Chemical process knowledge extraction method fusing auxiliary annotation and dependency analysis

Abstract

The invention discloses a chemical process knowledge extraction method fusing auxiliary annotation and dependency analysis, and belongs to the technical field of natural language processing. The method comprises the following steps: collecting production materials and operation and maintenance materials of a chemical process as original corpora; processing and annotating partial data in the original corpus, including data set segmentation, keyword acquisition, keyword annotation, data set annotation and data set format conversion to obtain a data set with a tag; inputting the data set with the label and the original data into a knowledge extraction model to perform preliminary knowledge extraction, including BiLSTM-CRF model training, named entity identification by using the trained model, and relation extraction by using a dependency syntax analysis method; the initial results are fused, and a knowledge extraction task is completed; the extracted knowledge is stored in a graph database in a triple format, which is convenient for workers to inquire and learn, so as to reduce the dependence of operation and maintenance work on experts, improve the operation and maintenance efficiency and reduce the personnel training cost.

Description

Technical Field

[0001] The present invention relates to the technical field of natural language processing, and specifically to a knowledge extraction method in a chemical production process, and in particular to a chemical process knowledge extraction method that integrates auxiliary annotation and dependency analysis. Background Art

[0002] Chemical processes are an integral part of modern society, playing a key role in the production of chemicals, pharmaceuticals, food, and materials, among other fields. Due to the complex chemical reactions and physical processes involved, the large number of equipment involved, and the vast amount of associated production and operation and maintenance data, chemical process operations are challenging. The textual information contained within these numerous production and operation and maintenance documents is often difficult for laypeople to understand, resulting in high learning costs for new employees and an over-reliance on experienced staff and experts for chemical process operations, which limits production efficiency and increases labor costs for enterprises.

[0003] To address these issues, it is necessary to extract knowledge from this information and convert it into structured data that is easy to understand and query. Knowledge extraction technologies can generally be divided into three categories: rule-based and dictionary-based methods, machine learning-based methods, and deep learning-based methods. Rule-based and dictionary-based methods rely on manually formulated rules or templates for knowledge extraction. They are effective for extracting knowledge in specific fields, but require manual intervention and customization, and have poor portability. Machine learning-based methods treat tasks such as named entity recognition as sequence labeling problems, using statistical models to identify entities. Compared with rule-based methods, they have better generalization capabilities but require large amounts of labeled data for training. Deep learning-based methods mainly rely on multiple neural networks to complete knowledge extraction tasks. This method can automatically learn the complex features of text and has achieved remarkable results in knowledge extraction tasks in recent years. However, the amount of data required for training is even larger, and it is difficult to collect data that meets the requirements in actual chemical production processes. Therefore, it is necessary to label the raw text data that can be collected.

[0004] Given the diverse and complex nature of Chinese text data, the current annotation of Chinese text relies primarily on manual annotation using certain annotation software or platforms. These software programs are not only cumbersome to install but also have stringent requirements for the operating environment. The annotation process is also independent of the knowledge extraction process, resulting in poor overall integrity. Furthermore, when the same keyword appears in different sentences, manual annotation often needs to be repeated, which not only significantly wastes time but also reduces the overall efficiency of the knowledge extraction task. Therefore, it is particularly important to develop efficient and holistic data annotation methods. Summary of the Invention

[0005] The object of the present invention is to provide a method for extracting chemical process knowledge by integrating auxiliary annotation and dependency analysis, to overcome the technical problems in the above technologies, and to accurately complete the task of extracting knowledge from production materials and operation and maintenance materials. At the same time, it has the characteristics of good integrity and high efficiency of the method.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] A method for extracting chemical process knowledge by integrating auxiliary annotation and dependency analysis includes the following steps:

[0008] S1, original corpus collection; collect the production materials and operation and maintenance materials involved in the chemical production process as the original corpus; the collected original corpus is pure text data without image information;

