Knowledge extraction method and device in field of cigarette cut tobacco processing

Abstract

The invention discloses a knowledge extraction method and device in the field of cigarette cut tobacco processing. The method comprises the steps that a to-be-extracted text in the field of cigarette cut tobacco processing is input into an entity extraction model; performing dependency syntactic analysis on the to-be-extracted text based on an entity extraction result; forming a triple structure based on a dependency analysis result; wherein the step of training the entity extraction model comprises the following steps of: dividing an unstructured knowledge text from data in the field of cigarette cut tobacco processing; constructing an ontology network structure according to cigarette cut tobacco data characteristics; marking the unstructured knowledge text based on the ontology network structure to form a data set; wherein the label of each entity comprises a start-stop position and an entity type; and training the entity extraction model based on the data set. According to the method, the ontology network structure is constructed according to the cigarette cut tobacco data characteristics, the entity extraction model is assisted to clearly divide the entity boundary, and the entity extraction accuracy can be improved.

Description

Technical Field

[0001] The present application relates to the field of knowledge extraction technology, and more specifically, to a knowledge extraction method and device in the field of cigarette shred processing. Background Art

[0002] Knowledge extraction is one of the key technologies required to build a knowledge graph. The main task of knowledge extraction is to extract structured information from unstructured text data, so as to obtain a triple structure representing the language relationship between entities and construct a knowledge graph.

[0003] The knowledge extraction of cigarette silk processing technology and equipment involves a wide range of manufacturers and production lines. The knowledge text of the processing technology is relatively complex. During the knowledge extraction process, the entity boundaries are blurred, so the accuracy of entity relationship extraction is low. Summary of the invention

[0004] The present application provides a knowledge extraction method and device in the field of cigarette silk processing, which constructs an ontology network structure based on the characteristics of cigarette silk processing data, assists the entity extraction model to clearly divide the entity boundaries, and helps to improve the accuracy of entity extraction. At the same time, knowledge extraction is combined with dependency syntactic analysis to improve the overall accuracy of knowledge extraction.

[0005] The present application provides a knowledge extraction method in the field of cigarette shred processing, including:

[0006] Input the text to be extracted in the field of cigarette shred processing into the entity extraction model to obtain the entity extraction result;

[0007] Perform dependency syntactic analysis on the extracted text based on the entity extraction results;

[0008] Form a triple structure based on the dependency analysis results;

[0009] Among them, the entity extraction model is trained, specifically including:

[0010] Separate unstructured knowledge text from the data in the field of cigarette thread processing;

[0011] Construct the ontology network structure according to the characteristics of cigarette thread making data;

[0012] Based on the ontology network structure, unstructured knowledge texts are annotated to form a data set; the annotation of each entity includes the start and end positions and entity type;

[0013] Train the entity extraction model based on the dataset.

[0014] Preferably, the entity extraction model includes a vector representation layer, a gating structure layer and a conditional random field model layer.

[0015] Preferably, before training the entity extraction model, the data set is fused with an external knowledge graph and / or multimodal cigarette making data.

[0016] Preferably, when performing dependency syntactic analysis on the text to be extracted based on the entities in the text to be extracted, the entity extraction results are imported into a natural language processing tool in the form of a dictionary for dependency syntactic analysis to achieve a word segmentation function for professional vocabulary.

[0017] Preferably, the triple structure is formed based on the dependency analysis result, specifically including:

[0018] Based on the dependency analysis result, each component in the extracted text is divided into upper and lower levels, and the upper and lower level division result includes a first entity, a second entity that is a hypernym of the first entity, and a dependency relationship between the first entity and the second entity;

[0019] The first entity and the second entity are regarded as the head entity and the tail entity respectively, and the dependency relationship is regarded as the relationship between the head entity and the tail entity.

