Method for extracting equipment maintenance support text entity nesting based on multi-technology fusion

Abstract

This invention relates to a method for extracting entity nesting from equipment maintenance and support text based on multi-technology fusion, belonging to the field of big data. The method includes: acquiring equipment maintenance and support text and performing word segmentation to obtain multiple sequence fragments; setting scanning windows of different lengths and scanning and filtering the sequence fragments according to configured filtering rules to obtain candidate sequences; inputting the candidate sequences into a BERT pre-trained model to convert them into feature vector sequences; using a BiLSTM neural network to extract features from the feature vector sequences and obtain sequence context information; using a TextCNN neural network to process the feature vector sequences and obtain local sequence features; concatenating the features to obtain fused features; mapping the fused features to label scores through a fully connected layer; calculating the label probability of a sequence fragment being an entity based on the label scores to predict the entity label corresponding to the sequence fragment. This invention's method can improve the efficiency and quality of equipment maintenance and support work by integrating multiple methods and strategies.

Description

Technical Field

[0001] This invention relates to the field of big data technology, and in particular to a method for extracting nested text entities for equipment maintenance and support based on multi-technology integration. Background Technology

[0002] Currently, equipment maintenance and support operational data is scattered across different equipment management departments and equipment-using units. Due to the relatively traditional and outdated data management methods used by the military, this data is typically recorded in text form, directly leading to severe fragmentation of equipment maintenance and support operational data, particularly with significant entity nesting within these texts. For example, a document might state: "The Army Equipment Department has subordinate units such as the Supply Support Bureau, Maintenance Support Bureau, and Aviation Equipment Bureau." Given the complexity and high level of specialization in equipment maintenance and support operations, extracting entities from these nested texts is extremely challenging and has become a key issue hindering the construction and development of the equipment maintenance and support data resource system. Summary of the Invention

[0003] The purpose of this invention is to at least address one of the aforementioned technical deficiencies.

[0004] Therefore, the purpose of this invention is to propose a method for extracting nested text entities of equipment maintenance and support based on multi-technology integration, which can improve the efficiency and quality of equipment maintenance and support work by integrating multiple methods and strategies.

[0005] To achieve the above objectives, embodiments of the present invention provide a method for extracting nested text entities of equipment maintenance and support based on multi-technology fusion, including:

[0006] S1, Obtain the equipment maintenance and support text, and use a word segmentation tool to segment the equipment maintenance and support text to obtain multiple sequence fragments; S2, set scanning windows of different lengths, and scan and filter sequence segments according to the configured filtering rules to obtain candidate sequences; S3, the candidate sequence is input into the Bert pre-trained model and converted into a feature vector sequence. The BiLSTM neural network is used to extract features from the feature vector sequence to obtain sequence context information. The TextCNN neural network is used to process the feature vector sequence to obtain local sequence features. S4. The sequence context information and the sequence local features are concatenated to obtain fused features. The fused features are mapped to label scores through a fully connected layer. The label probability of a sequence segment being an entity is calculated based on the label scores to predict the entity label corresponding to the sequence segment.

[0007] Furthermore, the word segmentation tool has a built-in dictionary for the equipment maintenance and support domain. Using this tool, the equipment maintenance and support text X is segmented to obtain a segmentation sequence C={c1,c2,c3,...cj,...cm}. Where X = {x1, x2, x3, ... xi, ... xn}, where xi is the i-th character in the text statement, n is the text length, and the number of sequence segments of X is n(n+1) / 2; cj is the j-th word in the word segmentation sequence, m is the number of words in the word segmentation sequence, and the number of sequence segments C is m(m+1) / 2, where m n.

[0008] Furthermore, the filtering rules are set based on word part of speech, grammatical rules, and semantic similarity.

[0009] Furthermore, in S2, when the length of the scanning window is 1, the filtering rule is set to directly filter non-noun words; when the length of the scanning window is greater than 1, sequences that meet the preset conditions, such as suffix nouns, adverb beginnings, and endings, are directly filtered.

[0010] Further, in S3, the candidate sequence is input into the BERT pre-trained model and converted into a feature vector sequence. for:

[0011] in, , Let N be the i-th character in the output sequence, and N be the length of the input sequence. , For BERT feature dimensions.

[0012] Furthermore, in S3, the feature vector output by the Bert pre-trained model is used... The input to the BiLSTM neural network is obtained by concatenating the outputs of the forward LSTM and the backward LSTM in the BiLSTM neural network. The output of the feature vector sequence after passing through the BiLSTM neural network is ,in, This is the hidden state sequence output by the BiLSTM, which incorporates contextual information from the input text sequence. Let be the hidden state corresponding to the t-th position in the sequence.