[0009] S2, data preprocessing and auxiliary annotation; data preprocessing includes: the original corpus is simultaneously subjected to two parts of operations. The first part is to perform a segmentation operation on the basis of the number of text natural paragraphs at a specified ratio to obtain a segmented data set; the second part is to perform operations such as word segmentation, stop word removal, duplicate word removal, and part-of-speech screening to obtain a list of words to be annotated; data auxiliary annotation annotates some words in the list of words to be annotated to obtain a list of words with labels; match the list of words with labels with the segmented data set, and the corresponding labels of the matched words will be annotated on the corresponding words in the segmented data set to obtain a data set with labels; perform format conversion on the data set with labels to generate a data set in BMEOS annotation format for subsequent model training;

[0010] S3, model training and preliminary knowledge extraction; the model is obtained by combining a bidirectional long short-term memory neural network and a stochastic conditional field, and this model is used to perform the named entity recognition task of chemical texts; first, use the data set in BMEOS annotation format generated in S2 to train this model, and the trained model is used to complete the named entity recognition of chemical texts to obtain a chemical text entity set and a corresponding label set; then, combine the entities in the chemical text entity set with their corresponding labels in pairs to obtain a plurality of binary sequence structures of entity-entity labels; combine all the binary sequence structures of entity-entity labels to obtain a chemical text entity quadruple sequence structure of head entity-head entity label-tail entity-tail entity label; use the dependency parsing method to perform relationship extraction to obtain a plurality of initial chemical process knowledge triple sequence structures of head entity-relationship-tail entity;

[0011] S4. Initial knowledge extraction result fusion: Calculate the similarity between the head and tail entities of the chemical text entity quadruple sequence structure and the initial chemical process knowledge triple sequence structure. For the data corresponding to the chemical text entity quadruple sequence structure and the initial chemical process knowledge triple sequence structure with a similarity reaching the set value, perform a fusion operation. Complete the "relationship" in the initial chemical process knowledge triple sequence structure to the chemical text entity quadruple sequence structure, and remove the head entity label and tail entity label to form a new chemical process knowledge triple sequence of head entity - relationship - tail entity, thus completing knowledge fusion.

[0012] S5. Knowledge storage and knowledge query: Store all the new chemical process knowledge triple sequence data obtained from knowledge fusion as a chemical process knowledge graph.

[0013] Furthermore, the chemical process includes large and small chemical process data that can collect chemical process technical regulations, technical manuals, and fault handling log text data, not limited to a certain production process. The production materials and operation and maintenance materials include chemical process technical regulations of chemical production lines, technical manuals of chemical production equipment, and production process fault handling logs.

[0014] Furthermore, S2 includes the following steps;

[0015] S21. Dataset segmentation: Segment the original corpus according to a specified ratio of the number of lines to obtain a training set, a validation set, and a test set to be labeled. The specified ratio is 6:2:2.

[0016] S22. Keyword acquisition: First, tokenize each sentence in the original corpus, and perform stop word removal and deduplication on the obtained list of words to obtain a new list of words. Then, perform part-of-speech analysis on the obtained new list of words, and filter out words with part-of-speech tags of noun n, person name nr, and place name ns to form the final output list of words to be labeled.

[0017] S23. Keyword annotation: Select each type of word in the list of words to be labeled obtained in S22, add labels to each selected type of word respectively, and assign the labels of each type of word to each word in that type of word to obtain a list of labeled words.

[0018] S24. Dataset annotation: Match the training set, validation set, and test set to be labeled in S21 with the list of labeled words in S23, add the labels corresponding to the matched words to the training set, validation set, and test set, and output the labeled training set, validation set, and test set, which are saved in json format.

[0019] S25, Dataset format conversion; perform format conversion on the labeled training set, validation set, and test set in JSON format in S24, convert them into data files labeled by the BMEOS annotation method, and save them as text data.

[0020] Furthermore, the JSON format in S24 includes: the sentence text "text" key, the sentence number "id" key, and the labels "label" in the sentence; among them, the "label" key includes the start position "start", the end position "end", the vocabulary name "text", and the vocabulary label "labels" of the first character of each label.

