Traditional skill entity identification method, system, equipment and medium

Abstract

The invention discloses a traditional skill entity recognition method, system and device and a medium. The method comprises the steps of obtaining a to-be-recognized text; the text to be recognized is input into the trained traditional technical entity recognition model based on the entity boundary features and the attention sequence structure to recognize the traditional technical entities in the text to be recognized, the method, the system, the equipment and the medium can achieve extraction of non-abandoned entities, and accuracy and generalization are high.

Description

Technical Field

[0001] The present invention belongs to the field of big data technology and relates to a traditional art entity recognition method, system, equipment and medium. Background Art

[0002] As an important part of human cultural heritage, traditional folk culture is mostly spread in the form of handicrafts, performance activities, and skill inheritance. With the rapid development of digital technology, the research on entity extraction of traditional skills culture has received widespread attention. Some scholars have explored the inheritance methods of traditional skills by conducting in-depth analysis of the historical background, inheritance methods, and skill characteristics of traditional intangible cultural heritage skills. In contrast, foreign scholars focus on using digital storage technology to collect information on traditional skills, thereby deeply exploring the cultural connotations and skill characteristics of traditional skills.

[0003] In the application research of information extraction, named entity recognition (NER) identifies entity information of specific meaning from a large amount of corpus text, which is an important step to understand the meaning of text and analyze the deep relationship between entities. At present, scholars mainly study the recognition of intangible cultural heritage text entities from the aspects of entity boundary recognition, data sparsity, and sequence model application. On the one hand, methods based on supervised learning and semantic search can effectively solve the problem of identifying entity boundary information. In order to reduce the interference of redundant vocabulary, entity extraction methods of word combination have been proposed successively. On the other hand, sequence labeling methods based on deep learning frameworks have gradually become the mainstream methods of entity recognition and part-of-speech tagging. Traditional manual labeling relies on specific grammatical rules to extract key entity information, which is difficult to apply to the research of traditional skills intangible cultural heritage.

[0004] Although some progress has been made in the research of named entity recognition in intangible cultural heritage texts, there are still some challenges in the recognition of specific cultural entities. The diverse and complex entity information and semantic structure of traditional cultural skills have put forward higher extraction requirements for its entity recognition research. Existing sequence models capture the semantic features of entity information by learning the dependencies between text elements, and are thus widely used in knowledge extraction and intangible cultural heritage information recognition research, such as bidirectional long short-term memory networks, conditional random fields, hidden Markov models, and recurrent neural networks. However, when the above sequence labeling methods are applied to the research on traditional skill entity recognition with complex text features, it may be difficult to accurately identify its entity features and boundary attributes.

[0005] The traditional skills intangible cultural heritage named entity recognition mainly extracts key entity information such as the names of various traditional skills, related figures or the origin of skills from a large amount of text corpus. Traditional entity recognition relies on prior rules or data statistics to match text content, thereby learning sequence annotation. At present, relevant scholars at home and abroad have conducted research on the problem of traditional skills entity recognition from the perspective of the progress of traditional skills cultural research and intangible cultural heritage entity recognition.

[0006] Traditional folk culture is mostly in the form of oral transmission, handicrafts, performances, visual arts, etc. With the rise of digital technology, the research on traditional skills and culture at home and abroad has continued to deepen and expand, but the research focus is slightly different.

[0007] Domestic intangible cultural research mainly focuses on the cultural protection of traditional skills. By integrating the historical background, inheritance methods and skill characteristics of traditional intangible cultural heritage, the storage method of oral archives is improved, thus providing a new theoretical method and research direction for the protection of folk skills. With the continuous deepening of intangible cultural heritage research, intangible cultural heritage is no longer limited to simple conceptual theories. On the contrary, in-depth analysis is conducted on folk skills of handicrafts. Fang Lili took Jingdezhen ceramic handicrafts as the starting point to explore how intangible cultural heritage can cope with the trend of global integration, and emphasized that handicrafts are a symbol of post-industrial civilization and regional culture. Wei Rong proposed that ceramics, due to its unique characteristics such as texture, shape, and decorative patterns, provide an important reference for understanding traditional skills and their regional culture. In view of the interdependent relationship between cultural heritage and industrial development, Sun Tian proposed a variety of feasible development paths, providing new ideas for the industrialization development of traditional skills and intangible cultural heritage. In addition, some scholars have proposed strengthening the digital archive protection of intangible cultural heritage of skills. For example, Huang Shuiqing built a place name recognition model based on the pre-Qin corpus to better identify the local entity names of the birthplace of skills; Fan Tao and others extracted the text attributes of intangible cultural heritage through remote supervision and pre-training methods, which is helpful to promote the construction of knowledge maps and cultural communication of skill culture. Therefore, traditional skills, as an important carrier of folk art and handicrafts, are crucial to the protection and inheritance of intangible cultural heritage.

[0008] In contrast, although the research on intangible culture abroad started earlier, the exploration of traditional skills is still in the data storage stage. Due to the diversity of traditional skills, a single research method often cannot fully cover all aspects, so foreign scholars tend to use digital storage and visualization methods for research. For example, Shanxiu He and Zeming Duan used hypertext technology and augmented reality technology to collect and organize various data information and store cultural heritage in digital form. N. Grammalidis et al. constructed i-Treasures including video images, manual annotations and other metadata, providing a rich data foundation for the study of various folk skills. Fan Tao et al. analyzed the intangible cultural heritage images and proposed a method based on attention mechanism and hierarchical fusion technology to combine data with text descriptions, thereby helping researchers better understand the semantic and visual features of images and provide technical support for the inheritance of intangible cultural heritage. Choi T et al. used a multi-sensor data collection method to capture the dynamic data of intangible cultural heritage performances in real time, and verified the feasibility of the method through statistical analysis. Chen et al. used case analysis to discuss in detail the historical inheritance of paper-cutting art and the application value of digital skills in the dissemination of intangible cultural heritage, and analyzed the dissemination methods of paper-cutting skills, thereby showing the diversity of intangible cultural heritage digital research.

[0009] Through the above analysis, domestic and foreign scholars have shown great interest in the study of traditional skills and intangible cultural heritage, and have explored the cultural connotation, skill characteristics and dissemination channels of traditional skills from multiple dimensions. In order to effectively convey the culture of folk handicrafts, some scholars focus on cultural heritage and handicrafts themselves, and especially point out the innovation and improvement in digital storage. These studies not only provide important theoretical support for the protection and inheritance of traditional skills, but also pave the way for in-depth exploration of traditional folk skills and culture.

[0010] In the field of information extraction, named entity recognition (Ner) has made significant progress as a key step in obtaining semantic relationships between entities from unstructured data. Traditional named entity recognition methods mainly perform entity recognition through rule grammar, unsupervised learning, and feature-based learning. Liu Liu and Wang Dongbo discussed the Chinese text annotation situation from the perspective of rule statistics and machine learning representation methods, respectively, and believed that the selection of text feature representation can significantly improve the entity recognition effect. Zhentao et al. faced the limitation of traditional extraction methods to artificial feature engineering, so they compared traditional recognition methods with typical deep learning models to verify the overall recognition effect of the model. When extracting geological related entities, Qinjun et al. found that there were irregular information in their text corpus, such as nested entities and redundant entities, which seriously affected the entity recognition effect. Scholars combined deep learning pre-training models to represent vectors from byte sequences to obtain entity category results, but they were limited to geological related fields.

