Method for intelligently extracting ship collision accident risk influence factor propagation path

Abstract

The invention discloses a method for intelligently extracting a ship collision accident risk influence factor propagation path. The method comprises the following steps: 1, constructing a ship collision accident analysis report data set; 2, constructing a ship collision accident risk influence factor ontology based on a rooting theory; 3, performing fine adjustment and evaluation on the original IE model by using the annotation data in the data set, selecting an optimal fine adjustment IE model according to an evaluation result, and intelligently extracting required entities and relationships from the ship collision accident analysis report data set by using the optimal fine adjustment IE model; and 4, according to a semantic similarity model and a preset similarity threshold value, solving an expression co-reference problem of the extracted entities, and converting a propagation relationship, a causal relationship and related entity matching into an accident risk factor propagation path of which a node is an ontology code. According to the method, the ship collision accident risk influence factor propagation path can be intelligently extracted from a large batch of unstructured accident reports, the result form is simple, the semantic expression is clear, and there is no co-reference problem.

Description

Technical Field

[0001] The present invention belongs to the field of shipping safety and risk management, and specifically is a method for intelligently extracting the propagation path of risk influencing factors of ship collision accidents. Background Art

[0002] With the growth of economic trade and the development of China's Maritime Silk Road, maritime transport carries more than 80% of global trade volume and has a significant driving effect on the world economy. However, with the increase in sea transportation, the number of maritime accidents is very large, resulting in a large number of casualties and losses, among which ship collision accidents account for the largest proportion of maritime accidents. Ship collision accidents are different from other accidents. There are many types of risk influencing factors of ship collision accidents. There are a large number of important dynamic transmission relationships between different accident risk influencing factors. The dynamic transmission relationship is intricate and difficult to simulate using dynamic occurrence mechanisms. Moreover, ship collision accidents are easily affected by the water environment and are characterized by being greatly affected by subjective factors, high frequency of occurrence and large losses.

[0003] At present, the research on the risk influencing factors and blocking problems of ship collision accidents is mainly based on the internal occurrence mechanism of ship collision accidents, studying the various risk influencing factors that cause accidents and the importance of different risk influencing factors to ship collision accidents, and obtaining the optimal blocking path of accidents, thereby providing auxiliary decision-making for reducing ship collision accidents. This type of research mainly involves statistical analysis methods, accident causation models, knowledge graphs and other methods. However, these existing research results cannot simulate the dynamic transmission relationship of risk influencing factors of ship collision accidents, and the analysis of the internal dynamic occurrence mechanism of ship collision accidents is insufficient. Therefore, it is very difficult to analyze the optimal blocking path of ship collision accidents; in addition, there are currently few results on intelligently extracting the propagation path of accident risk influencing factors from unstructured data, and the methods for obtaining the propagation path of risk influencing factors are mostly manual, with low intelligence level, strong subjectivity and inaccurate results. Therefore, based on a large number of ship collision accident reports, intelligently extracting the propagation path of accident risk influencing factors is of great significance for studying the internal occurrence mechanism of accidents and the optimal blocking path of accidents.

