Method for intelligently extracting bilingual ship accident risk influence factor information

Abstract

The invention discloses a method for intelligently extracting bilingual ship accident risk influence factor information, which comprises the following steps of: 1, formulating a data labeling rule according to an extraction task; 2, ship accident sample data needing to be labeled are extracted from the data set through a stratified sampling method, and samples extracted each time contain Chinese and English corpora; 3, marking the extracted ship accident sample data according to a preset marking rule by utilizing marking software, and exporting the ship accident sample data in a json type to obtain a text sequence; 4, dividing the marked ship accident sample data into a training set, a verification set and a test set; 5, the UIE-Multilingual model is trained and evaluated, and parameters of the model are finely adjusted; 6, selecting a UIE-Multilingual model with the optimal parameter according to the F1 score, and storing the selected UIE-Multilingual model, so as to obtain a fine tuning UIE-Multilingual model; and 7, the accident risk influence factors are extracted from a large batch of bilingual unstructured ship accident reports by using a fine tuning UIE-Multilingual model, Chinese and English corpora do not need to be labeled respectively, and the method is more efficient and accurate.

Description

Technical Field

[0001] The invention relates to shipping safety and risk management, and in particular to a method for intelligently extracting bilingual information on ship accident risk influencing factors. Background Art

[0002] With the growth of economic trade and the development of China's Maritime Silk Road, maritime transport currently carries more than 80% of global trade volume, which has a significant driving effect on the world economy. However, with the increase in maritime transport, the number of maritime accidents is very large, resulting in a large number of casualties and losses. Therefore, the analysis of factors affecting ship accident risks has always been the focus of research in the ship and shipping related industries.

[0003] At present, most of the data sources in the shipping field are unstructured accident analysis reports. Accident risk influencing factors are contained in these accident analysis reports, and the dominant data sources (such as China MSA and MAIB) include English and Chinese. Intelligently extracting the required accident risk influencing factor information from these large-scale, unstructured bilingual data sources is a prerequisite for analyzing the risk influencing factors of ship accidents. Therefore, how to intelligently extract accident risk influencing factors from large quantities of bilingual ship accident analysis reports is worthy of in-depth study.

[0004] There are four main methods for extracting information on factors affecting ship accident risks, including manual extraction, rule-based methods, machine learning-based methods, and deep learning-based methods. Among them: manual extraction methods are aimed at scenarios with small data volumes and simple extraction tasks; rule-based methods generally extract information from texts through predefined rules or templates, and are suitable for specific task scenarios with simple tasks, small data volumes, and a small number of rules; machine learning-based methods are more flexible than rule-based methods and can learn implicit patterns; deep learning-based methods mainly use relevant information extraction models in the NLP field to extract information on factors affecting accident risks. They have automatic learning features, are suitable for complex sentences and long texts, have strong extraction capabilities, can handle complex context dependencies and nested relationships, and have certain knowledge transfer capabilities.

[0005] However, the above-mentioned methods for extracting information on factors affecting ship accident risks have the following problems: 1) The manual extraction method is time-consuming and labor-intensive, highly subjective and has a low level of intelligence for large quantities of accident reports in the field of ship accident risk factor analysis; 2) The rule-based method faces large amounts of data and complex information extraction requirements, and has problems such as high rule design and maintenance costs, poor versatility, and lack of context understanding capabilities; 3) The machine learning-based method requires the construction of feature engineering and the annotation of a large amount of data, which is time-consuming and labor-intensive, and the extraction accuracy depends on the quality of the annotated data and feature engineering; 4) The deep learning-based method can use a large number of information extraction models in the NLP field, but because there are many professional terms in the field of ship accidents and the data set is in both Chinese and English, the direct use of existing information extraction models is not accurate.

[0006] Therefore, it is necessary to develop an intelligent method for extracting information on factors affecting ship accident risks from large amounts of bilingual unstructured data. Summary of the invention

[0007] In view of the shortcomings of the prior art, the purpose of the present invention is to provide a method for intelligently extracting bilingual information on ship accident risk influencing factors, which can intelligently extract accident risk influencing factors from large quantities of bilingual unstructured ship accident reports without the need to separately label Chinese and English corpora, making it more efficient and accurate.

[0008] In order to achieve the above object, the present invention adopts the following technical solutions:

[0009] A method for intelligently extracting bilingual information on ship accident risk influencing factors includes the following steps:

[0010] Step 1: Formulate data annotation rules based on the extraction task;

[0011] Step 2: Use stratified sampling method to extract ship accident sample data that need to be labeled from the data set, and each sample drawn covers Chinese and English corpora;

[0012] Step 3: Use annotation software to annotate the extracted ship accident sample data according to the preset annotation rules, and export it in the "json" type to obtain a text sequence;

[0013] Step 4: Divide the labeled ship accident sample data into a training set, a validation set, and a test set;

[0014] Step 5: Use the training set to train the UIE-Multilingual model. During the training process, use the validation set to evaluate the performance of the UIE-Multilingual model and fine-tune the parameters of the UIE-Multilingual model.

