named entity is identified with the help of the classification module. This modular design with functional separation not only improves the efficiency of named entity recognition, but also enhances the stability and generalization ability of the system when processing multi-category named entities, and is suitable for text processing tasks in complex scenarios. Finally, the design of the output module can directly output the location information, length and category of the named entity, providing rich and intuitive data support for subsequent natural language processing tasks. This comprehensive information output method facilitates the integration of the system with other text analysis modules, improves the practicality and ease of use of the overall application, and is suitable for named entity recognition needs in multiple fields, such as information retrieval, question-answering system and knowledge graph construction. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present invention. Moreover, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings:

[0042] Figure 1 A functional block diagram of a target localization named entity recognition and classification system based on multi-layer feature fusion provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0043] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided in order to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art. It should be noted that, in the absence of conflict, the embodiments of the present invention and the features described in the embodiments can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0044] like Figure 1As shown, in some embodiments of the present application, this embodiment provides a target positioning named entity recognition and classification system based on multi-layer feature fusion, including: a full-text feature extraction module, a high-level feature extraction module, a fusion module, a recognition module, a classification module and an output module.

[0045] Specifically, the full-text feature extraction module is configured to extract named entity features with context features from the input text; the high-level feature extraction module is configured to extract multi-level features in the input text based on local convolution, and extract text features of the granularity corresponding to the features of each level, and the high-level feature extraction module is also configured to fuse the text features of the granularity corresponding to the features of each level to generate a comprehensive feature set; the fusion module is electrically connected to the full-text feature extraction module and the high-level feature extraction module respectively, and the fusion module is configured to splice and fuse the context features and the comprehensive feature set to generate recognition features; the recognition module is electrically connected to the fusion module, and the recognition module is configured to locate and recognize named entities in the input text based on the recognition features, and determine the starting position offset and length of each named entity; the classification module is electrically connected to the recognition module, and the classification module is configured to classify the named entities according to the results of the positioning of the recognition module, and determine the category of each named entity; the output module is electrically connected to the classification module, and the output module is configured to output the position information, length and category of each named entity in the input text.

[0046] It can be understood that the full-text feature extraction module and the high-level feature extraction module work together to extract the features of named entities from the global context and local semantics respectively. The full-text feature extraction module uses context information to extract the semantic dependency features of the text, while the high-level feature extraction module uses local convolution to perform multi-level processing on the input text to extract multi-dimensional semantic information from features of different granularities. Subsequently, the high-level feature extraction module is used to fuse these multi-level features to generate a comprehensive feature set, thereby achieving a comprehensive characterization of text features. In addition, the fusion module integrates the context features with the comprehensive feature set through a splicing operation to generate a unified recognition feature. Based on these recognition features, the recognition module locates and recognizes the named entities in the input text and determines the starting position offset and length of the named entity in the text. This process of feature fusion and positioning recognition relies on the high expressiveness of the features and the spatial inference ability of the model to ensure the accuracy of recognition. Finally, the classification module receives the positioning results provided by the recognition module, combines the recognized feature information, and performs category judgment and classified output on each named entity. Finally, the output module outputs the position information, length and category of each named entity recognized in the text. This process utilizes the category discrimination ability of the classification module and the intuitive presentation ability of the output module to complete the entire process from feature extraction to result output, and realizes a closed-loop technical architecture for named entity recognition and classification.

[0047] It can be seen that the complete process from feature extraction to entity recognition and classification is realized through the collaboration of multiple modules. The combination of the full-text feature extraction module and the high-level feature extraction module effectively improves the ability to comprehensively capture text semantic information; the fusion module enhances the expressiveness and recognition accuracy of the features by splicing and fusing contextual features and multi-level features; the recognition module and the classification module locate the named entities in turn and perform accurate classification, thereby realizing the accurate recognition and classification of named entities in the text; the output module presents the results in a structured manner for subsequent use.

[0048] Specifically, when the full-text feature extraction module is configured to extract named entity features with contextual features from the input text, it includes: the full-text feature extraction module is also configured to perform word segmentation on the input text and generate a number of semantic unit sequences; the full-text feature extraction module is also configured to encode a number of semantic unit sequences and extract feature sequences containing contextual dependencies; the full-text feature extraction module is also configured to obtain the dependencies between a number of semantic unit sequences based on a self-attention mechanism, and determine the named entity features with contextual features in the input text based on the dependencies between a number of semantic unit sequences and the feature sequences containing contextual dependencies.