[0004] The extraction of propagation paths of risk influencing factors of ship collision accidents is mainly divided into three steps, among which: the first step is to construct the ontology of risk influencing factors of ship collision accidents; the second step is to extract risk influencing factor related entities, propagation relationships and causal relationships from a large number of unstructured ship collision accident analysis reports based on the results of the accident risk influencing factors ontology; the third step is to convert the entity and relationship matching into the propagation path of accident risk influencing factors. Specifically, for the first step, it is necessary to construct the risk influencing factor ontology in the field of ship collision accidents. The existing ontology construction methods include SENSUS method, seven-step method, loop acquisition method and five-step loop method. Among them, the seven-step method, loop acquisition method and five-step loop method are relatively mature. The seven-step method supports the protégé ontology development method of the ontology development tool, which can be directly constructed in combination with the protégé software. It is more practical in the practice of domain knowledge ontology construction. However, the seven-step method lacks the "inspection and evaluation" link in the ontology construction process, and the loop acquisition method lacks specific methods and techniques; the five-step loop method has the disadvantages of insufficient initialization steps and poor operability. In addition, the existing ontology construction methods are all based on manual Or semi-automatically, so there is a certain degree of subjectivity and randomness in the construction process; for the second step, it is necessary to extract relevant entities and relationships from a large number of unstructured accident reports based on the ontology construction results. Most of the existing related studies are based on manual extraction, which has problems such as low intelligence level, time-consuming and labor-intensive, strong subjectivity and inaccurate results. Although the rapid development of relevant IE models in the NLP field, such as BERT, RoBERTa, GPT-3, ERNIE, T5 and UIE, has strong information extraction capabilities, it still cannot meet the specific information extraction tasks and strong domain adaptability requirements in the field of ship collision accident analysis. Therefore, it is still necessary to fine-tune and evaluate the relevant IE models, so as to achieve the purpose of fine-tuning and evaluating the IE model based on a small amount of labeled data to achieve the purpose of intelligently extracting the entities and dynamic transmission relationships of accident risk influencing factors; for the third step, in the process of converting the propagation path of accident risk influencing factors based on entity and relationship matching, there is a problem of expression co-reference. It is necessary to select appropriate semantic models and thresholds to solve such problems based on existing NLP field research.

[0005] In response to the above problems, it is urgent to propose a method for intelligently, objectively and accurately extracting the propagation paths of factors affecting the risk of ship collision accidents from large quantities of unstructured accident reports. Summary of the invention

[0006] In view of the shortcomings of the prior art, the purpose of the present invention is to provide a method for intelligently extracting the propagation paths of risk influencing factors of ship collision accidents, which can automatically and accurately extract the propagation paths of risk influencing factors of ship collision accidents from a large number of unstructured accident reports, and the result is concise in form, clear in semantic expression, and free of co-reference problems, which is convenient for subsequent identification of the internal occurrence mechanism of the accident and the optimal blocking path of the accident.

[0007] In order to achieve the above object, the present invention adopts the following technical solutions:

[0008] A method for intelligently extracting propagation paths of risk influencing factors of ship collision accidents comprises the following steps:

[0009] Step 1: Construct a ship collision accident analysis report dataset;

[0010] Step 2: Based on the grounded theory, the ship collision accident analysis report data set is used to construct the risk influencing factor ontology of ship collision accidents;

[0011] Step 3: Use the annotated data in the dataset to fine-tune and evaluate the original IE model, select the optimal fine-tuned IE model based on the evaluation results, and intelligently extract the required entities and relationships from the ship collision accident analysis report dataset using the optimal fine-tuned IE model based on the ontology of factors affecting the risk of ship collision accidents;

[0012] Step 4: Solve the co-reference problem of the extracted entities based on the semantic similarity model and the pre-set similarity threshold, and convert the propagation relationship, causal relationship and related entity matching into the propagation path of the accident risk factors encoded by the ontology.

[0013] Furthermore, the specific process of step 1 is: obtaining at least 50 different ship collision accident analysis reports, and naming them according to the name and date of the accident, and ensuring that each accident analysis report contains an explanation of the cause, process and result related to the accident, and storing them in "PDF" format to form a ship collision accident analysis report data set.

[0014] Furthermore, the specific process of step 2 is as follows:

[0015] Step 2.1: Open Coding

[0016] The initial concepts in the accident analysis report were extracted and coded by comparative analysis. During the coding process, multiple semantically similar initial concepts were merged into one initial category. The open coding process prohibited the setting of coding rules in advance.

[0017] Step 2.2: Spindle Coding

[0018] Further refine the initial categories and coding results, and merge the initial categories into main categories according to categories; Step 2.3, select coding

[0019] According to the purpose of ship collision accident research, the core categories are refined, the main categories are systematically connected, and the different initial concepts and initial categories are organized;

[0020] Step 2.4: Theoretical saturation detection

[0021] The reliability and adequacy of the initial categories, main categories, and core categories are tested through theoretical saturation testing. If the results of the theoretical saturation testing show new concepts and categories, return to step 2.1 and repeat steps 2.1 to

[0022] Step 2.4: If the theoretical saturation test results do not form new concepts and categories, it means that the ontology construction and coding process is over.