[0020] The present application also provides a knowledge extraction device in the field of cigarette shred processing, including an entity extraction module, a dependency analysis module, a triple formation module and a training module;

[0021] The entity extraction module is used to input the text to be extracted in the field of cigarette shred processing into the entity extraction model to obtain the entity extraction result;

[0022] The dependency analysis module is used to perform dependency syntactic analysis on the extracted text based on the entity extraction results;

[0023] The triple formation module is used to form a triple structure based on the dependency analysis results;

[0024] The training module is used to train the entity extraction model, including a text segmentation module, an ontology network construction module, an annotation module, and an iteration module;

[0025] The text segmentation module is used to segment unstructured knowledge text from the data in the field of cigarette thread processing;

[0026] The ontology network construction module is used to construct the ontology network structure according to the characteristics of cigarette thread making data;

[0027] The annotation module is used to annotate unstructured knowledge text based on the ontology network structure to form a data set; the annotation of each entity includes the start and end positions and the entity type;

[0028] The iteration module is used to iteratively train the entity extraction model based on the dataset.

[0029] Preferably, the entity extraction model includes a vector representation layer, a gating structure layer and a conditional random field model layer.

[0030] Preferably, the training module also includes a fusion module, which is used to fuse the data set with an external knowledge graph and / or multimodal cigarette making data.

[0031] Preferably, the dependency analysis module is used to import the entity extraction results in the form of a dictionary into a natural language processing tool for dependency syntactic analysis, thereby realizing the word segmentation function of professional field vocabulary.

[0032] Preferably, the triple formation module includes a hyponym division module and an entity relationship determination module;

[0033] The upper-lower-lower level division module is used to divide each component in the extracted text into upper-lower level divisions based on the dependency analysis result, and the upper-lower-higher level division result includes a first entity, a second entity that is a hypernym of the first entity, and a dependency relationship between the first entity and the second entity;

[0034] The entity relationship determination module is used to take the first entity and the second entity as the head entity and the tail entity respectively, and the dependency relationship as the relationship between the head entity and the tail entity.

[0035] Other features and advantages of the present application will become apparent from the following detailed description of exemplary embodiments of the present application with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The accompanying drawings, which are incorporated in and constitute a part of the specification, illustrate embodiments of the application and, together with the description, serve to explain the principles of the application.

[0037] Figure 1 A flow chart of the knowledge extraction method in the field of cigarette shred processing provided for this application;

[0038] Figure 2 The marking result of “actual value of moisture content of inlet material” provided for this application;

[0039] Figure 3 The network structure diagram of the entity extraction model provided for this application;

[0040] Figure 4 An example of dependency analysis results provided for this application;

[0041] Figure 5 This is a structural diagram of the knowledge extraction device in the field of cigarette shred processing provided in this application. DETAILED DESCRIPTION

[0042] Various exemplary embodiments of the present application will now be described in detail with reference to the accompanying drawings. It should be noted that unless otherwise specifically stated, the relative arrangement of components and steps, numerical expressions and numerical values ​​set forth in these embodiments do not limit the scope of the present application.

[0043] The following description of at least one exemplary embodiment is merely illustrative in nature and is in no way intended to limit the present application, its application, or uses.

[0044] Technologies, methods, and equipment known to ordinary technicians in the relevant art may not be discussed in detail, but where appropriate, the technologies, methods, and equipment should be considered part of the specification.

[0045] In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not limiting. Therefore, other examples of the exemplary embodiments may have different values.

[0046] The present application provides a knowledge extraction method and device in the field of cigarette silk processing, which constructs an ontology network structure based on the characteristics of cigarette silk processing data, assists the entity extraction model to clearly divide the entity boundaries, and helps to improve the accuracy of entity extraction. At the same time, knowledge extraction is combined with dependency syntactic analysis to improve the overall accuracy of knowledge extraction.

[0047] like Figure 1 As shown, the knowledge extraction method in the field of cigarette shred processing provided by this application includes:

[0048] S110: Input the text to be extracted in the field of tobacco shred processing into the entity extraction model to obtain the entity extraction result.

[0049] The entity extraction model is used to extract entities in sentences from the text to be extracted.

[0050] As an embodiment, training the entity extraction model specifically includes:

[0051] P1: Separate unstructured knowledge text from data in the field of cigarette wire processing.

[0052] As an example, unstructured knowledge text is divided from the data in the field of cigarette shred processing, specifically including:

[0053] P11: Collect source data in the field of cigarette cutting and processing.

[0054] Specifically, data (such as collected processing technology, processing equipment, processing mode, brand specifications, etc.) sorted out by a preset number of collection objects (such as a preset number of cigarette companies) on the production line is collected as source data.