[0015] Step 6: According to the F1 score, the fine-tuned UIE-Multilingual model is evaluated, and the UIE-Multilingual model with the best parameters is selected and stored to obtain the fine-tuned UIE-Multilingual model;

[0016] Step 7: Use the fine-tuned UIE-Multilingual model to extract accident risk influencing factors from a large number of bilingual unstructured ship accident reports.

[0017] Furthermore, the sample data of ship accidents that need to be labeled extracted in step 2 accounts for 20% of the total number of samples in the data set.

[0018] Furthermore, the annotation software in step 3 is doccano open source text annotation tool.

[0019] Furthermore, in step 4, the ratio of the number of samples in the training set, validation set and test set is 8:1:1.

[0020] Furthermore, the number of labeled samples in the training set is 5-shot, 10-shot or 15-shot.

[0021] Furthermore, in step 5, the specific process of training the UIE-Multilingual model is as follows:

[0022] Step 5.1: Input representation

[0023] Take SSI and the text sequence X in the training set i As the input of UIE-Multilingual model, the text sequence X i Convert to a continuous vector representation h 0 , expressed as:

[0024] h 0 =TokenEmbedding(x)+PositionalEmbedding(x)+SegmentEmbedding(x)

[0025] In the formula, Token Embedding(x) represents mapping the words in the input sequence to the vector space, Positional Embedding(x) and Segment Embedding(x) provide position encoding and paragraph encoding for the Transformer encoder respectively, and x is the text sequence X i A collection of;

[0026] Step 5.2: Feature extraction

[0027] Extract the vector representation h through the pre-trained ERNIE encoder 0 The ERNIE encoder uses a multi-layer Transformer encoder as the basic encoder. The Transformer encoder contains multiple self-attention mechanisms and feed-forward neural network layers. The Transformer encoder extracts the vector representation h 0 The process of contextual semantic features is expressed as follows:

[0028] h=Transformer(h 0 ;θ)

[0029] Where θ represents the linear transformation weight W in Transformer 1 and W 2 , query Q, key K, value matrix V, and bias vector b 1 and b 2 , h is the output of the Transformer encoder;

[0030] Step 5.3: Detect event trigger words and extract event arguments

[0031] The pooled output of the ERNIE encoder is used as the text sequence X i The global extraction result of ERNIE encoder is used as the sensitive representation result of each token. The pooling output and encoding output are used as the input of the extraction task-specific head to detect event trigger words and extract event arguments.

[0032] The extraction task-specific header includes an event extraction header and an entity extraction header, wherein: the event extraction header uses the pooled output of the ERNIE encoder as input to detect event trigger words; the entity extraction head uses the encoded output of the ERNIE encoder as input to extract event arguments;

[0033] The process of detecting event trigger words by the event extraction head is as follows:

[0034] The event extraction head detects whether each input position i is an event trigger word. If the position i is an event trigger word, the category of the trigger word is identified, which is expressed as:

[0035] p(t i |x)=Softmax(W t ·h i +b t )

[0036] In the formula, p(t i |x) represents the probability distribution of the event trigger word category at position i, t i Indicates event trigger words, hi represents the Transformer feature representation of position i, W t and b t Represent the weight and bias of the linear transformation, C t is the number of event trigger word categories, d is the dimension of the attention head;

[0037] In order to reduce the difference between the predicted output of the event extraction head and the true label, the cross entropy loss function is used as the trigger word detection loss, which is expressed as:

[0038]

[0039] The process of extracting event arguments from the entity extraction head is as follows:

[0040] The entity extraction head predicts the argument of each position i, expressed as:

[0041] p(a i |x,t)=Softmax(W a ·h i +b a )

[0042] In the formula, p(a i |x, t) is the probability distribution of the argument category at position i; a i represents the argument at position i, t is the trigger word position information, which is used as the conditional input; W a and b a Represent the weight and bias of the linear transformation, C a is the number of argument categories, d is the dimension of the attention head;

[0043] In order to reduce the difference between the predicted output of the entity extraction head and the true label, the cross entropy loss function is used as the argument extraction loss, which is expressed as:

[0044]

[0045] Step 5.4: Event type prediction

[0046] First, the event trigger word t i The expression h t and argument a i The expression h a Aggregate and get the overall representation of the event, expressed as:

[0047] h c =f(h t +h a )

[0048] In the formula, ht The event trigger word t i The expression of h a For the argument a i , f() is an aggregation function, which means weighted average or concatenation;

[0049] Then predict the probability distribution of event type, expressed as:

[0050] p(c|x,t,a)=Softmax(W c ·h c +b c )