[0023] Furthermore, the original IE model of step 3 is UIE.

[0024] Furthermore, the specific process of step 3 is as follows:

[0025] Step 3.1, use Doccano annotation software to annotate 20% of the samples in the ship collision accident analysis report dataset according to the classification rules of the required entities and relationships, obtain the annotated dataset, and then divide the annotated dataset into a training set, a validation set, and a test set, where: the entities that need to be annotated include participating objects, time, location, cause events, and result events, and the relationships that need to be annotated include causal relationships and propagation relationships;

[0026] Step 3.2: Use the labeled dataset to fine-tune and evaluate the UIE model, and select the best fine-tuned UIE model based on the evaluation results;

[0027] Step 3.3: Based on the ontology of factors affecting ship collision accident risk, the fine-tuned UIE model is used to intelligently extract the required entities and relationships from the ship collision accident analysis report dataset.

[0028] Furthermore, the specific process of step 3.2 is as follows:

[0029] First, export the labeled dataset to a jsonl format file, and then randomly extract 5, 10, and 15 samples from the labeled dataset at one time to form the datasets An{a}, An{b}, and An{c}, respectively. In order to reduce the chance, a total of 10 samples are extracted, thus obtaining the datasets An{a}, An{b}, and An{c}, which contain 10 groups of An{a}, 10 groups of An{b}, and 10 groups of An{c}, respectively. 1 ,a 2 …a 10 、An{b1 ,b 2 …b 10} and An{c 1 ,c 2 …c 10}, abbreviated as An{a i}、An{b i} and An{c i}, where 1≤i≤10, then, using An{a i}、An{b i} and An{c i} Carry out 10 rounds of fine-tuning and model evaluation experiments on the original IE model, record the results, evaluate the results, and conclude that the optimal fine-tuned IE model is the fine-tuned UIE model.

[0030] Furthermore, the specific process of step 3.3 is as follows:

[0031] The fine-tuned UIE model takes a given structural pattern director s and a text sequence x as input and generates a linearized structural extraction language (SEL, y), which contains the information extracted from the text sequence x according to the pattern director s, expressed as:

[0032]

[0033]

[0034] Where y is a structure extraction language (SEL) sequence that can be converted into extracted information records; spot is the target recognition name in a specific information extraction task; asso is the target association name; Represents vector concatenation; text represents text.

[0035] Furthermore, the specific process of step 4 is as follows:

[0036] Step 4.1: Use the cosine similarity model as the semantic similarity model, expressed as:

[0037]

[0038] In the formula, X i and Y i are vectors respectively;

[0039] Step 4.2, set the threshold of cosine similarity to 0.85. For related entities with co-reference problems, when the cosine similarity of the related entities is lower than 0.85, they are judged as dissimilar events. Otherwise, they are judged as similar events. By default, the first entity is retained as the final extraction result.

[0040] Step 4.3, refer to the ontology of factors affecting the risk of ship collision accidents, match the related entities of the problem of unexpressed co-reference, and then convert the propagation relationship, causal relationship and related entity matching into the propagation path of the accident risk factors whose nodes are encoded in the ontology.