[0049] It is understandable that the input text is divided into several semantic units (such as words or subwords), which are the basic components of the text. In this way, long texts can be converted into small units that are easier to handle, which is convenient for subsequent feature extraction and analysis. Secondly, the full-text feature extraction module extracts feature sequences containing contextual dependencies by encoding these semantic units. This means that the module not only focuses on the features of a single semantic unit, but also captures the interdependencies and contextual relationships between them, thereby improving the accuracy of named entity recognition. The encoding process may involve methods such as Word2Vec and BERT to learn the semantic representation of vocabulary. Finally, the module uses a self-attention mechanism to obtain the dependencies between semantic units. The self-attention mechanism allows the model to take into account the influence of all other words in the input sequence when processing each word, which helps to better understand the long-range dependencies in the text. By combining these dependencies and contextual information, the full-text feature extraction module can effectively determine the features of named entities and improve the effect and accuracy of named entity recognition.

[0050] Specifically: the high-level feature extraction module includes: a local feature extraction unit, which captures adjacent token features based on a one-dimensional convolutional network; a multi-layer convolution module, which extracts semantic features of different scales by expanding the number of channels and downsampling layer by layer; a normalization and activation module, which performs batch normalization and nonlinear activation after the convolution operation to enhance the feature expression capability.

[0051] It can be understood that the local feature extraction unit, which is based on a one-dimensional convolutional network, captures the features of adjacent tokens. The convolution operation gradually processes each local area in the input text through a sliding window, thereby capturing the local relationship between adjacent tokens. The advantage of this method is that it can effectively extract the relationship between each word and its neighboring words, especially for word sequence features in language models. The local convolutional network is particularly suitable for processing short-term dependencies because it can capture the connection between adjacent words within a certain window range, helping the model understand the meaning of each word in a specific context. Next, the multi-layer convolution module mentioned above extracts semantic features of different scales by expanding the number of channels and downsampling layer by layer. The design idea of ​​multi-layer convolution comes from the hierarchical feature learning of deep neural networks. Each layer of the network can further learn more abstract and representative features based on the features extracted by the previous layer. By expanding the number of channels layer by layer, the model can gradually form a more detailed and comprehensive semantic expression from low-level simple features to higher-level complex features. At the same time, the downsampling operation can help reduce the computational burden and retain the most important information by reducing the size of the feature map, thereby improving computational efficiency. Under the action of multi-layer convolution, the model can adaptively capture semantic features at different levels and handle complex contextual relationships. Finally, the role of the normalization and activation module is to stabilize the training process through batch normalization and enhance the expressiveness of features through nonlinear activation functions. Batch normalization mainly standardizes the input of each layer so that the data distribution remains consistent during the training process, thereby accelerating convergence and improving the stability of the model, especially in deep networks, which can effectively alleviate the problem of gradient disappearance or gradient explosion. In addition, nonlinear activation functions (such as ReLU, Leaky ReLU, etc.) introduce nonlinear transformations, allowing the model to learn and represent more complex nonlinear relationships, enhancing the neural network's ability to fit complex patterns. By combining the two, convolutional neural networks can better extract and express features.

[0052] Specifically, the fusion module includes: a high-level feature downsampling unit, which is configured to downsample the high-level features generated by the high-level feature extraction module so that the dimensions of the comprehensive feature set and the context features are consistent; a feature splicing unit, which is configured to splice the context features with the downsampled high-level features along the feature dimensions; and a weighted fusion unit, which is configured to fuse the spliced ​​features in a weighted manner and generate recognition features.

[0053] It can be understood that the high-level feature downsampling unit is configured to downsample the high-level features generated by the high-level feature extraction module so that the dimensions of the comprehensive feature set are consistent with those of the context features. The main purpose of the downsampling operation is to reduce the dimension of the high-level features so that they match the dimension of the context features, thereby facilitating subsequent fusion processing. Through downsampling, the model can effectively reduce the computational burden while retaining the key information in the high-level features. The downsampling method usually includes maximum pooling, average pooling, or dimensionality reduction operations through learning, thereby ensuring the simplification of information while not losing key information. Next, the function of the feature splicing unit is to splice the context features with the downsampled high-level features along the feature dimension. Feature splicing is a commonly used method in multimodal fusion. By directly splicing features from different sources in dimensions, multiple information can be combined into a more comprehensive feature expression. In this process, the model can take advantage of the complementarity of context features and high-level features to combine the advantages of both. This splicing operation allows the model to simultaneously utilize low-level context information and high-level abstract features, thereby improving the accuracy and robustness of named entity recognition. Finally, the weighted fusion unit generates the final recognition features by weighted fusion of the concatenated features. The technical principle of weighted fusion is that by assigning different weights to different features, the model can automatically learn the importance of each feature in the named entity recognition task. This weighting strategy enables the model to handle the importance of different features more flexibly, and can focus on certain features according to the needs of the task, thereby improving the overall recognition performance. Weighted fusion can fully tap the potential of various features and form recognition features with high expressiveness.