[0011] In the field of traditional intangible cultural heritage, entity recognition focuses more on identifying entities of specific significance, such as intangible cultural heritage projects, inheritors, regions, etc., through entity extraction and structuring of intangible cultural heritage text data, providing managers with structured data support on the historical inheritance and regional distribution of intangible cultural heritage, and becoming an important part of building the intangible cultural heritage knowledge graph. With regard to intangible cultural heritage text data, scholars have explored and improved it from many aspects. On the one hand, the accuracy of entity recognition is improved by improving entity extraction, supervised learning and semantic search methods. For example, Fan Qing et al. pointed out the important applications of entity recognition in the field of intangible cultural heritage, such as graph construction, relationship extraction and data storage, and improved related methods. In the field of traditional Chinese medicine text, in order to improve the accuracy of named entity recognition of traditional Chinese medicine text, Yang Hang, Peng Yehui et al. constructed a BRL medical case recognition model that integrates multiple medical text features by manually annotating medical case corpus, and emphasized that the text integrating related words can effectively improve the accuracy of the model in entity recognition. On the other hand, in response to problems such as blurred entity boundaries and inter-entity nesting, scholars such as Wang Jin and Wang Mengqi believe that named entity recognition mainly extracts and annotates key information from a large amount of unstructured text. The multi-head attention entity recognition method based on word-word union can effectively reduce the interference of redundant vocabulary and improve the recognition performance of text entities. At the same time, some scholars have also paid attention to the limitations of traditional word segmentation methods. A single word segmentation method is easily affected by word meaning and lack of syntactic information. In order to overcome these shortcomings, Zhang and Yang named entity recognition methods based on sequence models, by splicing characters and word vectors to improve the accuracy of entity semantic labels. Starting from syntactic information and prior information, Xu Xiaobo, Wang Tao and others integrated dictionary and glyph features in the representation layer, and used ON-LSTM and sequence encoding methods to enhance the entity recognition effect, thereby making up for the shortcomings of text recognition methods.

[0012] With the development of deep learning technology, entity recognition methods based on sequence models have gradually become mainstream methods. Compared with traditional pipeline models, entity extraction methods that integrate sequence models can establish a unified model for specific research objects so that different subtasks can interact with each other. Studies by Cao Mingyu, Yang Zhihao, etc. show that the associated data extraction method based on deep neural networks has a higher degree of information recognition. Specifically, sequence models similar to hidden Markov models (HMM), long short-term memory networks (LSTM), conditional random fields (CRF), etc., capture the temporal relationship and front-to-back dependencies of entity data, model traditional skill entities, learn the dependencies between elements of text sequences, and accurately apply them to the recognition of traditional skill-related intangible cultural heritage projects. Liu Peiyu, Xun Jing, etc. flexibly use hidden Markov models to accurately extract the subjective and objective features of text from complex corpora and effectively identify the entity features of given texts. Zhu Jizhao, Zhao Yilin, etc. combine attention mechanisms with long short-term memory networks to construct EPSA entity relationship extraction models. This sequence model calculates the correlation between entities and sentences, explicitly integrates entity information, effectively solves the problem of single entity overlap, and obtains an enhanced sentence representation. Hasna uses transfer learning and pre-trained models to build a BERT-CRF joint extraction model to perform term extraction and polarity detection on multi-language texts. Xiang, Shi, and others found that the information extraction task was significantly affected when faced with the difficult problem of complex data and unclear detection features. In order to solve the problem of data sparsity, they built the APT text corpus and combined the BiGRU and CRF models according to the characteristics of the gated recurrent unit to quickly capture the long-term dependencies of sequence data and achieve efficient sequence labeling of text entities.

[0013] Overall, domestic and foreign scholars have conducted research on the problem of intangible cultural heritage entity extraction from multiple perspectives, such as entity boundary recognition, data sparsity, and the application of sequence models, and have tended to improve their understanding of entity internal attributes, irregular text of intangible cultural heritage, and contextual environment. Although researchers have conducted entity recognition research based on sequence models, they have not considered the semantic structure and cultural background unique to the field of intangible cultural heritage, and the existing sequence methods still need to be optimized in capturing deep semantic information. Secondly, some methods do not fully consider the impact of contextual environment and entity boundary attributes on entity recognition, which affects the accuracy and generalization of intangible cultural heritage entity extraction. Summary of the invention

[0014] The purpose of the present invention is to overcome the shortcomings of the above-mentioned prior art and provide a method, system, device and medium for traditional art entity recognition. The method, system, device and medium can realize the extraction of intangible cultural heritage entities with high accuracy and generalization.

[0015] In order to achieve the above object, the present invention adopts the following technical scheme:

[0016] In one aspect, the present invention provides a method for identifying a traditional art entity, comprising:

[0017] Get the text to be recognized;

[0018] The text to be recognized is input into a trained traditional art entity recognition model based on entity boundary features and attention sequence structure to recognize the traditional art entities in the text to be recognized.

[0019] The conventional art entity recognition method of the present invention is further improved in that:

[0020] Furthermore, before inputting the text to be recognized into the trained traditional entity recognition model based on entity boundary features and attention sequence structure, the method further includes:

[0021] Construct a traditional skills dataset;

[0022] Construct a traditional art entity recognition model based on entity boundary features and attention sequence structure;

[0023] The traditional skills data set is used to train a traditional skills entity recognition model based on entity boundary features and attention sequence structure, thereby obtaining a trained traditional skills entity recognition model based on entity boundary features and attention sequence structure.

[0024] Furthermore, the process of constructing the traditional skills dataset is as follows:

[0025] Collect relevant data from national and local intangible cultural heritage websites of traditional skills through crawler technology;

[0026] Preprocess the collected relevant corpus to obtain several pieces of text data;

[0027] The traditional skills entities are annotated for each text data, and then a traditional skills dataset is constructed based on the annotated text data.

[0028] Furthermore, the traditional skills entity recognition model based on entity boundary features and attention sequence structure includes a text representation layer, a sequence modeling layer and a label decoding layer, wherein the text representation layer determines the BIO type according to the word position attribute, and uses the attention mechanism to perform text vector representation of traditional skills for the entity character information; the sequence modeling layer dynamically captures the context information and dependency relationships of the intangible cultural heritage entities through the BiLSTM unit, and obtains the possible entity labels and probability distributions for each character in the text sequence; the label decoding layer accepts the probability distribution of sequence modeling, and uses the potential association between the character position and the entity category label to predict the traditional skills entities at each position in the intangible cultural heritage text sequence.

[0029] Furthermore, the text representation layer includes a word embedding layer, a multi-layer feedforward network and a multi-head attention mechanism.

[0030] Furthermore, the loss function in the process of training the traditional skill entity recognition model based on entity boundary features and attention sequence structure using the traditional skill dataset is:

[0031]

[0032] In a second aspect of the present invention, the present invention provides a traditional art entity recognition system, comprising:

[0033] An acquisition module is used to acquire the text to be recognized;

[0034] The recognition module is used to input the text to be recognized into the trained traditional art entity recognition model based on entity boundary features and attention sequence structure to recognize the traditional art entities in the text to be recognized.