[0041] Compared with the prior art, the present invention has the following technical effects:

[0042] On the one hand, the present invention proposes a method for constructing an ontology of risk influencing factors of ship collision accidents based on grounded theory, which can objectively, comprehensively and accurately construct the ontology of accident risk influencing factors based entirely on the ship collision accident data set used, which is conducive to the subsequent acquisition of related entities without coreference problems, and lays a good foundation for accurately extracting the propagation path of risk influencing factors of ship collision accidents. Moreover, it can be combined with protégé software and has strong practicality; on the other hand, based on the fine-tuning IE model, the required entities and relationships are intelligently extracted, which greatly improves the intelligence level of the information extraction link compared with the traditional manual extraction method, and is conducive to automatically, accurately and efficiently extracting the propagation path of risk influencing factors of ship collision accidents from a large number of unstructured accident reports; on the other hand, the extraction result of the present invention is the propagation path of accident risk influencing factors with ontology codes as nodes, the form of the result is more concise, the semantic expression is clear, and there is no coreference problem, which is easier to be referenced and used for related research, and provides a reliable data basis for finding out the internal mechanism of accidents and the optimal blocking path of accidents in the future, which is of great significance to maritime shipping safety.

[0043] The present invention selects the fine-tuning UIE model as the optimal fine-tuning IE model. Compared with other IE models, it has the following three advantages: first, it is no longer necessary to develop three complex structures to meet the needs of different information extraction tasks, settings and scenarios respectively, which is more conducive to knowledge sharing, thereby unifying different types of information extraction tasks. Moreover, compared with task-specific information extraction models, it is not necessary to build three different data sets and knowledge sources, thereby saving more time and energy, unifying different types of information extraction tasks, and making the application more convenient and efficient; second, UIE introduces the structure extraction language (SEL), which can effectively encode different information extraction structures into a unified representation form, and in order to adaptively generate target structures for different information extraction tasks, UIE adopts a structure pattern guide (SSI), which is a pattern-based prompt mechanism for controlling the content to be identified, associated and generated in UIE, so UIE has stronger adaptability to information extraction tasks based on (SEL) and (SSI); third, by using the D pair , D Record and D TextThe extensive pre-training corpus trains UIE's text-to-structure conversion ability, structure decoding ability, and semantic encoding ability, further giving UIE a powerful information extraction ability; in short, fine-tuning UIE has the advantages of requiring less labeled data, being able to meet multi-task extraction requirements at the same time, strong domain adaptability, and high extraction accuracy. Therefore, it is more accurate, efficient, and convenient to extract entities and relationships from large quantities of unstructured accident reports. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 :The present invention constructs a flow chart of the risk influencing factors of ship collision accidents;

[0045] Figure 2 : Flowchart of fine-tuning and evaluating the original IE model of the present invention;

[0046] Figure 3 : The principle example of extracting entities and relations by fine-tuning the UIE model of the present invention;

[0047] Figure 4 : The framework diagram of the fine-tuning UIE model of the present invention for the task of extracting information on ship collision accidents. DETAILED DESCRIPTION

[0048] The specific contents of the present invention are further explained in detail below in conjunction with embodiments.

[0049] A method for intelligently extracting propagation paths of risk influencing factors of ship collision accidents comprises the following steps:

[0050] Step 1: Construct a ship collision accident analysis report dataset. The specific process is as follows:

[0051] First, obtain at least 50 different ship collision accident analysis reports to prevent the small amount of data from affecting the construction of the ship collision accident risk influencing factor ontology and the IE model fine-tuning. Then, name the ship collision accident analysis report according to the name and date of the accident, and ensure that each accident analysis report contains an explanation of the cause, process and result of the accident, and store it in "PDF" format to form a ship collision accident analysis report dataset;

[0052] Step 2: Based on the grounded theory, the risk influencing factors of ship collision accidents are constructed based on the ship collision accident analysis report data set. Figure 1 As shown, the following steps are included:

[0053] Step 2.1: Open Coding

[0054] The initial concepts in the accident analysis report are extracted and coded by comparative analysis. During the coding process, multiple initial concepts with similar semantics are merged into one initial category. The coding process is open, and it is forbidden to set coding rules in advance. Analysts must maintain an objective attitude.