[0021] Furthermore, the text data in S25 includes: splitting the vocabulary by characters, and labeling the tags of each character of the vocabulary according to the BMEOS format; if the vocabulary has no label, all Chinese characters are labeled as O; if the label of the vocabulary is T, all Chinese characters are labeled as T, single-character vocabulary is labeled as S-T, and the tags of each character of multi-character vocabulary are expressed as:

[0022]

[0023] Among them, is the tag of the th character, is the position of the first Chinese character, is the position of the ending Chinese character; B means the Chinese character representing the mark is at the beginning position of the vocabulary; M means the Chinese character representing the mark is in the middle of the vocabulary; E means the Chinese character representing the mark is at the end of the vocabulary.

[0024] Furthermore, in S3,

[0025] S31, input the chemical engineering text data in the dataset labeled in BMEOS annotation format into the model for training; the model includes an input layer, a long short-term memory network layer, a random conditional field layer, and an output layer;

[0026] Furthermore, the rule matching method in S33 includes using the dependency syntactic analysis method to analyze the dependency relationship of the vocabulary in the sentence, and extracting the relationship according to the matching rule for the obtained vocabulary and dependency relationship, to obtain the initial chemical engineering process knowledge triple sequence structure of head entity - relationship - tail entity.

[0027] Furthermore, the matching rule includes using the verb as the core relationship, and the combination rule of "relationship" in the initial chemical engineering process knowledge triple sequence structure of head entity - relationship - tail entity is "the adverbial-middle relationship word of the core relationship word + the core relationship word / the attributive-middle relationship word + the verb-complement structure of the core relationship word + the right-adjunct relationship of the core relationship word";

[0028] The entity matching rules include the following:

[0029] Entity matching rule one: The subject-predicate word of the core relational word serves as the head entity, and the object-predicate word of the core relational word serves as the tail entity;

[0030] Entity matching rule two: The subject-predicate word of the core relational word serves as the head entity, and the object-prepositional word of the complement structure of the core relational word serves as the tail entity;

[0031] Entity matching rule three: The object-prepositional word of the adverbial-center structure of the core relational word serves as the head entity, and the prepositional object of the core relational word serves as the tail entity;

[0032] Entity matching rule four: The core relational word serves as the head entity, and the object-predicate relationship of the core relational word serves as the tail entity.

[0033] Furthermore, S4 includes the following steps;

[0034] S41: Use word2vec to obtain the vector representations of the entity words in the two chemical knowledge tuple sequences in S3;

[0035] S42: Use the vector representations of the entity words in the two chemical knowledge tuple sequences in S41 to calculate the similarity of the head entity and the tail entity respectively, obtaining the head similarity and the tail similarity;

[0036] S43: If both the head similarity and the tail similarity are greater than the set threshold, then supplement the relationship of the initial chemical process knowledge triple sequence structure of head entity - relationship - tail entity to the chemical text entity quadruple sequence structure of head entity - head entity label - tail entity - tail entity label, forming: head entity - head entity label - relationship - tail entity - tail entity label structure, and at the same time remove the head entity label and the tail entity label, forming: head entity - relationship - tail entity, to complete knowledge extraction; otherwise, remove this quadruple and perform knowledge fusion on the next quadruple in the chemical text entity quadruple sequence.

[0037] Furthermore, the formula for similarity calculation in S42 is as follows:

[0038]

[0039] Wherein, represents the vector of the entity in the chemical text entity quadruple, represents the vector of the initial chemical process knowledge triple, and respectively represent the th chemical entity feature vector data in the two vectors, Represents the length of the chemical entity feature vector.

[0040] If both the head similarity and the tail similarity are greater than the set threshold of 0.8, the relationship of the corresponding triple is supplemented to the quadruple with which the similarity is calculated, forming a quintuple with the structure "head entity, head entity label, relationship, tail entity, tail entity label". At the same time, the labels in the tuple are removed, forming a triple with the structure "head entity, relationship, tail entity" to complete knowledge extraction; otherwise, the quadruple is removed and knowledge fusion of the next quadruple in the chemical text entity quadruple sequence is performed.

[0041] The beneficial effects of the present invention are as follows:

[0042] 1. The knowledge extraction method for chemical processes proposed by the present invention does not require a large amount of standardized labeled data as a training basis, does not involve cumbersome data sorting work, and does not require a large amount of manual labeling work. The knowledge extraction task can be completed by collecting Chinese text data of production materials and operation and maintenance materials in chemical plants (examples: process technical regulations of chemical production lines, operation manuals of chemical production equipment, fault handling logs), which has strong versatility and feasibility.