[0051] Where p(c|x,t,a) is the probability distribution of event type, W c and b c Respectively represent the weight and bias of the linear transformation, h c represents the overall representation of the event, and c represents the event type;

[0052] In order to reduce the difference between the predicted output event type and the true label value, the cross entropy loss function is used as the event type prediction loss, which is expressed as:

[0053] L type =-logp(c|x,y,a)

[0054] Step 5.5: Joint Optimization

[0055] The loss L for trigger word detection trigger , argument extraction loss L argument and event type prediction loss L type The weighted summation gives the total loss, expressed as:

[0056] L total =λ 1 L trigger +λ 2 L argument +λ 3 L type

[0057] In the formula, λ 1 , 2 and λ 3 All represent weight coefficients;

[0058] Furthermore, in step 5.2, the processing process of the self-attention mechanism on the input text sequence x is expressed as:

[0059]

[0060] Where Q, K, and V represent query, key, and value matrices, respectively, and Q, K, and V are all represented by the input vector h0 Linear transformation yields, d k is the dimension of the attention head.

[0061] Furthermore, in step 5.2, the processing process of the feedforward neural network layer on the input text sequence x is expressed as:

[0062] FFN(x)=ReLU(xW 1 +b 1 )W 2 +b 2

[0063] Where W 1 and W 2 is the linear transformation weight, b 1 and b 2 is the bias vector.

[0064] Furthermore, in step 5, the AdamW optimizer is used to adaptively fine-tune the parameters of the UIE-Multilingual model.

[0065] Compared with the prior art, the present invention has the following technical effects:

[0066] 1) The fine-tuned UIE-Multilingual model proposed in the present invention combines the context modeling capability of Transformer and the joint prediction of event trigger words, arguments and event types, and realizes the task of intelligently and efficiently extracting accident risk influencing factors from a large number of bilingual and unstructured ship accident reports. Compared with the manual extraction method, the intelligence level is greatly improved. Compared with the conventional information extraction model in the field of natural language processing, there is no need to separately annotate the Chinese and English corpora, nor is there a need to separately train the Chinese information extraction model and the English information extraction model, which is more time-saving and labor-saving. In addition, the use of Chinese and English corpora for simultaneous training of the model can improve the knowledge transfer capability, thereby improving the extraction accuracy.

[0067] 2) The fine-tuned UIE-Multilingual model proposed in the present invention is more suitable for extracting information on risk influencing factors and accident results of various types of ship accidents. It is not only suitable for information extraction from Chinese-English bilingual data sets, but also suitable for information extraction from Chinese or English monolingual data sets. Different fine-tuning strategies can be adopted according to different information extraction tasks, thereby expanding to accident information extraction scenarios in other fields.

[0068] 3) The fine-tuning UIE-Multilingual model proposed in the present invention makes the results easier to preserve and more reusable. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1: Flowchart of the present invention for fine-tuning the UIE-Multilingual model. DETAILED DESCRIPTION

[0070] The specific contents of the present invention are further explained in detail below in conjunction with embodiments.

[0071] A method for intelligently extracting bilingual information on ship accident risk influencing factors includes the following steps:

[0072] Step 1: Formulate data annotation rules based on the extraction task;

[0073] Step 2: Use stratified sampling method to extract ship accident sample data that need to be labeled from the data set. The extracted ship accident sample data that need to be labeled accounts for 20% of the total sample in the data set, and each sample drawn covers Chinese and English corpora;

[0074] The data in the dataset comes from China MSA (China Maritime Safety Administration) and MAIB (Marine Accident Investigation Branch);

[0075] Step 3: Use the open source text annotation tool doccano to annotate the extracted ship accident sample data according to the preset annotation rules, and export it in the "json" type to obtain a text sequence;

[0076] Step 4: Divide the labeled ship accident sample data into training set, validation set and test set in a ratio of 8:1:1;

[0077] Step 5: Use the training set to train the UIE-Multilingual model (Universal Information Extraction Multilingual Model). The number of labeled samples in the training set is 5-shot, 10-shot, or 15-shot. During the training process, the validation set is used to evaluate the performance of the UIE-Multilingual model and fine-tune the parameters of the UIE-Multilingual model. The specific process of training the UIE-Multilingual model is as follows: Figure 1 As shown, including:

[0078] Step 5.1: Input representation

[0079] Take SSI and the text sequence X in the training set i As the input of UIE-Multilingual model, the text sequence X iConvert to a continuous vector representation h 0 , expressed as:

[0080] h 0 =Token Embedding(x)+Positional Embedding(x)+Segment Embedding(x)

[0081] In the formula, Token Embedding(x) represents mapping the words in the input sequence to the vector space, Positional Embedding(x) and Segment Embedding(x) provide position encoding and paragraph encoding for the Transformer encoder respectively, and x is the text sequence X i A collection of;

[0082] Step 5.2: Feature extraction

[0083] Extract the vector representation h through the pre-trained ERNIE encoder (Enhanced Representation through kNowledgeIntegration) 0 The ERNIE encoder is developed by Baidu and uses a multi-layer standard Transformer encoder as the basic encoder, which is convenient for docking with existing pre-trained models and transfer learning.