[0035] The conventional art entity recognition system of the present invention is further improved in that:

[0036] Furthermore, the traditional skills entity recognition model based on entity boundary features and attention sequence structure includes a text representation layer, a sequence modeling layer and a label decoding layer, wherein the text representation layer determines the BIO type according to the word position attribute, and uses the attention mechanism to perform text vector representation of traditional skills for the entity character information; the sequence modeling layer dynamically captures the context information and dependency relationships of the intangible cultural heritage entities through the BiLSTM unit, and obtains the possible entity labels and probability distributions for each character in the text sequence; the label decoding layer accepts the probability distribution of sequence modeling, and uses the potential association between the character position and the entity category label to predict the traditional skills entities at each position in the intangible cultural heritage text sequence.

[0037] In a third aspect of the present invention, a computer device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the traditional art entity recognition method when executing the computer program.

[0038] In a fourth aspect of the present invention, the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the traditional art entity recognition method are implemented.

[0039] The present invention has the following beneficial effects:

[0040] During specific operation, the traditional skill entity recognition method, system, device and medium described in the present invention integrate entity boundary features into text feature representation based on the pre-trained attention mechanism to construct a traditional skill intangible cultural heritage entity recognition model that integrates entity boundary attribute features and attention sequence structure, thereby identifying complex types of traditional skill entity features. It should be noted that the present invention integrates entity boundary attributes into the training model, uses the attention mechanism to generate a feature vector representation of the target dimension for the intangible cultural heritage text corpus, captures the temporal correlation features of the text through the sequence modeling layer and the label decoding layer, and uses the hidden state vector representation to predict the optimal label for each position in the text sequence, thereby ensuring that the marked traditional skill label is globally optimal. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The accompanying drawings constituting a part of the present invention are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the accompanying drawings:

[0042] Figure 1 This is the structural diagram of the ER-BFAS text classification model;

[0043] Figure 2 This is the structural diagram of the text representation layer;

[0044] Figure 3 Schematic diagram of the global feature extraction process;

[0045] Figure 4 This is the structural diagram of the label decoding layer;

[0046] Figure 5 This is a comparison chart of the training situation before and after parameter adjustment;

[0047] Figure 6 Renderings of traditional craft physical evaluations of each part of the structure;

[0048] Figure 7 This is a diagram of the training status of each model on the data set;

[0049] Figure 8 This is a graph showing the results of a comprehensive evaluation of each indicator. DETAILED DESCRIPTION

[0050] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiment of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiment of the present invention. Obviously, the described embodiment is only an embodiment of a part of the present invention, not all embodiments, and is not intended to limit the scope of the present invention. In addition, in the following description, the description of well-known structures and technologies is omitted to avoid unnecessary confusion of the concepts disclosed in the present invention. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work should fall within the scope of protection of the present invention.

[0051] The accompanying drawings show schematic diagrams of structures according to embodiments disclosed in the present invention. These figures are not drawn to scale, and some details are magnified and some details may be omitted for the purpose of clear expression. The shapes of various regions and layers shown in the figures and the relative sizes and positional relationships therebetween are only exemplary, and may deviate in practice due to manufacturing tolerances or technical limitations, and those skilled in the art may additionally design regions / layers with different shapes, sizes, and relative positions according to actual needs.

[0052] Embodiment 1

[0053] refer to Figure 1 The traditional art entity recognition method of the present invention comprises the following steps:

[0054] 1) Construction and feature analysis of traditional skills dataset;

[0055] 11) Original corpus collection;

[0056] Huang Yonglin and others clarified in the "National Cultural Heritage Masterpiece Application and Assessment Method" that most of the traditional skills and intangible cultural heritage are the manifestations of various traditional cultures that have been passed down from generation to generation by various ethnic groups and are closely related to life, such as folk activities, performing arts, cultural spaces or utensils, etc. This characteristic has promoted the recognition of traditional skills semantic entities. Research. Therefore, the present invention focuses on traditional skills. Unlike traditional classification methods, the corpus divides the original data of traditional skills based on the characteristics of skills and forms of expression. Relevant corpora of national and local intangible cultural heritage websites of traditional skills are collected using relevant crawler technology. Text content that can be applied to the recognition of traditional skills and intangible cultural heritage entities is extracted based on semantic features. After pre-processing steps such as eliminating noise data and deleting key information, 6341 pieces of original data are obtained, and nearly 6000 pieces of text data are obtained through further screening. The text corpus data of traditional skills is shown in Table 1.

[0057] Table 1 Some traditional skills corpus data

[0058]

[0059] 12) Named entity annotation;

[0060] The traditional skills intangible cultural heritage named entities have rich semantic information, covering the production process, materials and tools, practical uses and other factors of the skills. This embodiment selects 9 main representative traditional skills entity categories based on the actual use and resource types of traditional skills. The specific traditional skills entity labels and their identification examples are shown in Table 2. These traditional skills entity types have their own characteristics, almost covering the traditional folk handicrafts including "textile", "sculpture", "paper cutting", "ceramics" and "modeling". By identifying the representative skill entities and their label types, it is convenient for people to intuitively understand the entity characteristics of each skill category, and provide data support for the research on the recognition of traditional skills intangible cultural heritage entities.

[0061] Table 2

[0062]

[0063]

[0064] In order to better understand the semantic features of traditional skill entities, this embodiment extracts the association information between characters and text sequences based on the BIO (Beginning Inside Outside, BIO) annotation method of sequence characters. In the process of intangible cultural heritage sequence annotation, characters are used as the smallest unit, and each character containing the traditional skill entity vocabulary is annotated as "BX", "IX" and "O" in turn. Among them, "BX" means that the character is the starting character of the annotated entity; "IX" means that the character is the middle or end character of the annotated entity; "O" means that it does not belong to any established traditional skill type. Combined with the identified entity labels, the traditional skill named entity category labels suitable for the fusion of the attention mechanism and the sequence model are further determined. The specific labels and their meanings are shown in Table 3.

[0065] Table 3

[0066]

[0067]

[0068] When annotating the named entities of the traditional craft corpus, words that are semantically close to the entity type label are selected as the annotation unit. Traverse the same text sequence and mark the intangible cultural heritage entities in the current sentence one by one until the last character of the sentence is used as the end mark. Taking the corpus in Table 4 as an example, "Zhidan embroidery retains the more original embroidery style and characteristics of my country, and mostly uses the wrapping embroidery method." In the sentence, (Zhidan embroidery, B-Embroidery I-EmbroideryI-EmbroideryI-Embroidery) can be marked; if its boundary attribute characteristics are to be identified, it is marked as (Embroidery, Zhidan embroidery, 0).