[0055] Step 2.2: Spindle Coding

[0056] Further refine the initial categories and coding results, and merge the initial categories into main categories according to categories; Step 2.3, select coding

[0057] According to the purpose of the ship collision accident research, the core categories are refined, the main categories are systematically connected, and different initial concepts and initial categories are organized. In this embodiment, the core categories of risk influencing factors of ship collision accidents are summarized as external factors and human factors;

[0058] Step 2.4: Theoretical saturation detection

[0059] The reliability and sufficiency of the initial categories, main categories and core categories are tested through theoretical saturation detection. The number of samples used in the theoretical saturation detection in this embodiment accounts for 20% of the total number of samples in the data set; if the theoretical saturation detection result has new concepts and categories, then return to step 2.1 and repeat steps 2.1 to 2.4; if the theoretical saturation detection result does not form new concepts and categories, it means that the ontology construction and coding process is over;

[0060] from Figure 1 It can be seen that compared with the currently commonly used domain ontology construction methods such as the seven-step method, the circular acquisition method and the five-step circular method, the method of constructing the ontology of factors affecting the risk of ship collision accidents based on the grounded theory can objectively, comprehensively and accurately construct the ontology of factors affecting the risk of accidents based on the data set used, and it can be combined with protégé software and has strong practicality;

[0061] Step 3: Use the labeled data in the dataset to fine-tune and evaluate the original IE model. Select the optimal fine-tuned IE model based on the evaluation results. Based on the ontology of factors affecting the risk of ship collision accidents, use the optimal fine-tuned IE model to intelligently extract the required entities and relationships from the ship collision accident analysis report dataset. The specific process is as follows: Figure 2 As shown, the following steps are included:

[0062] Step 3.1, use Doccano annotation software to annotate 20% of the samples in the ship collision accident analysis report dataset according to the classification rules of the required entities and relationships, obtain the annotated dataset, and then divide the annotated dataset into a training set, a validation set, and a test set, where: the entities that need to be annotated include participating objects, time, location, cause events, and result events, and the relationships that need to be annotated include causal relationships and propagation relationships;

[0063] Step 3.2: Use the labeled dataset to fine-tune and evaluate the IE model. According to the evaluation results, the optimal fine-tuned IE model is finally selected. The specific process is as follows: first export the labeled dataset as a jsonl format file, and then randomly extract 5, 10, and 15 samples from the labeled dataset at one time to form the datasets An{a}, An{b}, and An{c} respectively. In order to reduce the contingency, a total of 10 samples are extracted, thus obtaining the datasets An{a}, An{b}, and An{c}, which contain 10 groups of An{a}, 10 groups of An{b}, and 10 groups of An{c} respectively. 1 ,a 2 …a 10 、An{b 1 ,b 2 …b 10} and An{c 1 ,c 2 …c 10}, abbreviated as An{a i}、An{b i} and An{c i}, where 1≤i≤10, then, using An{a i}、An{b i} and An{c i} Carry out 10 rounds of fine-tuning and model evaluation experiments on the original IE model, record the results, evaluate the results, and obtain the optimal fine-tuned IE model;

[0064] Since the original IE model selected in this embodiment is the UIE model, the optimal fine-tuning IE model finally obtained is the fine-tuning UIE model. Compared with other conventional IE models, the UIE model has the following three advantages:

[0065] 1) It can unify different types of information extraction tasks, making the application more convenient and efficient

[0066] like Figure 3 As shown in the figure, first, it is no longer necessary to develop three complex structures to meet the needs of different information extraction tasks, settings, and scenarios respectively; second, the UIE model, as a model that integrates extraction tasks, settings, and scenario requirements, is more conducive to knowledge sharing; finally, compared with task-specific information extraction models, the UIE model does not need to build three different data sets and knowledge sources, thus saving more time and effort;

[0067] 2) Stronger adaptability to information extraction tasks based on SEL and SSI

[0068] In order to model heterogeneous information extraction structures, UIE introduces a structure extraction language (SEL), which can effectively encode different information extraction structures into a unified representation. In addition, in order to adaptively generate target structures for different information extraction tasks, UIE adopts a structure pattern instructor (SSI). The structure pattern instructor is a pattern-based prompt mechanism used to control the content to be identified, associated, and generated in UIE. The overall framework of applying it to the ship collision accident information extraction task is as follows: Figure 4 As shown;