[0055] P12: Preprocess the source data and build the basic data set.

[0056] Preprocessing can include data cleaning, processing error values, missing values ​​and outliers in the data, using machine learning to smooth and reduce noise in the data, converting the data into a unified format and unit, adding corresponding labels, categories and annotations to different types of data, and integrating the data into the data text of the corresponding category for comprehensive analysis.

[0057] After data preprocessing, the data is aggregated to form a data pool to build a basic data set. The aggregated data includes source data, preprocessed data, fused data, and data after relationship analysis.

[0058] P13: Divide the basic data set to obtain unstructured knowledge text.

[0059] Specifically, according to the data characteristics of various modes, the multimodal data in the basic data set is divided into structured data, image data and unstructured data. The entity extraction model mainly extracts unstructured text knowledge.

[0060] P2: Construct an ontology network structure based on the characteristics of cigarette making data.

[0061] The ontology of silk-making technology and equipment provides a structured way to define and organize concepts and their relationships within the field. This structured knowledge helps the knowledge extraction system to more accurately identify entities in the text and correctly understand the relationships between them. By constructing the silk-making and equipment domain ontology, it helps the system understand the deep meaning of the text. For example, the ontology can associate "silk cutting machine" with "silk cutting equipment", enabling the system to identify and integrate related information, making this information appear in different terms or expressions.

[0062] As an embodiment, when constructing the ontology network structure, first determine the main category concepts in the silk-making process and equipment ontology based on the characteristics of cigarette silk-making data, clarify the domain and scope of the ontology, and then construct the ontology network structure based on the relationship between the main category concepts.

[0063] In this application, as an embodiment, Protégé software is used to design and implement classes and class construction levels. The construction of the silk-making process and equipment ontology can be carried out through the graphical page of the software to design classes and inter-class levels, and the constructed model can be converted into various data formats such as OWL, RDF, etc. for storage.

[0064] In Protégé software, there are concepts of super- and hyponyms, i.e., parent and child classes. There is a top-level class "OWL: Thing", and all ontologies are subclasses of this class, such as "process flow" is a subclass of "processing mode", and "process parameters" is a subclass of "processing procedures". Unstructured text is annotated with this hierarchical structural concept as a constraint.

[0065] As an example, based on the characteristics of cigarette silk making, OWL: Thing is selected as the top-level class, and different manufacturing companies are distinguished through the "Brand" class. The "Leaf Group Formula" class determines the specific products of different brands. The "Production Implementation Standard" class sets requirements for the production process from three aspects: silk making, rolling and packing, and filter rods. The subclasses of "Silk Making Requirements" include "Processing Mode". From the experience given by experts in the field, it is concluded that there need to be corresponding silk making requirements for different processing modes. For different processing modes, the process flow and guarantee parameters are also different. Multiple major categories are set under the "Process Flow" class, and molecular categories are continued on this basis. On this basis, according to the concepts of ontology and semantic relationships between ontologies, the construction of the ontology network structure is completed in Protégé.

[0066] On this basis, "actual value of moisture content of inlet material", "outlet material temperature" and "exhaust air volume" meet the constraints defined in the process parameter class, while "thin plate drying wire" meets the definition of the processing procedure. Therefore, when judging its class, "thin plate drying wire" belongs to the parent class entity of "exhaust air volume".

[0067] P3: Annotate unstructured knowledge text based on the ontology network structure to form a data set.

[0068] Specifically, unstructured knowledge texts are annotated according to ontology constraints.

[0069] As an embodiment, the Label Studio text annotation tool is used, and the BMEO annotation strategy is adopted, that is, the first character of the target entity is annotated with B, the middle characters are all M, the ending character is E, and the non-target entity character is O. The entity category follows the letter, such as B-processing technology. Specifically, the required entity relationship type is first set in the Label Studio tool, and then the entity in the text is selected. Finally, the annotated text is converted into a bems file, and the annotation of each entity consists of two parts: the start and end positions and the entity type. Figure 2 The marking result of “actual value of moisture content of inlet material” is shown.