[0069] Table 3

[0070]

[0071] 2) Construct a traditional skill entity recognition model based on entity boundary features and attention sequence structure (ER-BFAS model);

[0072] From the perspective of vocabulary enhancement, this embodiment combines the boundary attributes of intangible cultural heritage entities with the pre-training model, performs word vector extraction at the semantic level of multi-head attention, and greatly improves the model's ability to recognize entity boundaries. At the same time, the BiLSTM and CRF sequence models are introduced to further reduce the interference of redundant vocabulary and accurately recognize entities in traditional art texts. Therefore, the present invention integrates the attention mechanism and sequence model features, effectively solves the problem of entity boundary recognition model and entity nesting, and constructs a multi-head attention sequence entity recognition model for traditional art. The specific ER-BFAS model structure is shown in the figure. The ER-BFAS model includes a text representation layer, a sequence modeling layer, and a label decoding layer (CRF layer). Among them, the text representation layer determines the BIO type according to the word position attribute, and uses the attention mechanism to perform text vector representation of traditional art on the entity character information; through the BiLSTM sequence structure, the context information and dependency of the intangible cultural heritage entity are dynamically captured, and the possible entity label and probability distribution of each character in the text sequence are obtained; the CRF layer accepts the probability distribution of sequence modeling, and uses the potential association between the character position and the entity category label to predict the traditional art entity at each position in the intangible cultural heritage text sequence. The following is a detailed introduction to each layer:

[0073] 21) Text presentation layer

[0074] This embodiment selects the text-label joint embedding method to combine the intangible cultural heritage entities of each mark with the category label for joint mapping. Figure 2The text representation part is based on a pre-trained bidirectional encoder, which mainly includes a word embedding layer, a multi-layer feedforward network, and a multi-head attention mechanism. When processing text sequences, the characters in the intangible cultural heritage text sequence are mapped one by one into a fixed vector space through a pre-trained character-word embedding layer, thereby obtaining a text representation containing vocabulary-level information. The multi-head attention mechanism captures the semantic information of different positions in the text sequence at the same time through multiple sets of parallel attention mechanisms, and the stacked feedforward network layers are used to enhance the representation ability of the pre-trained model.

[0075] For entity information in the intangible cultural heritage text sequence, the entity boundary information is spliced ​​with the text body by extracting the corresponding position features and type features of the entity, and a complete input representation is formed by jointly embedding words. In order to identify the entity label features of the entity position and splicing in the text sequence, this model introduces specific identifiers "[CLS]" and "[SEP]" at the start and end positions of the text to mark the beginning and end of the text. Therefore, for any text sequence S, the maximum sequence length is N, and there is a corresponding character encoding representation t, then the input corpus is represented as:

[0076] S={[CLS],t1,t2,...,t N ,[SEP]}(1)

[0077] The boundary information of the entity is f, which refers to the position information feature of the specific entity in the sentence. By traversing the labels "BX" or "IX" in the sentence, the starting position Label of the traditional craft entity is calculated respectively. start and the end position Label end , thus forming the boundary information features of traditional skill entities. Finally, the text sequence is concatenated with the entity boundary attributes and entity labels to form the input feature S′ of the text representation layer.

[0078]

[0079] Among them, I(Bx,i) is the indicator function, which is used to record the first position where the "BX" label appears, ∑s i It indicates the number of times the "IX" label appears multiple times, and Label indicates the type label of the entity.

[0080] The traditional named entity recognition task has a high dependence on the model structure and requires a complex deep network structure to capture the key information in the intangible cultural heritage corpus. Among the many entity feature extraction structures, compared with the convolutional neural network, the BERT model with integrated attention mechanism can better adapt to the recognition task of traditional skills intangible cultural heritage entities through pre-training and fine-tuning of the mask mechanism. The attention mechanism based on entity boundary information constructs three different linear transformations (Q, K, V) to represent the current entity, the relationship between entities, and the unit feature vector expression respectively, and uses the multi-head attention layer to calculate the encoding output T[out] of the current attention layer respectively, and repeatedly iterates to perform deep feature extraction. The feature expression matrix can be expressed as:

[0081]

[0082] Therefore, for the input text sequence S′, given the feature dimension d k , then the activation function is used to calculate each head i The attention weight MultiHead(Q,K,V) is specifically expressed as:

[0083]

[0084] Among them, W Q ,W K ,W V is the learned weight matrix, W O Used to concatenate the output to the target dimension, Q i ,K i ,V i It is a matrix divided according to the multi-head attention dimension.

[0085] Finally, all heads are concatenated along the last feature dimension to form the output T of the attention layer out .

[0086] T out =MultiHead(Q,K,V)(5)

[0087] 22) Sequence Modeling Layer

[0088] As a kind of semantically complex text sequence data, the semantic features and the dependencies between entities should be fully considered in the entity extraction process. Common sequence annotation models include maximum entropy hidden Markov model (MEMM), gated recurrent neural network (GRU) and bidirectional long short-term memory network (BiLSTM). For example, Wang Jintao and Qin Ang et al. (Wang Jintao, Qin Ang, Zhang Yuan, et al. Chinese medical entity recognition based on attention enhancement and feature fusion [J / OL]. Computer Engineering: 1-11 [2024-06-25].) integrated the pinyin and stroke features of the entity through the iterative attention module, and used the long short-term memory network to capture the semantic features of the Chinese medical entity before and after, thereby improving the model's ability to recognize characters. Compared with the structures of MEMM and GRU, BiLSTM can bidirectionally traverse the contextual information of the intangible cultural heritage text sequence, thereby making up for the shortcomings of the maximum entropy MEMM model traversal. Combined with the task examples of traditional skill sequence annotation, the cyclic structure may have the problems of gradient explosion and gradient disappearance when processing long text sequences.

[0089] This embodiment refers to the sequence-based document-level entity recognition method mentioned by Ren Anqi et al. (Ren Anqi, Liu Lin, Wang Hailong, et al. A review of research on text-oriented entity relationship extraction [J / OL]. Computer Science and Exploration: 1-28 [2024-06-25].), and uses a bidirectional long short-term memory network for sequence annotation of intangible cultural heritage texts, remodels and encodes the text features that integrate entity boundary attributes, and further captures character-level semantic information, so that ER-BFAS can effectively identify the boundary information and type of traditional skills entities. Taking the traditional skill paper-cutting entity "Qianyang Paper-cutting" as an example, LSTM only recognizes the previous feature information "Yang" of "cut", while the bidirectional LSTM simultaneously captures the bidirectional information of "Yang" and "paper". Therefore, the sequence modeling layer mainly includes BiLSTM units, which realize the feature encoding and semantic feature capture of the traditional skill intangible cultural heritage text sequence through forward calculation, backward calculation and merging of hidden states. The specific process is as follows: Figure 3 shown.

[0090] In the LSTM basic unit, the calculation process of the hidden state at each time step t depends on the current time step h t and the hidden state x at the previous time step t. Specifically, the forward calculation process of LSTM mainly controls the information flow and storage of text sequences through the input gate, forget gate and output gate. In the sequence labeling task of traditional skill entity recognition, the calculation of the forward hidden state not only depends on the input at the current moment, but also is related to all previous inputs, so that the bidirectional LSTM can accurately identify the long-term dependencies contained in the traditional skill-related skill sequence. The backward hidden state calculation is similar to the forward calculation process, but the calculation direction is opposite. Taking the forward calculation of LSTM as an example, the information flow and storage of the text sequence are mainly controlled by the input gate, forget gate and output gate. In the LSTM basic unit, the calculation process of the hidden state at each time step t depends on the current time step h. t and the hidden state x at the previous time step t , the update process of the three gate control units and memory units can be expressed as:

[0091]

[0092] Among them, x t is the text sequence input at the current time step, f t ,i t ,o t ,C t They represent the forget gate, input gate, output gate and updated memory unit respectively.