[0043] 2. The knowledge extraction method for chemical processes proposed by the present invention is a chemical entity annotation method that integrates a highly automated annotation method, improving the integrity of the method. At the same time, it greatly reduces the number of times of manual annotation of the corpus, improves the efficiency of Chinese text annotation, and thus improves the overall task completion efficiency of chemical knowledge extraction, and is advanced and efficient in this task.

[0044] 3. The chemical knowledge extraction method proposed by the present invention combines rule matching, which ensures the extraction accuracy while reducing the requirements of the method for training data, reducing the data annotation pressure, and further improving the operation speed of the method and the efficiency of chemical knowledge extraction.

[0045] 4. The knowledge extraction method for chemical processes proposed by the present invention does not require additional processing of Chinese text data in chemical processes. Only by inputting text format data into the model can the knowledge in production materials and operation and maintenance materials (examples: process technical regulations, equipment operation manuals, fault handling logs) be extracted.

[0046] 5. The knowledge extraction method for chemical processes proposed by the present invention is a means of chemical process knowledge extraction with versatility and universality, and is not aimed at a specific production process or production unit. It is applicable to chemical processes with production materials and operation and maintenance materials (examples: process technical regulations of chemical production lines, operation manuals of chemical production equipment, fault handling logs), and has good versatility. Brief Description of the Drawings

[0047] Figure 1 It is a schematic flow chart of the knowledge extraction method for chemical processes according to the present invention.

[0048] Figure 2 It is a principle block diagram of the chemical entity annotation method with a high degree of automation according to the present invention.

[0049] Figure 3 It is a flow chart of step S23 according to the present invention.

[0050] Figure 4 It is a principle block diagram of the chemical knowledge extraction method according to the present invention, where Eh*, R*, and Et* are respectively the "head entity", "relation", and "tail entity" in the triple result of relation extraction; Eh, Lh, Et, and Lt are respectively the "head entity", "head entity label", "tail entity", and "tail entity label" in the quadruple result of named entity recognition; Sh and St are respectively the head entity similarity and the tail entity similarity. Detailed implementation manners

[0051] To make the objectives, technical solutions, and advantages of the present application clearer. The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0052] In view of the problem that it is difficult to learn and query production materials and operation and maintenance materials in the chemical production process, the present invention proposes a chemical process knowledge extraction method that combines assisted annotation and dependency analysis. The main implementation flow chart of this method is as shown in the appendix Figure 1 shown below. The following describes the entire process according to Figure 1 the system; it is mainly divided into five steps. The overall process of the knowledge extraction method for chemical processes can be described as follows:

[0053] First, collect the original corpus, then process and annotate some data in the original corpus to obtain a labeled data set. Input the labeled data set and the original data into the knowledge extraction model for preliminary knowledge extraction. Perform named entity recognition through the trained model, use the dependency syntax analysis method for relation extraction, fuse the initial results, and complete the knowledge extraction task. Finally, store the extracted knowledge in the graph database in the form of triples.

[0054] In the knowledge extraction method for chemical processes provided by the present invention, it is necessary to first collect production materials and operation and maintenance materials of chemical processes as the original corpus, and then perform operations on processing and annotation of some data in the original corpus. The operation steps include dataset segmentation, keyword acquisition, keyword annotation, dataset annotation, and dataset format conversion to obtain a labeled dataset. The labeled dataset and the original data are input into the knowledge extraction model for preliminary knowledge extraction, including BiLSTM-CRF model training, using the trained model for named entity recognition, and using the dependency parsing method for relation extraction; fusing the initial results to complete the knowledge extraction task; the extracted knowledge is stored in the graph database in the form of triples for easy query by staff, so as to reduce the dependence on experts during operation and maintenance work.

[0055] Specifically, S1 is to collect various production materials and operation and maintenance materials in chemical processes as the original corpus. These materials include but are not limited to the process technical regulations of chemical production lines, technical manuals of chemical production equipment, and fault handling logs. It mainly depends on the recording degree of the logs by previous staff and the completeness of chemical process design materials.