[0070] According to the constructed ontology, various types of entities in the complex cigarette making text are strictly labeled so that the entities have clear boundaries. This not only helps the deep learning network model to accurately judge the entity relationships hidden in unstructured knowledge, but also makes the knowledge graph constructed based on this more clear and organized, which is helpful for the subsequent association mining and analysis between processes through vectorization of the knowledge graph.

[0071] As an embodiment, the labeled data set is divided into a training set, a test set, and a validation set in a ratio of 7:2:1.

[0072] It should be noted that during the training process, when using the entity extraction model for prediction, the entity will be considered a successful prediction only when both the type and the start and end boundaries of the entity are successfully predicted.

[0073] P4: Train the entity extraction model based on the dataset.

[0074] As an embodiment, the entity extraction model includes a vector representation layer, a gating structure layer, and a conditional random field model layer.

[0075] Preferably, the entity extraction model adopts the ERNIE-BiGRU-CRF model.

[0076] The vector representation layer uses the ERNIE (Enhanced Representation through Knowledge Integration) model, which is a pre-trained language model that performs well in natural language processing tasks. The basic principle of the ERNIE model includes knowledge enhancement, that is, by introducing external knowledge (such as entity information in the knowledge graph or multimodal data, such as images, structured data, etc.), the model can use more background information when processing language tasks. In the case of existing relevant knowledge of cigarette making, it can help the model better identify professional terms and entities and enhance the model's ability to understand external knowledge. The ERNIE model uses hierarchical pre-training and uses multi-level pre-training tasks to improve the model's ability to understand text.

[0077] The gated structure layer uses the BiGRU (bidirectional gated neural network) layer. BiGRU is an important variant of the recursive neural network RNN. BiGRU uses two GRU layers to process the forward and backward information of the sequence respectively. This structure enables it to capture richer contextual information when processing sequences, and is particularly suitable for tasks such as natural language processing, time series prediction, and speech recognition. In terms of unit structure, GRU is a simplified version of LSTM. Each LSTM unit has three gated structures (input gate, forget gate, output gate), while GRU has only two gated structures (update gate, reset gate). Due to its simple structure, fewer parameters, and higher computational efficiency, the GRU unit can more efficiently solve the problem of gradient disappearance or explosion in the RNN network structure. For BiGRU, the specific formulas involved are as follows:

[0078] r t =σ(W r ·[h t-1 , x t ])

[0079] z t =σ(W z ·[h t-1 , x t ])

[0080]

[0081]

[0082]

[0083] In the formula, x is input from the current position t and the output h of the previous hidden layer t-1 After adding the linear changes, it is connected to the sigmoid activation function σ, W r and W z Represent the weights of the reset gate and the update gate respectively. Due to the existence of σ, its output value is between 0 and 1, which is used to select how much information to keep. The calculation process of the update gate is similar. is the updated value, which is determined by the weight W of the new candidate state h , reset gate t 、Previous position output h t-1 , input x at this position t Together, the final output of the forward GRU at the current position is By the updated value Update gate z t , enter h in the previous position t-1 The calculation process of the reverse GRU is the same as the forward one, and the final output h of BiGRU is tUsually the forward hidden state and the reverse hidden state splicing.

[0084] In the task of entity recognition, the label of a character is often closely related to the features of the characters before and after it. Therefore, the use of a bidirectional gated neural network can well learn the text features of the context, thereby improving the model's accuracy in dividing entity boundaries.

[0085] The conditional random field model (CRF) is a logarithmic linear model for sequence labeling problems. Its function is to constrain the labels by capturing the dependencies between the context. As a conditional probability distribution model, the core principle formula for the linear chain conditional random field is as follows:

[0086]

[0087] P(y|x) represents the probability of labeling sequence y given the observation sequence x. Where x is the input variable, which represents the labeled observation sequence; y is the output sequence, which represents the label sequence corresponding to x; f k represents the characteristic function; ω k Represents the weight corresponding to the feature function. The training set is estimated through maximum likelihood to obtain the conditional probability model, and then the Viterbi algorithm is used to output the label sequence y with the maximum conditional probability according to the given observation sequence.