[0093] Calculate the forward hidden state and backward hidden state of the current time step respectively, and obtain the overall hidden state h of each time step by merging the hidden states. t , and its calculation process can be expressed as shown in formula (7). At this time, the context information of each character in the text sequence is captured, so that the model can better understand the previous and next dependencies and semantic information between the individual characters of the traditional art entity.

[0094]

[0095] 23) Label decoding layer

[0096] After passing through the attention mechanism and the BiLSTM layer successively, the attention weights on different parts of the text sequence are allocated, so as to obtain the label score vector of each word in the intangible cultural heritage text sequence. The label decoding layer calculates all possible labels of the given text sequence and selects the label entity with the largest conditional probability as the output. In order to avoid the label conflict problem caused by the sliding window, CRF is selected in this embodiment to output possible predicted labels from the traditional skills text sequence. For example, in the traditional paper-cutting skill corpus "Qianyang paper-cutting is divided according to different craftsmanship", when predicting the "Qianyang paper-cutting" entity label, the "B-PaperCutting" and "I-PaperCutting" labels should be continuous to ensure the integrity of the predicted labels. When performing BIO tag recognition for traditional skill entities, the CRF layer is used to generate the globally optimal sequence label. The specific label layer structure is as follows: Figure 4 shown.

[0097] The traditional skills entity scores output by the sequence modeling layer are used as the input of the label decoding layer. CRF comprehensively marks the feature representation of the fusion entity boundary features and global information to mark the traditional skills entities of the intangible cultural heritage sequence. The probability score of the text entity label position is:

[0098]

[0099] Among them, y is the label sequence, H is the input sequence, i is the index of each word in the text sequence, and A is the transfer score of the transfer matrix, which is used to represent y i to y i+1 The movement probability of the label.

[0100] According to the scores of the traditional skill label sequences, this embodiment uses the Viterbi algorithm to find the label sequence with the highest score, thereby improving the accuracy of the traditional skill named entity recognition. The conditional probability of the characters in the text sequence can be expressed as:

[0101]

[0102] In the process of identifying traditional skill entity labels, data loss is inevitable. The pre-trained text representation layer generates an embedded representation of the contextual information of the skill entity. In order to further capture the long dependencies in the sequence text, the feature vector of the attention layer is passed to the BiLSTM structure, so that the sequence modeling layer can calculate the start and end losses for each skill category label. These key steps may have some factors that cause the model to fail to accurately capture all relevant skill entity information. Different from the traditional loss calculation method, the model calculates the negative log-likelihood loss of each real entity at the corresponding position in the Logits sequence according to the starting position of the real skill label in iterative training, and ensures that the start position label is before the end label so that ER-BFAS can accurately learn the boundary features of the entity.

[0103] Assume that there are N technical entity categories in a given text sequence. For any label label∈[1,N], the text sequence feature vector T is obtained by BERT representation. out , the LSTM unit further captures the previous and next dependencies to obtain the overall hidden vector h lstm Therefore, for each traditional craft label c i ∈label, the logits of specific positions are calculated through linear transformation, that is, the fully connected layer, that is,

[0104]

[0105] Among them, W c start and W c end is the weight matrix, b is the bias term, s label 、e label Indicates the loss value at the starting and ending positions.

[0106] In addition, in order to only consider the valid positions of the intangible cultural heritage text sequence, the attention mask m is used to filter the loss calculation of invalid positions in the text sequence to ensure that only the actual skill entity information of the text is optimized during the iteration process, where y label Record the real start and end labels of the label, [m] means only retaining characters with a mask of 1, so the loss value loss for each category can be expressed as:

[0107]

[0108] Finally, all skill label losses are added together to get the total loss total_loss to ensure that the data loss is at the lowest level. This model ER-BFAS combines entity boundary features with the attention sequence structure to generate more accurate entity label features through feature fusion. Its core is to use cross entropy correlation loss as the optimization target, continuously iterate and optimize model performance, thereby improving the model's recognition accuracy and efficiency for specific traditional skill entities.

[0109]

[0110] 3) Traditional skill entity recognition is performed based on the trained traditional skill entity recognition model based on entity boundary features and attention sequence structure (ER-BFAS model).

[0111] Comparative experiment

[0112] This experiment selected the text corpus related to the representative traditional skills list of the Intangible Cultural Heritage Network as the data source of the annotation sequence. According to the actual use and common types of traditional handicrafts, 11 traditional skills label categories including performance skills, ceramic skills, textile skills and paper-cutting skills were specified. After data preprocessing, more than 14,127 traditional skills intangible cultural heritage entities were identified through BIO annotation. In order to further integrate the boundary attribute characteristics of the entity, the starting position of the traditional skills entity was introduced as the boundary feature of data analysis. In addition, the data set was divided according to the ratio of 7:1:2 to construct the experimental data set. The specific division statistical results are shown in Table 4.

[0113] Table 4

[0114]

[0115] Among them, X represents a specific traditional skill category, B represents the starting position of the identification entity, and I represents the middle or end position of the identification entity.

[0116] This experiment is based on pre-trained models, bidirectional long short-term memory neural networks, and conditional random fields, and is compared and verified with other benchmark models. In terms of experimental parameters, several key parameters are set to ensure the effectiveness and stability of model entity recognition. The maximum length of the input text is limited to 512, the number of fixed samples in the batch is 12, and the number of data categories is determined by the number of labels in the data set. During the model training process, the basic learning rate is 3e-5, the learning rate for the sequence modeling layer is set to 3e-3, and AdamW is used to update the weights to prevent overfitting. In order to ensure that the model training is more stable in the early stage, the warmup_proportion is set to 0.01 to ensure that the learning rate increases according to the predefined strategy. Other environmental parameters are referenced.

[0117] Table 5.

[0118] Table 5

[0119]

[0120] In order to intuitively compare the performance of models in identifying different traditional art entities, this experiment selected recall, precision and F1 score to evaluate different models.

[0121] When analyzing the entity prediction performance of the model, accuracy is an important indicator specifically used to quantify the proportion of accurate positive examples when identifying traditional skills entities. Its numerical value directly reflects the accuracy of the model. The higher the value, the better the effect of entity prediction and recognition. When calculating the specific value, the confusion matrix (as shown in Table 7) can be used to count the number of true positives, true negatives, false positives, and false negatives, so as to understand the overall accuracy of the model in identifying traditional skills.

[0122] Table 6

[0123]

[0124]

[0125] The recall rate is mainly used to evaluate the model's ability to identify entity labels related to traditional skills, and can count the correct entities correctly identified. Since the identified entity type covers multiple entity labels, it is necessary to combine the annotated data set to compare the model's predicted labels with the traditional skill entities in the data set, so as to calculate the proportion of the number of entities correctly identified by the model to the positive examples. A higher value means that the model can comprehensively identify relevant intangible cultural heritage entities from a large number of intangible cultural heritage text sequences to ensure the integrity of the data information.