[0069] 3) Powerful information extraction capabilities based on large-scale corpus pre-training

[0070] The UIE model has a powerful information extraction capability and utilizes the D pair , D Record and D Text The extensive pre-training corpus is used to train UIE’s text-to-structure conversion ability, structure decoding ability, and semantic encoding ability. Three loss functions are used in the training process, namely L Pair , L Record and L Text , the total loss Loss is the sum of these three losses, L Pair , L Record , L Text And the total loss Loss are expressed as:

[0071]

[0072] Loss = L Pair +L Record +L Text

[0073] In the formula, θ e and θ d are the parameters of the encoder and decoder respectively, x′ is the corrupted source text, and x" is the corrupted target segment;

[0074] Step 3.3: Based on the ontology of factors affecting the risk of ship collision accidents, the fine-tuned UIE model is used to intelligently extract the required entities and relationships from the ship collision accident analysis report dataset. The specific process is as follows:

[0075] The fine-tuned UIE model takes a given structural pattern director s and a text sequence x as input and generates a linearized structural extraction language (SEL, y), which contains the information extracted from the text sequence x according to the pattern director s, expressed as:

[0076]

[0077] Where y is a structure extraction language (SEL) sequence that can be converted into extracted information records; spot is the target recognition name in a specific information extraction task, such as "ship" or "location"; asso is the target association name, such as "collision"; Represents vector concatenation; text represents text;

[0078] Step 4: Solve the co-reference problem of the extracted entities based on the semantic similarity model and the pre-set similarity threshold, and convert the propagation relationship, causal relationship and related entity matching into the propagation path of the accident risk factors whose nodes are encoded in the ontology. The specific process is as follows:

[0079] Step 4.1: Solve the co-reference problem of the extracted entities according to the semantic similarity model and the set threshold;

[0080] Because in the text description of the cause event and the result event in the ship collision accident report, the same cause or the same result may appear multiple times in different writing methods, such as "negligence in lookout", "negligence in lookout", "failure to maintain regular lookout", "failure to maintain continuous lookout", "minor damage to the hull", "minor damage to the hull". For such event co-reference problems, the event trigger words have similar meanings, the event structure is not complicated, and there is little connection with the context semantics. They are basically caused by differences in the way of textual description. Therefore, the event co-reference problem can be solved by determining the semantic similarity of the events;

[0081] Common semantic similarity models include word embedding-based models (Word2Vec, GloVe), context-based embedding models (BERT, GPT), sentence embedding-based models (Universal Sentence Encoder, Sentence-BERT), graph-based models (WordNet, Knowledge Graphs), deep learning models (T5, XLNet) and cosine similarity models. Considering that the amount of data to be judged is not large, and the calculation speed needs to be fast and the computing power is low, combined with the actual application scenario, this embodiment selects the cosine similarity model as the semantic similarity model, which is expressed as:

[0082]

[0083] In the formula, X i and Y i are vectors respectively;

[0084] After many experiments, the threshold of cosine similarity is set to 0.85. For related entities with co-reference problems, when the cosine similarity of related entities is lower than 0.85, they are judged as dissimilar events. Otherwise, they are judged as similar events. By default, the first entity is retained as the final extraction result.

[0085] Step 4.2: Convert the propagation relationship, causal relationship and related entity matching into nodes, which are the propagation paths of accident risk factors encoded in the ontology.

[0086] The results of entities that have solved the coreference problem (i.e., extracted entities without the coreference problem) are matched with reference to the ontology coding of ship collision accident risk influencing factors, and then converted into propagation paths of ship collision accident risk factors with nodes as ontology coding.