[0088] After the feature vector is transformed in dimension by the fully connected layer, the CRF layer is responsible for decoding the predicted label sequence and realizing the conversion from feature to probability. It not only considers the emission scores of a single label, but also combines the transfer scores between labels to correct these emission scores. This enables CRF to capture the structural correlations and constraints between labels. For example, in some contexts, one entity category may be more likely to follow another category. By ensuring the structural rationality of the label sequence, CRF can significantly improve the performance of the model.

[0089] When the entity extraction model is trained based on the data set, the evaluation criteria for the prediction results output by the model are: when both the start and end boundaries and the entity type are successfully predicted, it is considered a successful prediction.

[0090] The common evaluation indicators of entity labeling tasks include precision P, recall R and F1 value to comprehensively evaluate the experimental results. The calculation formula is as follows:

[0091]

[0092]

[0093]

[0094] Preferably, before training the entity extraction model, the dataset is fused with an external knowledge graph and / or multimodal cigarette shred data. During the training process, ERNIE can process text and knowledge graph data simultaneously, and improve the overall performance of the model through a joint learning method. This method can combine the contextual information of the text with the entity information in the knowledge graph to improve the model's understanding and reasoning capabilities.

[0095] As an example, the entity extraction model includes an ERNIE vector representation layer, two GRU layers, and a CRF layer. Its network structure is as follows: Figure 3 shown. Figure 3 The prediction results when the input is "drum drying method" are also shown.

[0096] The F1 value, accuracy, and recall rate obtained by the entity extraction model are shown in the following table. From the table, it can be seen that the training accuracy F1 value of the model reaches 0.929, the accuracy reaches 0.939, and the recall rate reaches 0.920, which are significantly higher than the traditional neural network model. The results show that the ERNIE-BiGRU-CRF combination model provided by this application has a very good entity extraction effect on the cigarette shred processing technology.

[0097] Model Accuracy Recall F1 value ERNIE-BiGRU-CRF 0.939 0.920 0.929

[0098] S120: Perform dependency syntactic analysis on the text to be extracted based on the entity extraction result.

[0099] When performing dependency syntactic analysis on the text to be extracted based on the entities in the text to be extracted, the entity extraction results are imported into the natural language processing tool in the form of a dictionary for dependency syntactic analysis to achieve the word segmentation function of professional vocabulary.

[0100] Figure 4 The dependency analysis result of "the material flow rate should be set reasonably, not exceed the process manufacturing capacity, and remain continuous and stable" is given.

[0101] S130: A triple structure is formed based on the dependency analysis result, and then the triple structure is imported into the graph database to complete the knowledge storage work.

[0102] As an embodiment, forming a triple structure based on the dependency analysis result specifically includes:

[0103] Q1: Based on the dependency analysis results, each component in the extracted text is divided into upper and lower levels. The upper and lower level division results include a first entity, a second entity that is a hypernym of the first entity, and a dependency relationship between the first entity and the second entity.

[0104] According to the rules of dependency syntactic analysis, any word has a dependency relationship with the root word directly or indirectly. The root points to the core predicate of the sentence, which is usually the first predicate of the sentence. The dependency relationship is HED, and the dependency relationship of the entire sentence is constructed by expanding outward from the core predicate. The results of the dependency analysis of the sentence show that there is a hierarchical influence relationship between entities.

[0105] exist Figure 4 In the example shown, the components in the sentence can be divided into upper and lower levels based on the dependency analysis results. The tool numbers each word and punctuation mark in the sentence from front to back. The first number in the dependency relationship represents the sequence number of the hypernym that the word depends on, and the following letters represent the dependency relationship between the two words. The results are as follows:

[0106] Material flow: Dependencies: (3, 'SBV');

[0107] Should: Dependencies: (3, 'ADV');

[0108] Rational: Dependencies: (3, 'ADV');

[0109] Set: Dependencies: (0, 'HED');

[0110] No more than: Dependencies: (3, 'COO');

[0111] Process manufacturing capability: Dependency: (5, 'VOB');

[0112] AND:Dependencies:(8,'ADV');

[0113] Keep: Dependencies: (3, 'COO');

[0114] Continuous Stable: Dependency: (8, 'VOB');

[0115] The serial number of the core predicate is 3, SBV represents the subject-predicate relationship, ADV represents the adverbial-verb relationship, COO represents the parallel relationship, and VOB represents the verb-object relationship.