[0126]

[0127] Combined with the actual application of traditional skill entity recognition, the F1 score, as the harmonic mean of precision and recall, can comprehensively measure the balance between the accuracy and recall of the model in identifying positive entity samples. By comprehensively considering the model's prediction of positive samples, it is ensured that the model can fully and accurately identify the relevant traditional skill entities, especially the non-legacy labels of skills with unbalanced categories, providing a reference for the subsequent evaluation of the overall situation of each model.

[0128]

[0129] In order to compare the effectiveness and recognition ability of this model, models that have been used in entity recognition research in recent years are selected as references. The benchmark models include WL-SIE, Trans-NER, CNN, TFBERT, TextCNN, BiLSTM, and Multi_LSTM.

[0130] (1) Trans-NER (Wang Yinrui, Peng Dunlu, Chen Zhang, et al. Trans-NER: A Chinese named entity recognition model supported by transfer learning [J]. Small and Micro Computer Systems, 2019, 40(08): 1622-1626.): Wang Yinrui et al. proposed an entity recognition method that supports transfer algorithms to reduce the need for tedious data labeling work. However, the training parameters of this model are inefficient and require a lot of computing resources.

[0131] (2) Trans-HMM (Li Jian, Xiong Qi, Hu Yating, et al. Chinese named entity recognition method based on Transformer and hidden Markov model [J]. Journal of Jilin University (Engineering Edition), 2023, 53(05): 1427-1434.): Li Jian et al. introduced a position encoding function to represent the relative position information features between characters, and used the hidden Markov model to generate relevant entity soft labels for text sequences, thereby improving the strong label representation ability of the model. However, it may not be interchangeable in capturing semantic features, which affects the recognition of relative position information of text sequences.

[0132] (3) SciBERT-BiLSTM (Pan Xuelian, Qian Yufei, Wang Xianyu. Research on software entity recognition based on SciBERT-BiLSTM-CRF-wordMixup [J / OL]. Modern Intelligence: 1-18 [2024-07-01].): Pan Xuelian et al. conducted in-depth research on relevant entities in the software field. They defined relevant entities, used entity annotation software to build a corpus, and used wordMixup for data enhancement to make up for the shortcomings of SciBERTNer's entity recognition ability. However, its generalization ability and overfitting problems still need to be solved.

[0133] (4) MHA-CRF (Wu Shuai, Yang Xiuzhang, He Lin, et al. Research on fine-grained ancient book entity recognition based on syntactic features and Bert-BiLSTM-MHA-CRF [J / OL]. Data Analysis and Knowledge Discovery: 1-16 [2024-07-01].): Wu Shuai et al. combined the complex sentence structure characteristics of ancient book texts, used high-frequency relation words as key feature words for recognition, and constructed a fact-based feature extraction paradigm; they used the MHA-CRF structure to perform deep fine-grained entity recognition on the test text. However, it is only applicable to a specific ancient text sample set in terms of syntactic features.

[0134] (5) BERT-OnLSTM (Li Chunnan, Wang Lei, Sun Yuanyuan, et al. Named Entity Recognition Method for Legal Documents on Theft Crimes Based on BERT [J]. Journal of Chinese Information Science, 2021, 35(08): 73-81.): Li Chunnan et al. constructed a text entity corpus in the professional legal field to solve the problem of polysemy. They combined BERT-OnLSTM-CRF to further capture relevant entity text features, but there are certain limitations in semantic extraction.

[0135] (6) WL-SIE (Wang Haipeng, Du Fang, Song Lijuan, et al. Chinese medical named entity recognition integrating word-level paragraph information [J / OL]. Computer Technology and Development: 1-8 [2024-07-01].): Wang Haipeng et al. integrated word-level information with paragraph information and proposed a new word-level paragraph entity recognition method for the Chinese named entity task, but it cannot handle long and complex texts well, and its accuracy is easily affected.

[0136] Results and Discussion

[0137] According to the boundary characteristics of traditional skill entities, the structure of attention mechanism and sequence modeling is integrated, and the pre-trained model is used to represent the input text sequence as a feature vector, so as to obtain the deep semantic information of the text. In order to further improve the model's ability to extract the front and back dependencies of traditional skill entity labels, a bidirectional LSTM structure is added to bidirectionally traverse the context information of entity labels. The model divides the training set into multiple batches according to the pre-set batch size, and continuously accepts data to generate predicted labels in each batch iterative training. After that, the model statistically calculates the loss value between the predicted and true label values, and performs back propagation to update the internal parameters to optimize the model's entity recognition ability. Therefore, under the condition that the pre-set training parameters remain unchanged, appropriately adjust the parameters such as the batch number, the maximum length of the text limit, and the learning rate to observe the optimal parameters suitable for the traditional skill entity recognition model.

[0138] In order to deeply analyze the relationship between different training parameters of the ER-BFAS model and entity label prediction, this paper attempts to adjust multiple parameter combinations to find the optimal parameter combination suitable for capturing traditional art entities and achieve the best entity extraction effect. Figure 6 It can be seen that this experiment attempts to analyze the impact of different parameter combinations on the overall extraction ability of the model, with the characteristics of focusing on the key parameters of batch size, maximum text limit length, and learning rate. At the same time, the curve changes are shown in the form of batch_size=12,max_seq_len=150,lr=1e-4 and batch_size=32,max_seq_len=512,lr=3e -5Two sets of parameter combinations and the approximate training status of the model in the other five evaluation indicators.

[0139] By comparing the evaluation results of the chart, the following conclusions can be drawn:

[0140] When the maximum text length is 512, the batch size is 32, and the learning rate is 3e -5 When , the training results of ER-BFAS after adjusting the explanation are better than those before the adjustment, which verifies that different learning rates and processing text lengths will affect the model's recognition of traditional skills entities in intangible cultural heritage text sequences. When epochs exceed 7, the average loss value of the model tends to 0, and the change trends of the prediction accuracy and recall rate are basically consistent, indicating that the traditional skills recognition model constructed in this study fully utilizes the label prediction ability of sequence modeling to deeply extract the intangible cultural heritage text corpus in the experimental data.

[0141] As the training cycle, text limit length and batch size continue to increase, the average loss of the model, prediction accuracy, recall rate and F1 score change trends are different. Among them, sub-graph (a) shows the change trend of the average loss of each node before adjusting the parameters, but compared with another parameter combination, it can be clearly found that the loss between the prediction result and the true label of the latter is smaller, and there is no large fluctuation. For the overall change of precision, recall rate and F1 score, the change degree of the latter three curves is basically consistent, and when the batch size is small, the model needs to undergo multiple iterations of training to effectively recognize traditional art entities in text sequences.

[0142] Ablation experiment

[0143] In order to further study the effectiveness of each module of this model on the task of named entity recognition of traditional skills, this study conducted a series of ablation experiments on five modules based on the annotated traditional skills intangible cultural heritage corpus, including the sequence modeling part, the attention layer that integrates entity boundary information, and other parts. To further verify the influence of each part of this model on the intangible cultural heritage classification prediction task, this paper conducted a series of ablation experiments on each module based on the constructed intangible cultural heritage dataset, including the attention module and the semantic extraction part. The specific experimental results can be seen in

[0144] Table 7.