[0013] Furthermore, in S3, the feature vector output by the Bert pre-trained model is used... This refers to the input to the TextCNN neural network. Represented as:

[0014] The computation process of the TextCNN neural network is as follows:

[0015]

[0016]

[0017] in, W is the weight matrix of the convolution kernel. h is the kernel height; b is the bias; Ci is the value obtained from the i-th convolution; C is the feature vector obtained through convolution. ; Max pooling of C yields , q represents the number of convolution kernels.

[0018] Furthermore, in S4, the output of the BiLSTM neural network is... and the output of the TextCNN neural network By splicing the data, the fused features are obtained. , to fuse Combined features are mapped to label scores through a fully connected layer:

[0019] Where M represents the number of actual entity labels.

[0020] Furthermore, in S4, a softmax classifier is used to compute sequence fragments. Predicted label probability of an entity :

[0021] Furthermore, in S4, the true label y of the sequence segment is one-hot encoded, and the training objective function is defined as the cross-entropy loss between the predicted label probability and the true label. The calculation formula is as follows:

[0022] in, express The corresponding real labels are 1 for positive classes and 0 for negative classes.

[0023] Compared with the prior art, the advantages and beneficial effects of the present invention are as follows: This invention proposes a method for extracting nested entities from equipment maintenance and support text based on multi-technology fusion. This method uses a neural network model to extract nested entities from equipment maintenance and support text. First, it segments the text information using a domain dictionary to obtain word sequences. Second, it uses a sliding window to scan and sample word sequences to obtain sequence fragments. Then, it filters candidate entity sequences by limiting the size of the scanning window and establishing filtering rules, thereby reducing the number of negative samples and improving model recognition efficiency. Next, the classification model uses a BERT pre-trained model as the text representation layer, and extracts text feature vectors through a fusion of BiLSTM and TextCNN. Finally, it uses softmax to decode the sequence entity labels.

[0024] This invention verifies the superiority of the model in fault corpora by comparing it with commonly used baseline models, and demonstrates through ablation experiments that the model using dual neural networks outperforms the single-neural model. Compared with existing technologies, the equipment maintenance and support text entity nesting extraction method based on multi-technology fusion proposed in this invention has the following significant advantages and technical effects: (1) Sample optimization is achieved in the preprocessing stage, which significantly improves the model recognition efficiency. In the text preprocessing stage, this invention first segments the equipment maintenance and support text information into word sequences by combining a domain dictionary, and then obtains sequence fragments by scanning and sampling through a sliding window. At the same time, by limiting the size of the grass temple window and formulating filtering rules to screen out candidate entity sequences, the number of negative samples is greatly reduced, effectively reducing the computational overhead of the model training and inference stages and significantly improving the model recognition efficiency.

[0025] (2) Feature extraction method of multi-model fusion improves entity recognition performance This invention employs a BERT pre-trained model as the text representation layer, extracts text feature vectors through a fusion of BiLSTM and TextCNN neural networks, and then uses softmax to decode the sequence entity labels. Comparison with commonly used baseline models verifies the superiority of this invention's model in fault prediction. Furthermore, ablation experiments demonstrate that the model using dual neural network fusion significantly outperforms the single neural network model, substantially improving the accuracy and robustness of nested entity extraction.

[0026] (3) Supporting the standardization and value-added aspects of equipment maintenance and support text information This invention provides a novel text-embedded entity extraction technique to effectively support the transformation and loading of textual information in equipment maintenance and support into standardized and regulated value information, support the construction of a big data resource system for equipment maintenance and support, and realize the transformation of equipment maintenance and support from data advantages and information advantages to application advantages and decision-making advantages. It provides a high-quality data foundation for subsequent applications such as intelligent decision-making, knowledge graph construction, and fault analysis in equipment maintenance and support.

[0027] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0028] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which: Figure 1 A flowchart of a method for extracting nested text entities for equipment maintenance and support based on multi-technology fusion according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the overall process of extracting nested entities from equipment maintenance and support text based on a neural network according to an embodiment of the present invention; Figure 3 This is a schematic diagram of a neural network-based sequence fragment classification model according to an embodiment of the present invention. Detailed Implementation

[0029] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0030] This invention proposes a method for extracting nested entities from equipment maintenance and support text based on multi-technology fusion. It involves equipment maintenance and support big data ETL (Extract, Transform, Load) technology, as well as a technology for extracting nested entities from equipment maintenance and support text based on multi-method fusion.