[0056] S2 is data preprocessing and auxiliary annotation, and its main implementation framework is as shown in the appendix Figure 2 It mainly includes dataset segmentation, keyword acquisition, keyword annotation, dataset annotation, and dataset format conversion.

[0057] S21 is dataset segmentation, detecting the number of natural paragraphs in the original corpus , the length of the training set after segmentation is , the length of the validation set is , the length of the test set is . The segmentation ratio is 6:2:2, 、 and The calculation formulas of are expressed as:

[0058]

[0059]

[0060]

[0061] S22 is keyword acquisition. First, perform data preprocessing on the original corpus . The model splits the input original corpus sentence by sentence, and then performs word segmentation, stop word removal, and duplicate removal operations on each sentence to obtain a preliminary processed vocabulary list , expressed as:

[0062]

[0063]

[0064]

[0065] Among them, represents the original corpus, represents the th word in the th sentence of the chemical engineering corpus, represents the deduplication operation, represents the list of chemical engineering vocabulary after deduplication for the th sentence, represents the list of stop words in the chemical engineering text,

[0066] Then, for the obtained vocabulary list perform part-of-speech analysis, filter out words with parts of speech of noun "n", person name "nr", and place name "ns", and form the final output vocabulary list to be labeled.

[0067] S23 is keyword annotation, and the process flow of this step is as shown in the appendix Figure 3 as follows. First, read the vocabulary list to be labeled obtained in S22, select the words with the same label to be labeled, insert and save the label. If further annotation is needed, the already annotated words will be removed and the vocabulary list will be updated; if no further annotation is needed, the labeled vocabulary list will be output.

[0068] S24 is dataset annotation. Match the training set, validation set, and test set to be labeled in S21 with the labeled vocabulary list in S23, add the labels corresponding to the matched words to the training set, validation set, and test set, and output the labeled training set, validation set, and test set, which are saved in json format. The saved json format: sentence text "text" key, sentence number "id" key, and labels in the sentence "label" key, where the "label" key includes the starting position "start" of the first character of each label word, the ending position "end" of the last character of the word, the word name "text", and the word label "labels".

[0069] S25 is the dataset format conversion, which converts the labeled training set, validation set, and test set in JSON format in S24 and saves them as text data. During annotation, the vocabulary is first split by character, and then each character of the vocabulary is annotated according to the BMEOS format; if the vocabulary has no label, all Chinese characters are annotated as O; if the label of the vocabulary is T, all Chinese characters are annotated as T, and single-character vocabulary is annotated as S-T, and each character label of multi-character vocabulary is expressed as:

[0070]

[0071] Among them, is the label of the th character, is the position of the first Chinese character, is the position of the ending Chinese character.

[0072] The principle block diagram of the knowledge extraction method is shown in Appendix Figure 4 as shown, including two parts: preliminary knowledge extraction and fusion of preliminary knowledge extraction results.

[0073] S3 is the preliminary knowledge extraction, which mainly includes three parts: model training, named entity recognition, and relation extraction.

[0074] S31 is model training. The labeled data is input into a model combined with a bidirectional long short-term memory neural network (BiLSTM) and a conditional random field (CRF) for training. The model includes an input layer, a long short-term memory network layer, a conditional random field layer, and an output layer. The bidirectional long short-term memory network (BiLSTM) captures the context information in the text. The output of BiLSTM generates the scores of each label through a fully connected layer. The CRF layer receives the scores from BiLSTM and considers the connections between labels to find the sequence with the highest probability among all possible label sequences. Input the BiLSTM-CRF neural network in batches, set the batch sampling to 32, and set parameters and hyperparameters such as the learning rate, pulse neuron threshold, and number of graph convolutional layers to train the model.

[0075] S32 is named entity recognition. Use the trained model for named entity recognition. The input of the model is sentence-level text information, and the output is a sequence of multiple quadruples with the structure of "head entity, head entity label, tail entity, tail entity label".