[0054] Specifically, when the recognition module is configured to locate and recognize named entities in the input text based on the recognition features and determine the starting position offset and length of each named entity, the recognition module includes: the recognition module is also configured to calculate the starting position offset and length of the named entity through linear transformation;

[0055] y0=(x center -x token )+Δ offset ;

[0056] y1=Length;

[0057] Among them, y0 is the entity starting position offset, y1 is the entity length, and x center is the center point of the entity, x token is the center point of the text, Δ offset is the offset; the recognition module is also configured to determine the feature distribution of the named entity based on the starting position offset and length of the named entity.

[0058] It can be understood that the starting position offset and length of the named entity are calculated by linear transformation. This process uses a positioning method based on position offset and length to accurately determine the position of the named entity in the text. The y0 in the formula represents the starting position offset of the entity, y1 represents the length of the entity, and by calculating the distance between the center point of the entity and the center point of the text, combined with the offset Δoffset, the model can accurately obtain the position of the named entity in the input text. Next, y0 (entity starting position offset) is obtained by calculating the formula y0 = (xcenter-xtoken) + Δoffset. Here, xcenter represents the center point of the named entity, xtoken represents the center point in the text, and Δoffset is an adjustable offset used to adjust the position of the named entity. This linear transformation ensures the accurate positioning of the starting position of the named entity by measuring the distance between the center points and adjusting the offset. In this way, the model can capture small changes in the relative position of the named entity in the text, improving the accuracy and reliability of recognition. Finally, the recognition module is also configured to further determine the feature distribution of the named entity based on the starting position offset and length of the named entity. This means that in addition to locating the position of the named entity, the model can also further analyze the feature distribution of the named entity based on the determined entity range. The principle of this operation is to analyze the relationship between the identified named entity and the context in the text by using the offset and length information, so as to deeply understand the semantic features of the entity.

[0059] Specifically, when the classification module is configured to classify the named entity according to the positioning result of the recognition module, the classification module is further configured to calculate the category probability distribution using the softmax function according to the named entity positioning result provided by the recognition module:

[0060] y2=p(ci|x)(Wx+b);

[0061] Among them, y2 is the category probability distribution, W is the weight matrix, b is the bias, and P(ci|x) is the mathematical function of the probability distribution, where:

[0062]

[0063] Among them, zi is the score of the named entity ci, It is expressed as the exponential sum of the scores for all categories, which is used for normalization so that the sum of the probabilities of all categories is 1, and n is the total number of categories; the classification module is also configured to determine the category label of the named entity according to the category probability distribution, and pass the classification result to the output module.

[0064] It can be understood that the classification module first obtains the location information of the named entity from the recognition module, including the starting position and length of the named entity. Then, based on this location information, the classification module further processes the features of each named entity and calculates the score of the entity belonging to different categories. By passing these scores to the softmax function, the module generates a normalized probability distribution that accurately represents the relative likelihood of each category. This method can not only evaluate the likelihood of a single category, but also help the system comprehensively determine which category is more representative of the named entity. Secondly, through the calculation of softmax, the classification module can select the one with the highest probability from multiple possible categories as the final category of the named entity. This process makes the classification decision more accurate, because the softmax function itself can effectively avoid bias when dealing with multi-classification problems, ensuring the fairness and rationality of the classification results. More importantly, this method can also adapt to various complex text environments and make in-depth distinctions between the subtle differences between multiple categories, thereby improving the robustness and performance of the classification system. Finally, the classification module passes the determined category label to the output module, which is responsible for feeding back the final classification results to the user or other application systems. This process ensures the efficient operation of the entire process from input text to final classification results, and also enables the system to provide accurate named entity classification information in practical applications.

[0065] Specifically, when the output module is configured to output the position information, length and category of each named entity in the input text, it includes: the output module is also configured to obtain the downstream task input format, and convert the starting position offset and length of each named entity in the input text and the category information of each named entity into the downstream task input format for output.

[0066] In the above embodiment, the present invention extracts the named entity features of the input text from the context dependency and multi-level semantic granularity respectively by setting a full-text feature extraction module and a high-level feature extraction module, and splices and fuses these features through a fusion module. This comprehensive processing method of multi-level features improves the accuracy of named entity recognition, so that the system can more accurately capture entity features in a complex semantic environment. Secondly, the starting position offset and length of the named entity are accurately located by the recognition module, and the category of the named entity is identified with the help of the classification module. This modular design with functional separation not only improves the efficiency of named entity recognition, but also enhances the stability and generalization ability of the system when processing multi-category named entities, and is suitable for text processing tasks in complex scenarios. Finally, the design of the output module can directly output the location information, length and category of the named entity, providing rich and intuitive data support for subsequent natural language processing tasks. This comprehensive information output method facilitates the integration of the system with other text analysis modules, improves the practicality and ease of use of the overall application, and is suitable for named entity recognition needs in multiple fields, such as information retrieval, question-answering system and knowledge graph construction.