[0145] contrast

[0146] Table 7 shows the evaluation results of multiple model structures. Micro-average, macro-average and weighted-average are used to conduct an in-depth analysis of the experimental results. The current optimal results are indicated in bold. From the perspective of entity recognition, the BiLSTM, CRF and BiLSTM-CRF models have similar recognition effects on Label Avg and Micro avg, but poor performance on Macro avg and Weighted avg, indicating that the single sequence labeling structure has high data requirements during model training and performs poorly for small category entities. The values ​​of Micro avg and Weighted avg of BERT-CRF and ER-BFAS are relatively close, but the former is slightly lower than the ER-BFAS module. From the perspective of category average and macro-average values, the ER-BFAS model that integrates entity boundary features is more balanced in processing traditional skill entities of various categories, even when processing small sample test corpora.

[0147] Table 7

[0148]

[0149] In order to intuitively compare the effects of each part on traditional technical entity recognition from the evaluation indicators of precision, recall rate and F1 value, the length of the horizontal line is used to represent the specific value of each indicator, thus drawing Figure 6 . Comparing multiple subgraphs of (a)-(g), it is found that BiLSTM and CRF perform better in terms of precision and recall, but the predicted value of Macro avg is lower, indicating that a single sequence structure may have limitations when dealing with fewer entity categories. In addition, the values ​​on the right side of the image of BERT-CRF and ER-BFAS are relatively close, but the latter's precise accuracy in Micro avg and Weighted avg both reach 0.88, which is 0.06 and 0.05 higher than BERT-CRF, indicating that the use of pre-trained attention mechanism on a large amount of intangible cultural heritage text corpus can capture richer semantic information. However, in terms of extracting entity boundary features, it is second only to the ER-BFAS model.

[0150] Therefore, the BERT-CRF based on pre-training and label dependency of a large number of intangible cultural heritage corpora has achieved an accuracy of 0.84 and 0.63 or above, indicating that the pre-trained attention mechanism can capture rich semantic information and generate text representation vectors; the introduction of CRF effectively handles the dependency relationship of traditional craft labels, thereby improving the accuracy of entity recognition. On this basis, by integrating entity boundary attribute features and bidirectional LSTM structure, the macro-average values ​​of Precision and Recall of ER-BFAS have increased by 0.08 and 0.15 respectively, enhancing the model's ability to handle complex traditional craft entities, so as to more accurately identify entity boundary information and handle entity overlap.

[0151] Overall performance comparison

[0152] The entity recognition effects between different benchmark models and this model are compared and analyzed. In order to reduce the error caused by the training iteration process, the hyperparameter combination of the benchmark model is adjusted with reference to the basic parameter mechanism. In order to reduce the error caused by parameter adjustment, the hyperparameters of the benchmark model are adjusted with reference to the training parameters of this model. At the same time, this experiment uses 5 evaluation indicators to evaluate the prediction results of the traditional skills test corpus. Among them, Avg time and Avg loss respectively represent the average training time and training loss value of each node during the model training process. The lower loss value indicates that the model has a higher fitting effect. Precision, Recall and F1score are replaced by the weighted average of the indicator, which is used to describe the basic performance of different models in identifying different traditional skills labels on the test corpus.

[0153] Table 8

[0154]

[0155]

[0156] Through

[0157] Through in-depth analysis of the data in Table 8, it can be observed that different models show different differences in traditional art entity recognition. Compared with other models, SciBERT-BiLSTM has the lowest precision, which may be due to the high number of false positives in the entity labels, while Trans-HMM, MHA-CRF, Trans-Ner and BERT-OnLSTM have outstanding performance in precision and recall, especially Trans-HMM and Trans-Ner, which have values ​​above 96% when identifying traditional art entity labels, showing high accuracy. However, the results of these four prediction models are not ideal in terms of training loss and time efficiency, especially Trans-HMM and Trans-Ner, which consume more computing resources in inferring the potential entity labels of traditional art corpus. In contrast, this model performs best in terms of time cost and average loss, while the prediction accuracy and recall are slightly lower than those of several other models. Therefore, it is shown that the sequence and training model that integrates entity boundary attribute features has better named entity recognition effect in the entity recognition task of traditional art.

[0158] From the perspective of resource loss and prediction accuracy alone, it may not be possible to show the model's extraction effect on various traditional art entities. Therefore, we choose precision, recall and F1 score to evaluate the recognition of different entity labels. Figure 8The training status of each model on the test corpus. Among the six sub-graphs (a)-(f), it can be clearly found that the ER-BFAS, SciBERT-BiLSTM and Trans-HMM models have better entity recognition effects, especially the accuracy of each label of ER-BFAS is basically maintained above 0.7, indicating that this model can effectively identify the relevant entity information characteristics of traditional skills from the intangible cultural heritage corpus. In contrast, SciBERT-BiLSTM in sub-graph (b) is more sensitive to data requirements. When there is an imbalance in the number distribution of the data set, the accuracy of a certain skill type will be low. On the contrary, although Trans-HMM has a simple structure and relatively low evaluation results for each label, it can well identify the continuity and sequence features that may appear in the intangible cultural heritage corpus for the traditional skill entity label. In addition, the accuracy of Trans-NER based on the attention mechanism on the training set is second only to the first three, but its F1 score and recall rate need to be improved. Therefore, the model structure of BiLSTM based on attention mechanism or bidirectional traversal can better adapt to the data distribution and feature recognition in different traditional technology fields.

[0159] At the same time, three different averaging methods, micro avg, macro avg and weightedavg, are used to calculate precision, recall and F1 scores respectively, and their values ​​are expressed as "T1", "T2" and "T3". The overall comparison can be seen in Figure 8 . From the above conclusions, it can be concluded that ER-BFAS and SciBERT-BiLSTM have better entity recognition effects at present. Here, we add the model comparison of WL-SIE under different optimization algorithms. Comparing the average results of precision, recall and F1 score, the micro-average, macro-average and weighted average of ER-BFAS and SciBERTNer are relatively close overall, but the recall rate of the former is slightly higher than that of the latter, indicating that the ER-BFAS proposed in this paper not only uses the BiLSTM structure to perform sequence modeling of traditional skill labels, but also combines the boundary information and attribute characteristics of the entity to enhance the entity recognition ability, providing a comprehensive and accurate optimization model for the study of traditional skill entity recognition. In contrast, although the WL-SIE model can recognize some traditional skill entities, its recall rate and F1 score are extremely low, and it may still need to adjust parameters or optimize the model structure multiple times.

[0160] In summary, through the comparison of the training results of multiple models, it can be found that BERT can better capture the front-end dependency of text through pre-training of large-scale text data, but the structure has high requirements for computing resources and cannot process long text sequences alone; BiLSTM solves the limitations of traditional LSTM unidirectional memory units and improves the classification prediction ability of the model for text sequences, but the training time is long and the local information of the data cannot be extracted; and CNN learns the expression ability of intangible cultural heritage text representation by stacking multiple convolutional layers, and improves the representation ability of the model layer by layer. Compared with recurrent neural networks and long short-term memory neural networks, it is impossible to directly analyze the time relationship in the text sequence, and has certain limitations in the classification prediction task of the intangible cultural heritage list. Therefore, the present invention relies on the pre-trained BERT model, combined with bidirectional LSTM and one-dimensional convolutional layers, so as to make up for the problems of insufficient front-end dependency and local feature extraction of text sequences, and constructs the ER-BFAS classification prediction model. Through multiple experimental comparisons and verifications, compared with a single deep neural network model, this model shows better performance in the classification prediction task of intangible cultural heritage text resources.