[0031] like Figure 1 and Figure 2 As shown, the method for extracting nested text entities of equipment maintenance and support based on multi-technology fusion in this embodiment of the invention includes: S1. Obtain the equipment maintenance and support text, and use a word segmentation tool to segment the equipment maintenance and support text to obtain multiple sequence fragments.

[0032] According to the equipment maintenance and support log, the equipment log content is as follows: X={x1,x2,x3,...xi,...xn}(1) Where xi represents the i-th character in the text statement, n represents the text length, and the number of sequence segments of X is n(n+1) / 2.

[0033] Directly identifying equipment maintenance and support entities from all sequence fragments would increase the training difficulty of the classification model and reduce classification efficiency. To overcome this problem, this invention designs a method that uses words as the basic unit to label the sampled sequences.

[0034] Specifically, first, a dictionary for the equipment maintenance and support domain is determined, and then this dictionary is imported into the word segmentation tool. That is, the word segmentation tool has a built-in dictionary for the equipment maintenance and support domain. The Elasticsearch word segmentation tool of the equipment maintenance and support big data management and analysis platform is used to segment the text content X, resulting in the following word segmentation sequence: C={c1,c2,c3,...cj,...cm}(2) Where cj represents the j-th word in the segmented sequence, and m represents the number of words in the segmented sequence. By using words as the basic unit, the granularity of the equipment maintenance and support text is increased, and the collection density is reduced. After segmentation, the number of sequence fragments of C is m(m+1) / 2, and m must be less than n (i.e., m n).

[0035] refer to Figure 2 For example, after segmenting the phrase "The Army Equipment Department has subordinate units such as the Supply Support Bureau, Maintenance Support Bureau, and Aviation Equipment Bureau," the resulting sequence fragment is "Army / Equipment Department / subordinate units / Supply Support Bureau / Maintenance Support Bureau / Aviation Equipment Bureau / etc." The use of labeled sampled sequences significantly reduces the number of sequence fragments, thereby reducing the model's classification pressure and the number of negative samples, and providing input for subsequent sequence fragment scanning and filtering.

[0036] S2, set scanning windows of different lengths, and scan and filter sequence segments according to the configured filtering rules to obtain candidate sequences.

[0037] Specifically, since the sequence fragments obtained from sampled sequence labeling still contain a large number of negative samples, classifying all sequence fragments would not only increase the model's computation time but also cause overfitting. Therefore, this invention obtains candidate sequences by reasonably setting the scanning window size (i.e., the number of words m in the segmented sequence) and then scanning and filtering the sequence fragments according to the configured filtering rules, which serve as the input to the natural language training model.

[0038] As shown in Table 1, the filtering rules are set based on attributes such as word part-of-speech, grammatical rules, and semantic similarity in the sequence fragments, as detailed below: 1. When the length of the scanning window is 1, the filtering rule is set to directly filter non-noun words. That is, if the sequence fragment is a single word, the probability of a non-noun word being an equipment maintenance and support entity is 0, so it can be directly filtered out.

[0039] 2. When the length of the scanning window is greater than 1, sequences beginning with or ending with suffix nouns or adverbs that meet the preset conditions are directly filtered. That is, if a sequence fragment contains words of multiple parts of speech, it needs to be further processed by filtering rules. Although the filtering rules here cannot completely cover all non-entity sequence fragments, they can still filter out most meaningless sequence fragments.

[0040] Table 1 Filtering Rules

[0041] S3. The candidate sequence is input into the BERT pre-trained model and converted into a feature vector sequence. The BiLSTM neural network is used to extract features from the feature vector sequence to obtain sequence context information. The TextCNN neural network is used to process the feature vector sequence to obtain local sequence features.

[0042] Specifically, the BERT pre-trained model can transform input text sequences into feature vector sequences, possessing strong semantic representation capabilities. This invention employs the Next Sentence Prediction (NSP) task to implement the BERT pre-training process. The NSP task is primarily used to determine whether two input texts are related as consecutive sentences. Typically, a classification character [CLS] is inserted before the two sentences, and a separator character [SEP] is inserted between them. The semantic vector C contains the semantic information of the entire input sample and can be used to determine whether there is a contextual relationship between the two sentences. Since the input to BERT in this paper is a single text sequence, the separator character [SEP] is not needed during model training, and the semantic vector output by BERT is selected as the input to the subsequent model.