[0076] S33 is relation extraction, which uses the method of rule matching for relation extraction. The input is sentence-level text information, and the output is multiple triples with the structure of "head entity, relation, tail entity". This method usually uses verbs as the core relation words. The combination rule of "relation" in the triple is "adverbial modifier of the core relation word + core relation word / attributive of the core relation word + verb-complement structure of the core relation word + right adjunct of the core relation word".

[0077] The entity matching rules in the triple are as follows:

[0078] Entity matching rule one: The subject-predicate relation word of the core relation word is used as the head entity, and the object-verb relation word of the core relation word is used as the tail entity; Entity matching rule two: The subject-predicate relation word of the core relation word is used as the head entity, and the object-preposition relation word of the verb-complement structure of the core relation word is used as the tail entity; Entity matching rule three: The object-preposition relation word of the adverbial-modifier structure of the core relation word is used as the head entity, and the prepositional object of the core relation word is used as the tail entity; Entity matching rule four: The core relation word is used as the head entity, and the object-verb relation of the core relation word is used as the tail entity.

[0079] S4 is the fusion of preliminary knowledge extraction results, which includes three parts: entity word embedding, similarity calculation, and knowledge fusion.

[0080] S41 is entity word embedding, which uses word2vec to obtain the vector representation of the entity vocabulary in the two chemical knowledge tuple sequences in S3.

[0081] S42 is similarity calculation, which uses the vector representation of the entity vocabulary in the two chemical knowledge tuple sequences in S41 to calculate the similarity of the head entity and the tail entity respectively, and obtains the head similarity and the tail similarity. The similarity calculation formula is:

[0082]

[0083] Among them, represents the vector of the entity in the quadruple of chemical text entities, represents the vector of the initial chemical process knowledge triple, and respectively represent the th chemical entity feature vector data in the two vectors, represents the length of the chemical entity feature vector.

[0084] S43 is knowledge fusion. If both the head similarity and the tail similarity are greater than the set similarity threshold of 0.8, the relationship of the triple is supplemented into the quadruple to form a quintuple with the structure of "head entity, head entity label, relationship, tail entity, tail entity label", and at the same time, the labels in the tuple are removed to form a triple with the structure of "head entity, relationship, tail entity", completing knowledge extraction; otherwise, the quadruple is removed and knowledge fusion of the next quadruple in the chemical text entity quadruple sequence is performed.

[0085] S5 is knowledge storage. The knowledge extracted from production materials and operation and maintenance materials (examples: process technical regulations of chemical production lines, operation manuals of chemical production equipment, fault handling logs) is stored in a graph database, which is convenient for staff to consult and learn, improves operation and maintenance efficiency, and reduces personnel training costs.

[0086] The knowledge extraction method for chemical processes proposed by the present invention establishes a knowledge extraction model with a highly automated chemical entity annotation method. The collected original corpus is annotated by the annotation module, and knowledge extraction is performed using BiLSTM-CRF and rule matching to obtain knowledge in the form of triples, realizing efficient knowledge extraction. The obtained knowledge is stored in a graph database, which is convenient for staff to consult and learn, improves operation and maintenance efficiency, and reduces personnel training costs. The knowledge extraction method proposed by the present invention has advantages especially for a large number of production materials and operation and maintenance materials existing in chemical processes. Traditional deep learning models need to pre-annotate and process text data, and the integrity of the method is poor. In the face of the situation where the same keyword appears in different sentences, manual annotation often needs to be repeated, which not only greatly wastes time but also reduces the overall efficiency of the knowledge extraction task. The model proposed by the present invention obtains a labeled vocabulary list through operations such as word segmentation, part-of-speech screening, and class-based annotation of the original corpus, peels off the original data, automatically annotates the matched vocabulary, and finally performs data format conversion. The output data can be directly used for deep learning model training, greatly improving the completion efficiency of the knowledge extraction task and the integrity of the knowledge extraction model.

[0087] In the description of this specification, the descriptions referring to terms such as "one embodiment", "example", "specific example", etc. mean that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example.