[0067] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0068] The present application is described with reference to flowcharts and / or block diagrams of methods, devices (systems) and computer program products according to embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0069] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0070] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0071] 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 above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A target localization named entity recognition and classification system based on multi-layer feature fusion, characterized in that: include: A full-text feature extraction module, configured to extract named entity features with contextual features from an input text; A high-level feature extraction module is configured to extract multi-level features in the input text based on local convolution, and extract text features of the granularity corresponding to the features of each level, and the high-level feature extraction module is further configured to fuse the text features of the granularity corresponding to the features of each level to generate a comprehensive feature set; A fusion module, electrically connected to the full-text feature extraction module and the high-level feature extraction module, respectively, and configured to perform splicing and fusion between the context feature and the comprehensive feature set to generate recognition features; A recognition module, electrically connected to the fusion module, the recognition module being configured to locate and recognize named entities in the input text based on the recognition features, and determine a starting position offset and a length of each named entity; A classification module, electrically connected to the recognition module, the classification module being configured to classify the named entities according to the positioning result of the recognition module and determine the category of each named entity; An output module is electrically connected to the classification module, and is configured to output the location information, length and category of each named entity in the input text.

2. The target positioning named entity recognition and classification system based on multi-layer feature fusion as claimed in claim 1, characterized in that: When the full-text feature extraction module is configured to extract named entity features with contextual features from the input text, it includes: The full-text feature extraction module is further configured to perform word segmentation processing on the input text and generate a plurality of semantic unit sequences; The full-text feature extraction module is further configured to encode a plurality of the semantic unit sequences and extract feature sequences containing context dependencies; The full-text feature extraction module is also configured to obtain the dependency relationship between the plurality of semantic unit sequences based on a self-attention mechanism, and determine the named entity features with the context features in the input text based on the dependency relationship between the plurality of semantic unit sequences and the feature sequence containing the context dependency relationship.

3. The target positioning named entity recognition and classification system based on multi-layer feature fusion as claimed in claim 1, characterized in that: The high-level feature extraction module comprises: Local feature extraction unit, which captures adjacent token features based on a one-dimensional convolutional network; Multi-layer convolution module, which extracts semantic features of different scales by expanding the number of channels and downsampling layer by layer; The normalization and activation module performs batch normalization and nonlinear activation after the convolution operation to enhance the feature expression capability.

4. The target positioning named entity recognition and classification system based on multi-layer feature fusion as claimed in claim 1, characterized in that: The fusion module includes: A high-level feature downsampling unit, configured to perform downsampling processing on the high-level features generated by the high-level feature extraction module, so that the dimensions of the comprehensive feature set and the context feature are consistent; A feature concatenation unit is configured to concatenate the context feature with the downsampled high-level feature along a feature dimension; The weighted fusion unit is configured to fuse the spliced ​​features in a weighted manner and generate the recognition feature.

5. The target positioning named entity recognition and classification system based on multi-layer feature fusion as claimed in claim 1, characterized in that: When the recognition module is configured to locate and recognize named entities in the input text based on the recognition features and determine the starting position offset and length of each named entity, it includes: The recognition module is further configured to calculate the starting position offset and length of the named entity through linear transformation; y0=(x center -x token )+D offset ; y1=Length; Among them, y0 is the entity starting position offset, y1 is the entity length, and x center is the center point of the entity, x token is the center point of the text, Δ offset is the offset; The recognition module is further configured to determine a feature distribution of the named entity based on a starting position offset and a length of the named entity.

6. The target positioning named entity recognition and classification system based on multi-layer feature fusion as claimed in claim 1, characterized in that: When the classification module is configured to classify the named entity according to the positioning result of the recognition module, it includes: The classification module is further configured to calculate the category probability distribution using a softmax function according to the named entity location result provided by the recognition module: y2=p(ci|x)(Wx+b); Among them, y2 is the category probability distribution, W is the weight matrix, b is the bias, and P(ci|x) is the mathematical function of the probability distribution, where: Among them, zi is the score of the named entity ci, It is expressed as the exponential sum of the scores for all categories, which is used for normalization so that the sum of the probabilities of all categories is 1, and n is the total number of categories; The classification module is also configured to determine a category label of the named entity according to the category probability distribution, and transmit the classification result to the output module.

7. The target positioning named entity recognition and classification system based on multi-layer feature fusion as claimed in claim 6, characterized in that: When the output module is configured to output the position information, length and category of each named entity in the input text, it includes: The output module is also configured to obtain a downstream task input format, and convert the starting position offset and length of each named entity in the input text and the category information of each named entity into the downstream task input format for output.

Images