[0116] Taking "Material Flow: Dependency: (3, 'SBV')" as an example, the first entity is the material flow, the second entity is the entity "Setting" with sequence number 3, and the dependency is SBV.

[0117] Q2: The first entity and the second entity are regarded as the head entity and the tail entity respectively, and the dependency relationship is regarded as the relationship between the head entity and the tail entity.

[0118] For example, usually the noun that plays the role of subject in a sentence is the head entity of the triple structure, the object component of the sentence is the tail entity of the triple structure, and the predicate component connecting the subject and the object is generally directed from the head entity to the tail entity as the relationship of the triple structure.

[0119] exist Figure 4 In the example, a triple structure can be constructed according to the modifying and modified relationship of each component in the sentence. For example, the triple structure of material flow - not exceeding - process manufacturing capacity, material flow - maintain - stable, can also be the structure of material flow - reasonable - set, such as a name phrase plus an adjective phrase.

[0120] The prerequisite for performing dependency syntactic analysis on a sentence is to segment the sentence into words. However, the word segmentation tool of the tool itself is difficult to directly segment entity-level vocabulary such as "process manufacturing capability". Therefore, a deep learning network (the above-mentioned entity extraction model) is used to extract the entity and then import it into LTP to assist in word segmentation.

[0121] In this way, structured knowledge is extracted from unstructured text, and the constructed triple structure is imported into the Neo4j graph database through the Cypher language to complete the knowledge storage.

[0122] Based on the above, the present application also provides a knowledge extraction device in the field of cigarette shred processing. Figure 5 As shown, the knowledge extraction device includes an entity extraction module 510 , a dependency analysis module 520 , a triple formation module 530 and a training module 540 .

[0123] The entity extraction module 510 is used to input the text to be extracted in the field of cigarette shred processing into the entity extraction model to obtain the entity extraction result.

[0124] The dependency analysis module 520 is used to perform dependency syntactic analysis on the text to be extracted based on the entity extraction result.

[0125] The triple forming module 530 is used to form a triple structure based on the dependency analysis result.

[0126] The training module 540 is used to train the entity extraction model, and includes a text segmentation module 5401 , an ontology network construction module 5402 , an annotation module 5403 and an iteration module 5404 .

[0127] The text segmentation module 5401 is used to segment unstructured knowledge text from data in the field of cigarette thread processing.

[0128] The ontology network construction module 5402 is used to construct an ontology network structure according to the characteristics of cigarette thread making data.

[0129] The annotation module 5403 is used to annotate the unstructured knowledge text based on the ontology network structure to form a data set, wherein the annotation of each entity includes the start and end positions and the entity type.

[0130] The iteration module 5404 is used to iteratively train the entity extraction model based on the data set.

[0131] Preferably, the entity extraction model includes a vector representation layer, a gating structure layer and a conditional random field model layer.

[0132] Preferably, the training module 540 further includes a fusion module 5405, which is used to fuse the data set with an external knowledge graph and / or multimodal cigarette making data.

[0133] Preferably, the dependency analysis module 520 is used to import the entity extraction results in the form of a dictionary into a natural language processing tool for dependency syntactic analysis, thereby realizing the word segmentation function of professional domain vocabulary.

[0134] Preferably, the triple forming module 530 includes a hyponym division module 5301 and an entity relationship determination module 5302 .

[0135] The hyponym division module 5301 is used to perform hyponym division on each component in the extracted text based on the dependency analysis result. The hyponym division result includes a first entity, a second entity that is a hypernym of the first entity, and a dependency relationship between the first entity and the second entity.

[0136] The entity relationship determination module 5302 is used to take the first entity and the second entity as the head entity and the tail entity respectively, and the dependency relationship as the relationship between the head entity and the tail entity.

[0137] This application has a strong ability to extract data in the field of cigarette silk processing, and can provide a data basis and technical reference for the construction of a knowledge graph of cigarette silk processing technology.

[0138] Although some specific embodiments of the present application have been described in detail by way of example, it should be understood by those skilled in the art that the above examples are only for illustration, not for limiting the scope of the present application. It should be understood by those skilled in the art that the above embodiments may be modified without departing from the scope and spirit of the present application. The scope of the present application is defined by the appended claims.