[0043] refer to Figure 3 Viewing BERT as a non-linear mapping function from character sequences to semantic vectors, candidate sequences are input into the BERT pre-trained model and transformed into feature vector sequences. In other words, Bert's output can be represented as: (3) in, , Let N be the i-th character in the output sequence, and N be the length of the input sequence. , This represents the output feature vector. This represents the BERT feature dimension.

[0044] After obtaining the above feature vector sequence, the feature vector of the sequence segment is extracted by calculating semantic vectors using BiLSTM and TextCNN respectively and then fusing them.

[0045] 1. BiLSTM Neural Network LSTM is widely used in various NLP tasks, and it has significant advantages in sequence feature modeling, effectively capturing long dependencies in sequence data. An LSTM unit controls sequence information through input gates, output gates, and forget gates. Its computation process is as follows:

[0046]

[0047] (4)

[0048]

[0049]

[0050] in, This represents the forget gate, input gate, and output gate. Represents the transition vector; Indicates the state of the cell at time t; represents the t-th character in the text statement, and is the feature vector input to the LSTM unit at time t; This represents the hidden state of the LSTM at time t, which is the feature information output by the cell at time t and is used to pass it to the next time step. The weight matrix of the forget gate is used to... and Perform a linear transformation to control the information filtering of the forget gate; The weight matrix represents the input gate, which controls the absorption of information by the input gate; The weight matrix represents the output gate, which controls the output ratio of the cell state.

[0051] A weight matrix representing the candidate cell state is used to generate candidate information at time t; This represents the bias term of the forget gate, used to adjust the activation threshold of the forget gate; This represents the bias term of the input gate, which adjusts the activation threshold of the input gate; This represents the bias term of the output gate, which adjusts the activation threshold of the output gate; Bias terms representing candidate cell states, adjusting the generation of candidate states; and It is a non-linear activation function.

[0052] This invention employs BiLSTM to extract features from the input vector sequence. BiLSTM consists of a forward LSTM and a backward LSTM, which can better capture sequence context information. The feature vector output by the BERT pre-trained model is used as an example. The input to the BiLSTM neural network is the forward LSTM output of the BiLSTM neural network. and inverse LSTM output spliced ​​together The output of the feature vector sequence after passing through the BiLSTM neural network is: (5) in, This is the hidden state sequence output by the BiLSTM, which incorporates contextual information from the input text sequence. Let be the hidden state corresponding to the t-th position in the sequence.

[0053] That is, the output of the BERT layer is used as the input of the BiLSTM, and the output of the forward LSTM is used as the input of the forward LSTM. and inverse LSTM output spliced ​​together The output of the feature vector sequence after passing through BiLSTM is .

[0054] 2. TextCNN Neural Network TextCNN has a simple structure, can quickly acquire local features of sequences, and performs excellently in feature modeling of short texts. The feature vector output by the BERT pre-trained model is used as an example. This refers to the input to the TextCNN neural network. Represented as: (6) in, This indicates vector concatenation.

[0055] The computation process of the TextCNN neural network is as follows: (7) (8) (9) (10) Where W is the weight matrix of the convolution kernel, h is the kernel height; b is the bias; Ci is the value obtained from the i-th convolution; C is the feature vector obtained through convolution. Max pooling is performed on C to obtain , q represents the number of convolution kernels.

[0056] S4. The sequence context information and the sequence local features are concatenated to obtain the fused features. The fused features are then mapped to label scores through a fully connected layer. The label probability of a sequence segment being an entity is calculated based on the label scores to predict the entity label corresponding to the sequence segment.

[0057] Specifically, the output of the BiLSTM neural network and the output of the TextCNN neural network By splicing the data, the fused features are obtained. , will integrate features After being mapped to label scores by a fully connected layer: (11) Where M represents the number of actual entity labels.

[0058] To predict the label of a sequence fragment, a softmax classifier is used to compute the sequence fragment's label. Predicted label probability of an entity : (12) To measure the difference between predicted and true labels during model training, the true label y of the sequence segment is one-hot encoded. The training objective function is defined as the cross-entropy loss between the predicted label probability and the true label, and the calculation formula is as follows: (13) in, express The corresponding real labels are 1 for positive classes and 0 for negative classes.

[0059] The filtered sequence fragments are input into a sequence classification model to predict their corresponding entity labels, thus realizing the automated processing of nested entities in text information.