[0087] In order to verify the performance of the fine-tuned UIE model, this embodiment selects the T5 model commonly used in the field of information extraction as the basic model, and fine-tunes the T5 model and UIE model based on 5-shot, 10-shot and 15-shot small batch labeled samples, respectively. The extraction results of event trigger words, event arguments and relations are compared and analyzed using the fine-tuned T5 model and the fine-tuned UIE model. The hardware configuration of the experiment is Tesla V100 32G, RAM 32G, CPU Core 4. The relevant parameters in the fine-tuning process are: batch-size is 16; train-epoch is 10; learning-rate is 1e -5 The F1 scores of the two models after fine-tuning are shown in Table 1;

[0088] Table 1 Comparison of F1 scores between UIE model and T5 model after fine-tuning

[0089]

[0090] As can be seen from Table 1, compared with the fine-tuned T5 model, the fine-tuned UIE model achieved higher F1 scores for the extraction of event trigger words, event arguments and relations, indicating that the fine-tuned UIE model has the best information extraction effect on the ship collision accident analysis report dataset.

[0091] In order to verify the reliability of the method for intelligently extracting the propagation path of risk influencing factors of ship collision accidents proposed in this embodiment, the investigation report of the collision accident between the "Sanshui 805" ship and the "Zhesanyu 00046" ship is taken as an example. After information extraction and solving the problem of co-reference of expressions, the results are shown in Table 2, and the propagation path results of accident risk factors are shown in Table 3;

[0092] In Table 3, the meanings of the ontology codes are: IS01 for lookout error; IS06 for illegal operation; IS07 for speeding; IS09 for improper avoidance measures; IS11 for lack of experience; IS13 for lack of safety awareness; IS15 for unqualified or incompetent crew / captain; and IS18 for heavy fog / low visibility.

[0093] It can be seen from Tables 1 and 2 that the method for intelligently extracting the propagation path of factors affecting the risk of ship collision accidents proposed in this embodiment can not only solve the problem of co-reference of the extracted entities and construct an ontology of factors affecting the risk of ship collision accidents, but also the extraction result is the propagation path of the accident risk influencing factors with the ontology code as the node. The form of the result is more concise and the semantic expression is clear, which is easier to refer to and use for related research.

[0094] Table 2 Results of information extraction and coreference elimination

[0095]

[0096] Table 3. Results of transmission paths of accident risk factors

[0097] coding Transmission Path 1 IS18=>IS07=>IS14=>Collision accident 3 2 IS09=>Collision accident 3 3 IS01=>IS09=>Collision accident 3 4 IS13=>IS15=>IS11=>IS06=>Collision accident 3 5 IS15=>IS06=>Collision accident 3

Claims

1. A method for intelligently extracting the propagation path of factors affecting the risk of ship collision accidents, characterized in that: The steps include: Step 1: Construct a ship collision accident analysis report dataset; Step 2: Based on the grounded theory, the ship collision accident analysis report data set is used to construct the risk influencing factor ontology of ship collision accidents; Step 3: Use the annotated data in the dataset to fine-tune and evaluate the original IE model, select the optimal fine-tuned IE model based on the evaluation results, and intelligently extract the required entities and relationships from the ship collision accident analysis report dataset using the optimal fine-tuned IE model based on the ontology of factors affecting the risk of ship collision accidents; Step 4: Solve the co-reference problem of the extracted entities based on the semantic similarity model and the pre-set similarity threshold, and convert the propagation relationship, causal relationship and related entity matching into the propagation path of the accident risk factors encoded by the ontology.

2. The method for intelligently extracting the propagation path of factors affecting the risk of ship collision accidents according to claim 1 is characterized in that: The specific process of step 1 is: obtain at least 50 different ship collision accident analysis reports, and name them according to the name and date of the accident, and ensure that each accident analysis report contains a description of the cause, process and result of the accident, and store them in "PDF" format to form a ship collision accident analysis report data set.