[0088] The preferred embodiments of the present invention disclosed above are only used to help illustrate the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to the specific embodiments described. Obviously, many modifications and variations can be made according to the content of this specification. These embodiments are selected and specifically described in this specification to better explain the principles and practical applications of the present invention, so that those skilled in the art can well understand and utilize the present invention. The present invention is only limited by the claims and their full scope and equivalents.

[0089] Finally, it should be noted that the above describes the basic principles, main features and advantages of the present invention. The above embodiments are only used to illustrate the technical solutions of the present application, and are not intended to limit them. Although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features. These modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the various embodiments of the present application. Without departing from the spirit and scope of the present invention, the present invention will also have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed.

Claims

1. A chemical process knowledge extraction method integrating auxiliary annotation and dependency analysis, characterized in that It includes the following steps: S1. Original corpus collection: Collect production materials and operation and maintenance materials involved in the chemical production process as the original corpus. The original corpus is pure text data. S2. Data preprocessing and data-assisted annotation: The data preprocessing obtains two parts of content. The first part is the segmented data set, and the second part is the vocabulary list to be annotated. The data-assisted annotation annotates some of the vocabulary in the vocabulary list to be annotated to obtain a tagged vocabulary list. The tagged vocabulary list and the segmented data set are annotated to obtain a data set annotated in the BMEOS format. S3. Model training and preliminary knowledge extraction: The model is obtained by combining a bidirectional long short-term memory neural network and a stochastic conditional field. The model is used to perform the named entity recognition task on chemical texts. First, the model is trained using the data set annotated in the BMEOS format generated in S2. The trained model is used to complete the named entity recognition of chemical texts to obtain a chemical text entity set and a corresponding label set. Then, the entities in the chemical text entity set are combined with their corresponding labels to obtain a chemical text entity quadruple sequence structure of head entity - head entity label - tail entity - tail entity label. The dependency syntax analysis method is used for relationship extraction to obtain an initial chemical process knowledge triple sequence structure of multiple head entity - relationship - tail entity. S4. Fusion of preliminary knowledge extraction results: Calculate the similarity between the head entities and tail entities of the chemical text entity quadruple sequence structure and the initial chemical process knowledge triple sequence structure, and perform a fusion operation on the corresponding data of the chemical text entity quadruple sequence structure and the initial chemical process knowledge triple sequence structure whose similarity reaches the set value to complete knowledge fusion. S5: Knowledge storage and knowledge query: Store the data in all new chemical process knowledge triple sequences of head entity - relationship - tail entity obtained by knowledge fusion as a chemical process knowledge graph.

2. The chemical process knowledge extraction method integrating auxiliary annotation and dependency analysis according to claim 1, characterized in that The chemical process includes large and small chemical process data such as chemical process technical regulations, technical manuals that can be collected, and fault handling log text data. The production materials and operation and maintenance materials include chemical production line process technical regulations, technical manuals of chemical production equipment, and production process fault handling logs.

3. The chemical process knowledge extraction method integrating auxiliary annotation and dependency analysis according to claim 1, characterized in that S2 includes the following steps; S21. Data set segmentation: Segment the original corpus according to a specified ratio of the number of lines to obtain a training set, a validation set, and a test set to be annotated. The specified ratio is 6:2:

2. S22. Keyword acquisition: First, segment each sentence in the original corpus, and perform stop word removal and duplicate removal on the obtained vocabulary list to obtain a new vocabulary list. Then, perform part-of-speech analysis on the obtained new vocabulary list, and screen out the vocabulary with part-of-speech of noun n, person name nr, and place name ns to form the final output vocabulary list to be annotated. S23, Keyword annotation: Select each category of words in the list of words to be annotated obtained in S22, add tags to each selected category of words respectively, and assign the tags of each category of words to each word in that category to obtain a list of tagged words; S24, Dataset annotation; Match the training set, validation set, and test set to be annotated in S21 with the list of tagged words in S23, add the tags corresponding to the matched words to the training set, validation set, and test set, and output the annotated training set, validation set, and test set, which are saved in JSON format; S25, Dataset format conversion: Convert the JSON-format training set, validation set, and test set annotated in S24 into data files annotated by the BMEOS annotation method and save them as text data.