[0060] In summary, the implementation process of this invention is as follows: First, a domain dictionary is determined. Then, the dictionary is imported into a word segmentation tool to segment the text records of equipment maintenance and support. The sampling granularity of the text sequence is increased by using words as the basic unit. Sequence fragments are obtained through scanning windows of different lengths, and then filtered according to filtering rules to reduce the number of negative samples. Finally, the filtered sequence fragments are input into a sequence classification model to predict their corresponding entity labels, thereby completely solving the problem of automated processing of nested text information entities during the ETL process of equipment maintenance and support data.

[0061] The following example, using a document recording of equipment maintenance and support for a certain military branch, illustrates the implementation process of the method of this invention.

[0062] Step 1: Preparation of Experimental Data and Evaluation Indicators An experiment was conducted using the equipment maintenance and support text records of a certain military branch as the subject. 850 text records related to the organizational structure of equipment maintenance and support were manually compiled, such as: "The location of the Equipment Maintenance Department of the XX Group Army of the Eastern Theater Command of the Army is XX City, XX Province, China". The text content was segmented and sequence-labeled using the Elasticsearch word segmentation tool of the Equipment Maintenance and Support Big Data Management and Analysis Platform. The results are shown in Table 2.

[0063] Table 2. Schematic diagram of the organizational structure for equipment maintenance and support.

[0064] The labeled data is divided into training set, validation set and test set in a ratio of 7:1:2, and the amount of entity data in each dataset is shown in Table 3.

[0065] Table 3 shows the number of institutional entities in the dataset.

[0066] In summary, precision, recall, and F1 score are used as evaluation metrics. Precision represents the percentage of correctly identified entities out of all identified entities, recall represents the percentage of correctly identified entities out of all identified entities, and the F1 score is a combined representation of precision and recall. The overall precision and recall of the model are the average of the precision and recall for different entity types. The calculation formula is as follows:

[0067] (14)

[0068] Where M represents the number of entity categories; This represents the number of entities in the i-th class that are correctly classified in the identification results. This represents the number of entities of class i that were misidentified in the classification results. Let be the number of entities of type i that are misclassified as other entities.

[0069] Step Two: Parameter Analysis and Selection The experimental hyperparameters were set using the intrinsic parameters of the BERT model: embedding dimension of 768, character embedding dimension of 128, BiLSTM hidden layer units of 256, input batch size of 16, learning rate of 1×10⁻⁵, and training epochs of 50. For TextCNN, three different convolutional kernels were set: 2×768, 3×768, and 4×768, with each kernel having 256 units. The maximum scan window size was set to 6. Specific parameter settings are shown in Table 4.

[0070] Table 4 Experimental Parameter Settings

[0071] Step 3: Comparative Experiment Neural network models such as TextCNN, RCNN, DPCNN, BiLSTM, and Transformer were used as baseline models for experimental comparison to fully verify the performance of the model of the present invention. The experimental results are shown in Table 5.

[0072] Table 5 Comparison of experimental results

[0073] As shown in Table 5, using the F1 score as the evaluation criterion for the test set, TextCNN performed the worst in the test set with an F1 score of only 79.50%, while BiLSTM performed better with an F1 score of 90.34%. The F1 scores of DPCNN and RCNN were similar. However, the overall results show that the model proposed in this invention has the highest F1 score, and the recall and precision of the model are also relatively high.

[0074] Step 4 Ablation Experiment The model based on neural network extraction of nested entities in equipment maintenance and support text is composed of multiple algorithms and models. Each component has a significant impact on the performance of the model. Experiments were conducted to investigate the impact of each model component using the ablation method. The ablation experiment results are shown in Table 6.

[0075] Table 6 Ablation Experiment Results

[0076] As shown in Table 6, the model performed worst after removing the BERT embedding layer, with an F1 score decrease of 4.78% and a recall decrease of 7.69%. However, it still performed better than the single-channel model, demonstrating that the dual-channel feature fusion model can extract features from equipment maintenance and support text better. Removing both BiLSTM and TextCNN simultaneously, and using only the BERT embedding layer to classify equipment maintenance and support entities, significantly impacted model performance, with an F1 score decrease of 4.65% and a recall decrease of 7.52%. Removing TextCNN also reduced model performance, with an F1 score decrease of 2.52%. Further removal of BiLSTM only reduced the F1 score by 1.04%, indicating that TextCNN had a greater impact on model performance than BiLSTM.