Claims

1. A knowledge extraction method in the field of cigarette shred processing, characterized in that: include: Input the text to be extracted in the field of cigarette shred processing into the entity extraction model to obtain the entity extraction result; Performing dependency syntactic analysis on the text to be extracted based on the entity extraction result; Form a triple structure based on the dependency analysis results; The entity extraction model is trained, specifically including: Separate unstructured knowledge text from the data in the field of cigarette thread processing; Construct the ontology network structure according to the characteristics of cigarette thread making data; Annotating the unstructured knowledge text based on the ontology network structure to form a data set; wherein the annotation of each entity includes the start and end positions and the entity type; The entity extraction model is trained based on the data set.

2. The knowledge extraction method in the field of cigarette shred processing according to claim 1 is characterized in that: The entity extraction model includes a vector representation layer, a gated structure layer and a conditional random field model layer.

3. The knowledge extraction method in the field of cigarette shred processing according to claim 1 is characterized in that: Before training the entity extraction model, the data set is fused with an external knowledge graph and / or multimodal cigarette making data.

4. The knowledge extraction method in the field of cigarette shred processing according to claim 1 is characterized in that: When performing dependency syntactic analysis on the text to be extracted based on the entities in the text to be extracted, the entity extraction result is imported into a natural language processing tool in the form of a dictionary to perform dependency syntactic analysis, thereby realizing the word segmentation function of professional field vocabulary.

5. The knowledge extraction method in the field of cigarette shred processing according to claim 1 is characterized in that: The triple structure is formed based on the dependency analysis results, including: Based on the dependency analysis result, each component in the to-be-extracted text is divided into a hyponym and a hyponym, wherein the hyponym and hyponym division result includes a first entity, a second entity that is a hypernym of the first entity, and a dependency relationship between the first entity and the second entity; The first entity and the second entity are respectively used as a head entity and a tail entity, and the dependency relationship is used as a relationship between the head entity and the tail entity.

6. A knowledge extraction device in the field of cigarette shred processing, characterized in that: It includes entity extraction module, dependency analysis module, triple formation module and training module; The entity extraction module is used to input the text to be extracted in the field of tobacco shred processing into the entity extraction model to obtain the entity extraction result; The dependency analysis module is used to perform dependency syntactic analysis on the text to be extracted based on the entity extraction result; The triple formation module is used to form a triple structure based on the dependency analysis result; The training module is used to train the entity extraction model, including a text segmentation module, an ontology network construction module, a labeling module and an iteration module; The text segmentation module is used to segment unstructured knowledge text from data in the field of cigarette shred processing; The ontology network construction module is used to construct an ontology network structure according to the characteristics of cigarette thread making data; The annotation module is used to annotate the unstructured knowledge text based on the ontology network structure to form a data set; wherein the annotation of each entity includes the start and end positions and the entity type; The iteration module is used to iteratively train the entity extraction model based on the data set.

7. The knowledge extraction device in the field of cigarette shred processing according to claim 6, characterized in that: The entity extraction model includes a vector representation layer, a gated structure layer and a conditional random field model layer.

8. The knowledge extraction device in the field of cigarette shred processing according to claim 6, characterized in that: The training module also includes a fusion module, which is used to fuse the data set with an external knowledge graph and / or multimodal cigarette making data.

9. The knowledge extraction device in the field of cigarette shred processing according to claim 6, characterized in that: The dependency analysis module is used to import the entity extraction results into a natural language processing tool in the form of a dictionary to perform dependency syntactic analysis, thereby realizing the word segmentation function of professional field vocabulary.

10. The knowledge extraction device in the field of cigarette shred processing according to claim 6, characterized in that: The triple formation module includes a hierarchical division module and an entity relationship determination module; The hyponym division module is used to perform hyponym division on each component in the to-be-extracted text based on the dependency analysis result, wherein the hyponym division result includes a first entity, a second entity that is a hypernym of the first entity, and a dependency relationship between the first entity and the second entity; The entity relationship determination module is used to take the first entity and the second entity as the head entity and the tail entity respectively, and the dependency relationship as the relationship between the head entity and the tail entity.

Images