4. The chemical process knowledge extraction method integrating auxiliary annotation and dependency analysis according to claim 3, characterized in that The JSON format in S24 includes: the sentence text "text" key, the sentence number "id" key, and the tags "label" in the sentence; among them, the "label" key includes the starting position "start" of the first character of each tagged word, the ending position "end" of the last character of the word, the word name "text", and the word tag "labels".

5. The chemical process knowledge extraction method integrating auxiliary annotation and dependency analysis according to claim 3, characterized in that The text data in S25 includes: splitting the words by character, and annotating the tags of each character of the word in the BMEOS format; if the word has no tag, all Chinese characters are annotated as O; if the tag of the word is T, all Chinese characters are annotated as T, single-character words are annotated as S-T, and the tags of each character of multi-character words are expressed as: Among them, is the label of the th character, is the position of the first Chinese character, is the position of the ending Chinese character; B means the Chinese character representing the mark is at the beginning of the word; M means the Chinese character representing the mark is in the middle of the word; E means the Chinese character representing the mark is at the end of the word.

6. The chemical process knowledge extraction method integrating auxiliary annotation and dependency analysis according to claim 1, characterized in that In S3, S31, Input the chemical process text data in the dataset annotated in the BMEOS format into the model for training; the model includes an input layer, a long short-term memory network layer, a conditional random field layer, and an output layer.

7. The chemical process knowledge extraction method integrating auxiliary annotation and dependency analysis according to claim 6, characterized in that, The method for rule matching in S33 includes using the dependency parsing method to analyze the dependency relationships of the words in the sentence, and extracting the relationships of the obtained words and dependency relationships according to the matching rules to obtain the initial chemical process knowledge triple sequence structure of head entity - relationship - tail entity.

8. The chemical process knowledge extraction method integrating auxiliary annotation and dependency analysis according to claim 7, characterized in that, The matching rules include taking the verb as the core relationship, and the "relationship" combination rule in the initial chemical process knowledge triple sequence structure of head entity - relationship - tail entity is "the adverbial-middle relationship word of the core relationship word + the core relationship word / the attributive-middle relationship word + the verb-complement structure of the core relationship word + the right adjunct relationship of the core relationship word"; The entity matching rules include the following: Entity matching rule one, the subject-predicate relationship word of the core relationship word is used as the head entity, and the object-complement relationship word of the core relationship word is used as the tail entity; Entity matching rule two, the subject-predicate relationship word of the core relationship word is used as the head entity, and the object-complement structure's prepositional object of the core relationship word is used as the tail entity; Entity matching rule three, the prepositional object of the adverbial-middle structure of the core relationship word is used as the head entity, and the prepositional object before the core relationship word is used as the tail entity; Entity matching rule four, the core relationship word is used as the head entity, and the object-complement relationship of the core relationship word is used as the tail entity.

9. The chemical process knowledge extraction method integrating auxiliary annotation and dependency analysis according to claim 6, characterized in that S4 includes the following steps; S41. Obtain the vector representations of the entity words in the two chemical knowledge tuple sequences in S3 using word2vec; S42. Calculate the similarities of the head entity and the tail entity respectively using the vector representations of the entity words in the two chemical knowledge tuple sequences in S41 to obtain the head similarity and the tail similarity; S43. If both the head similarity and the tail similarity are greater than the set threshold, supplement the relationship in the initial chemical process knowledge triple sequence structure of head entity - relationship - tail entity to the chemical text entity quadruple sequence structure of head entity - head entity label - tail entity - tail entity label to form: head entity - head entity label - relationship - tail entity - tail entity label structure, remove the head entity label and the tail entity label to form: head entity - relationship - tail entity, and complete knowledge extraction; otherwise, remove this quadruple and perform knowledge fusion on the next quadruple in the chemical text entity quadruple sequence.

10. The chemical process knowledge extraction method integrating auxiliary annotation and dependency analysis according to claim 9, characterized in that The formula for similarity calculation in S42 is as follows: Among them, represents the vector of entities in the quadruple of chemical text entities, represents the vector of the initial chemical process knowledge triple, and respectively represent the th chemical entity feature vector data in the two vectors, represents the length of the chemical entity feature vector.

Images