[0077] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0078] It will be readily understood by those skilled in the art that this invention includes any combination of the inventive description and specific embodiments outlined in the foregoing specification, as well as the various parts shown in the accompanying drawings. Due to space limitations and for the sake of brevity, not all of these combinations have been described in detail. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

[0079] Although embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present invention without departing from the principles and spirit of the invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for extracting nested entities from equipment maintenance and support text based on multi-technology fusion, characterized in that, include: S1, Obtain the equipment maintenance and support text, and use a word segmentation tool to segment the equipment maintenance and support text to obtain multiple sequence fragments; S2, set scanning windows of different lengths, and scan and filter sequence segments according to the configured filtering rules to obtain candidate sequences; S3, the candidate sequence is input into the Bert pre-trained model and converted into a feature vector sequence. The BiLSTM neural network is used to extract features from the feature vector sequence to obtain sequence context information. The TextCNN neural network is used to process the feature vector sequence to obtain local sequence features. S4. The sequence context information and the sequence local features are concatenated to obtain fused features. The fused features are mapped to label scores through a fully connected layer. The label probability of a sequence segment being an entity is calculated based on the label scores to predict the entity label corresponding to the sequence segment.

2. The method for extracting nested text entities for equipment maintenance and support based on multi-technology fusion according to claim 1, characterized in that, S1, the word segmentation tool has a built-in dictionary for the equipment maintenance and support field. The word segmentation tool is used to segment the equipment maintenance and support text X, resulting in a word segmentation sequence C={c1,c2,c3,...cj,...cm}. Where X = {x1, x2, x3, ... xi, ... xn}, where xi is the i-th character in the text statement, n is the text length, and the number of sequence segments of X is n(n+1) / 2; cj is the j-th word in the word segmentation sequence, m is the number of words in the word segmentation sequence, and the number of sequence segments C is m(m+1) / 2, where m n.

3. The method for extracting nested text entities for equipment maintenance and support based on multi-technology fusion according to claim 1, characterized in that, The filtering rules are set based on word part of speech, grammatical rules, and semantic similarity.

4. The method for extracting nested text entities for equipment maintenance and support based on multi-technology fusion according to claim 1, characterized in that, In S2, when the length of the scanning window is 1, the filtering rule is set to directly filter non-noun words; when the length of the scanning window is greater than 1, sequences that meet the preset conditions, such as suffix nouns, adverb beginnings and endings, are directly filtered.

5. The method for extracting nested text entities for equipment maintenance and support based on multi-technology fusion according to claim 1, characterized in that, In S3, the candidate sequence is input into the BERT pre-trained model and converted into a feature vector sequence. for: in, , Let N be the i-th character in the output sequence, and N be the length of the input sequence. , For BERT feature dimensions.

6. The method for extracting nested text entities for equipment maintenance and support based on multi-technology fusion according to claim 1, characterized in that, In S3, the feature vector output by the Bert pre-trained model is used. The input to the BiLSTM neural network is obtained by concatenating the outputs of the forward LSTM and the backward LSTM in the BiLSTM neural network. The output of the feature vector sequence after passing through the BiLSTM neural network is ,in, This is the hidden state sequence output by the BiLSTM, which incorporates contextual information from the input text sequence. Let be the hidden state corresponding to the t-th position in the sequence.

7. The method for extracting nested text entities for equipment maintenance and support based on multi-technology fusion according to claim 1, characterized in that, In S3, the feature vector output by the Bert pre-trained model is used. This refers to the input to the TextCNN neural network. Represented as: The computation process of the TextCNN neural network is as follows: in, W is the weight matrix of the convolution kernel. h is the kernel height; b is the bias; Ci is the value obtained from the i-th convolution; C is the feature vector obtained through convolution. ; Max pooling of C yields , q represents the number of convolution kernels.

8. The method for extracting nested text entities for equipment maintenance and support based on multi-technology fusion according to claim 1, characterized in that, In S4, the output of the BiLSTM neural network and the output of the TextCNN neural network By splicing the data, the fused features are obtained. The fused features are mapped to label scores through a fully connected layer: Where M represents the number of actual entity labels.

9. The method for extracting nested text entities for equipment maintenance and support based on multi-technology fusion according to claim 1, characterized in that, In S4, a softmax classifier is used to calculate sequence fragments. Predicted label probability of an entity for: 。 10. The method for extracting nested text entities for equipment maintenance and support based on multi-technology fusion according to claim 1, characterized in that, In S4, the true label y of the sequence segment is one-hot encoded, and the training objective function is defined as the cross-entropy loss between the predicted label probability and the true label. The calculation formula is as follows: in, express The corresponding real labels are 1 for positive classes and 0 for negative classes.

Images