3. The method for intelligently extracting the propagation path of factors affecting the risk of ship collision accidents according to claim 1 is characterized in that: The specific process of step 2 is: Step 2.1: Open Coding The initial concepts in the accident analysis report were extracted and coded by comparative analysis. During the coding process, multiple semantically similar initial concepts were merged into one initial category. The open coding process prohibited the setting of coding rules in advance. Step 2.2: Spindle Coding The initial categories and coding results were further refined, and the initial categories were merged into main categories according to categories; Step 2.3, select encoding According to the purpose of ship collision accident research, the core categories are refined, the main categories are systematically connected, and the different initial concepts and initial categories are organized; Step 2.4: Theoretical saturation detection The reliability and adequacy of the initial categories, main categories, and core categories are tested through theoretical saturation detection. If the results of the theoretical saturation detection show new concepts and categories, return to step 2.1 and repeat steps 2.1 to 2.

4. If the results of the theoretical saturation detection do not form new concepts and categories, the ontology construction and coding process is completed.

4. The method for intelligently extracting the propagation path of the risk influencing factors of ship collision accidents according to claim 1 is characterized in that: The original IE model in step 3 is UIE.

5. The method for intelligently extracting the propagation path of the risk influencing factors of ship collision accidents according to claim 4 is characterized in that: The specific process of step 3 is as follows: Step 3.1, use Doccano annotation software to annotate 20% of the samples in the ship collision accident analysis report dataset according to the classification rules of the required entities and relationships, obtain the annotated dataset, and then divide the annotated dataset into a training set, a validation set, and a test set, where: the entities that need to be annotated include participating objects, time, location, cause events, and result events, and the relationships that need to be annotated include causal relationships and propagation relationships; Step 3.2: Use the labeled dataset to fine-tune and evaluate the UIE model, and select the best fine-tuned UIE model based on the evaluation results; Step 3.3: Based on the ontology of factors affecting ship collision accident risk, the fine-tuned UIE model is used to intelligently extract the required entities and relationships from the ship collision accident analysis report dataset.

6. The method for intelligently extracting the propagation path of the risk influencing factors of ship collision accidents according to claim 5 is characterized in that: The specific process of step 3.2 is as follows: First, export the labeled dataset to a jsonl format file, and then randomly extract 5, 10, and 15 samples from the labeled dataset at one time to form the datasets An{a}, An{b}, and An{c}, respectively. In order to reduce the chance, a total of 10 samples are extracted, thus obtaining the datasets An{a1,a2…a{a}, which contain 10 groups of An{a}, 10 groups of An{b}, and 10 groups of An{c}, respectively. 10 、An{b1,b2…b 10 } and An{c1,c2…c 10 }, abbreviated as An{a i }、An{b i } and An{c i }, where 1≤i≤10, then, using An{a i }、An{b i } and An{c i } Carry out 10 rounds of fine-tuning and model evaluation experiments on the original IE model, record the results, evaluate the results, and conclude that the optimal fine-tuned IE model is the fine-tuned UIE model.

7. The method for intelligently extracting the propagation path of the risk influencing factors of ship collision accidents according to claim 5 or 6, characterized in that: The specific process of step 3.3 is as follows: The fine-tuned UIE model takes a given structural pattern director s and a text sequence x as input and generates a linearized structural extraction language (SEL, y), which contains the information extracted from the text sequence x according to the pattern director s, expressed as: Where y is a structure extraction language (SEL) sequence that can be converted into extracted information records; spot is the target recognition name in a specific information extraction task; asso is the target association name; Represents vector concatenation; text represents text.

8. The method for intelligently extracting the propagation path of the risk influencing factors of ship collision accidents according to claim 5 or 6, characterized in that: The specific process of step 4 is as follows: Step 4.1: Use the cosine similarity model as the semantic similarity model, expressed as: In the formula, X i and Y i are vectors respectively; Step 4.2, set the threshold of cosine similarity to 0.

85. For related entities with co-reference problems, when the cosine similarity of the related entities is lower than 0.85, they are judged as dissimilar events. Otherwise, they are judged as similar events. By default, the first entity is retained as the final extraction result. Step 4.3, refer to the ontology of factors affecting the risk of ship collision accidents, match the related entities of the problem of unexpressed co-reference, and then convert the propagation relationship, causal relationship and related entity matching into the propagation path of the accident risk factors whose nodes are encoded in the ontology.

Images