[0161] In summary, by comparing multiple experiments, it can be found that Trans-NER, which is mainly based on the attention structure of the transfer algorithm, performs well in the expected accuracy of the test, but it is still insufficient for the traditional skill entity data; BERT-CRF obtains the deep semantic information of the text through pre-training, which helps to improve the accuracy of CRF sequence annotation; SciBERT-BiLSTM has a good pre-trained semantic representation effect, but it is more sensitive to the type distribution of experimental data; Trans-HMM structure is relatively simple, but it may not be able to extract the non-legacy entity information of traditional skills well. Therefore, compared with the sequence entity annotation model, the present invention enhances the model's processing ability for complex skill entities by integrating entity boundary attribute features, thereby using the pre-training attention mechanism to comprehensively obtain the vector representation information of the non-legacy text, and introduces the BiLSTM and CRF layers to add the front and back constraints of the skill class label, quickly capture the contextual information of the non-legacy sequence, so as to further improve the accuracy of traditional non-legacy entity recognition.

[0162] Embodiment 2

[0163] The traditional art entity recognition system of the present invention comprises:

[0164] An acquisition module is used to acquire the text to be recognized;

[0165] The recognition module is used to input the text to be recognized into the trained traditional art entity recognition model based on entity boundary features and attention sequence structure to recognize the traditional art entities in the text to be recognized.

[0166] In this embodiment, the traditional skills entity recognition model based on entity boundary features and attention sequence structure includes a text representation layer, a sequence modeling layer and a label decoding layer, wherein the text representation layer determines the BIO type according to the word position attribute, and uses the attention mechanism to perform text vector representation of traditional skills on the entity character information; the sequence modeling layer dynamically captures the context information and dependency relationships of the intangible cultural heritage entities through the BiLSTM unit, and obtains the possible entity labels and probability distributions of each character in the text sequence; the label decoding layer accepts the probability distribution of sequence modeling, and uses the potential association between the character position and the entity category label to predict the traditional skills entities at each position in the intangible cultural heritage text sequence.

[0167] The division of modules in the embodiments of the present application is schematic and is only a logical function division. There may be other division methods in actual implementation. In addition, each functional module in each embodiment of the present application may be integrated into a processor, or may exist physically separately, or two or more modules may be integrated into one module. The above-mentioned integrated modules may be implemented in the form of hardware or in the form of software functional modules.

[0168] Embodiment 3

[0169] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the traditional art entity recognition method are implemented. The memory may include a memory, such as a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk memory, etc. The processor, the network interface, and the memory are interconnected through an internal bus. The internal bus may be an industrial standard architecture bus, a peripheral component interconnection standard bus, an extended industrial standard structure bus, etc. The bus may be divided into an address bus, a data bus, a control bus, etc. The memory is used to store programs. Specifically, the program may include a program code, and the program code includes computer operation instructions. The memory may include a memory and a non-volatile memory, and provide instructions and data to the processor.

[0170] Embodiment 4

[0171] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the traditional art entity recognition method are implemented. Specifically, the computer-readable storage medium includes, but is not limited to, volatile memory and / or non-volatile memory. The volatile memory may include random access memory (RAM) and / or cache memory, etc. The non-volatile memory may include read-only memory (ROM), hard disk, flash memory, optical disk, magnetic disk, etc.

[0172] 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 in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0173] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the 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 the 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.

[0174] 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 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0175] 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.

[0176] 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 traditional art entity recognition method, characterized in that: include: Get the text to be recognized; The text to be recognized is input into a trained traditional art entity recognition model based on entity boundary features and attention sequence structure to recognize the traditional art entities in the text to be recognized.

2. The traditional art entity recognition method according to claim 1, characterized in that: Before inputting the text to be recognized into the trained traditional entity recognition model based on entity boundary features and attention sequence structure, the following steps are also included: Construct a traditional skills dataset; Construct a traditional art entity recognition model based on entity boundary features and attention sequence structure; The traditional skills data set is used to train a traditional skills entity recognition model based on entity boundary features and attention sequence structure, thereby obtaining a trained traditional skills entity recognition model based on entity boundary features and attention sequence structure.

3. The traditional art entity recognition method according to claim 2, characterized in that: The process of constructing the traditional skills dataset is as follows: Collect relevant data from national and local intangible cultural heritage websites of traditional skills through crawler technology; Preprocess the collected relevant corpus to obtain several pieces of text data; The traditional skills entities are annotated for each text data, and then a traditional skills dataset is constructed based on the annotated text data.

4. The traditional art entity recognition method according to claim 2, characterized in that: The traditional skills entity recognition model based on entity boundary features and attention sequence structure includes a text representation layer, a sequence modeling layer and a label decoding layer. The text representation layer determines the BIO type according to the word position attribute, and uses the attention mechanism to perform text vector representation of traditional skills for the entity character information; the sequence modeling layer dynamically captures the context information and dependency relationships of the intangible cultural heritage entities through the BiLSTM unit, and obtains the possible entity labels and probability distributions for each character in the text sequence; the label decoding layer accepts the probability distribution of sequence modeling, and uses the potential association between character positions and entity category labels to predict the traditional skills entities at each position in the intangible cultural heritage text sequence.

5. The traditional art entity recognition method according to claim 4, characterized in that: The text representation layer includes word embedding test, multi-layer feedforward network and multi-head attention mechanism.

6. The traditional art entity recognition method according to claim 2, characterized in that: The loss function in the process of training the traditional skill entity recognition model based on entity boundary features and attention sequence structure using the traditional skill dataset is:

7. A traditional art entity recognition system, characterized in that: include: An acquisition module is used to acquire the text to be recognized; The recognition module is used to input the text to be recognized into the trained traditional art entity recognition model based on entity boundary features and attention sequence structure to recognize the traditional art entities in the text to be recognized.

8. The traditional art entity recognition system according to claim 7, characterized in that: The traditional skills entity recognition model based on entity boundary features and attention sequence structure includes a text representation layer, a sequence modeling layer and a label decoding layer. The text representation layer determines the BIO type according to the word position attribute, and uses the attention mechanism to perform text vector representation of traditional skills for the entity character information; the sequence modeling layer dynamically captures the context information and dependency relationships of the intangible cultural heritage entities through the BiLSTM unit, and obtains the possible entity labels and probability distributions for each character in the text sequence; the label decoding layer accepts the probability distribution of sequence modeling, and uses the potential association between character positions and entity category labels to predict the traditional skills entities at each position in the intangible cultural heritage text sequence.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the traditional art entity recognition method as described in any one of claims 1-6 are implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method for identifying a traditional art entity as described in any one of claims 1 to 6 are implemented.

Images