[0084] The Transformer encoder contains multiple self-attention mechanisms and feed-forward neural network layers. The Transformer encoder extracts the vector representation h 0 The process of contextual semantic features is expressed as follows:

[0085] h=Transformer(h 0 ;θ)

[0086] Where θ represents the linear transformation weight W in Transformer 1 and W 2 , query matrix Q, key matrix K, value matrix V, and bias vector b 1 and b 2 , h is the output of the Transformer encoder;

[0087] The processing process of the self-attention mechanism on the input text sequence x is expressed as:

[0088]

[0089] Where Q, K, and V represent query, key, and value matrices, respectively, and Q, K, and V are all represented by the input vector h 0 Linear transformation yields, d k is the dimension of the attention head;

[0090] The processing process of the feedforward neural network layer on the input text sequence x is expressed as:

[0091] FFN(x)=ReLU(xW 1 +b 1 )W 2 +b 2

[0092] Where W 1 and W 2 is the linear transformation weight, b 1 and b 2 is the bias vector;

[0093] Step 5.3: Detect event trigger words and extract event arguments

[0094] The pooled output of the ERNIE encoder is used as the text sequence X i The global extraction result of ERNIE encoder is used as the sensitive representation result of each token. The pooling output and encoding output are used as the input of the extraction task-specific head to detect event trigger words and extract event arguments.

[0095] The extraction task-specific header includes an event extraction header and an entity extraction header, wherein: the event extraction header uses the pooled output of the ERNIE encoder as input to detect event trigger words; the entity extraction head uses the encoded output of the ERNIE encoder as input to extract event arguments;

[0096] The process of detecting event trigger words by the event extraction head is as follows:

[0097] The event extraction head detects whether each input position i is an event trigger word. If the position i is an event trigger word, the category of the trigger word is identified, which is expressed as:

[0098] p(t i |x)=Softmax(W t ·h i +b t )

[0099] In the formula, p(t i |x) represents the probability distribution of the event trigger word category at position i, t i Indicates event trigger words, h i represents the Transformer feature representation of position i, Wt and b t Represent the weight and bias of the linear transformation, C t is the number of event trigger word categories, d is the dimension of the attention head;

[0100] In order to reduce the difference between the predicted output of the event extraction head and the true label, the cross entropy loss function is used as the trigger word detection loss, which is expressed as:

[0101]

[0102] The process of extracting event arguments from the entity extraction head is as follows:

[0103] The entity extraction head predicts the argument of each position i, expressed as:

[0104] p(a i |x, t) = Sofmtax(W a ·h i +b a )

[0105] In the formula, p(a i |x, t) is the probability distribution of the argument category at position i; a i represents the argument at position i, t is the trigger word position information, which is used as the conditional input; W a and b a Represent the weight and bias of the linear transformation, C a is the number of argument categories, d is the dimension of the attention head;

[0106] In order to reduce the difference between the predicted output of the entity extraction head and the true label, the cross entropy loss function is used as the argument extraction loss, which is expressed as:

[0107]

[0108] Step 5.4: Event type prediction

[0109] Event-based trigger words i and argument a i Predict the event type. The specific process is as follows:

[0110] First, the event trigger word t i The expression h t and argument a i The expression h a Aggregate and get the overall representation of the event, expressed as:

[0111] h c =f(h t +h a )

[0112] In the formula, h t The event trigger word t i The expression of h a For the argument a i , f() is an aggregation function, which means weighted average or concatenation;

[0113] Then predict the probability distribution of event type, expressed as:

[0114] p(c|x,t,a)=Softmax(W c ·h c +b c )

[0115] Where p(c|x,t,a) is the probability distribution of event type, W c and b c Respectively represent the weight and bias of the linear transformation, h c represents the overall representation of the event, and c represents the event type;

[0116] In order to reduce the difference between the predicted output event type and the true label value, the cross entropy loss function is used as the event type prediction loss, which is expressed as:

[0117] L type = -logp(c|x,t,a)

[0118] Step 5.5: Joint Optimization

[0119] The loss L for trigger word detection trigger , argument extraction loss L argument and event type prediction loss L type The weighted sum is used to obtain the total loss, which is used to balance the losses of different tasks and is expressed as:

[0120] L total =λ 1 L trigger +λ 2 L argument +λ 3 L type

[0121] In the formula, λ 1 , 2 and λ 3 Both represent weight coefficients

[0122] This embodiment uses the AdamW optimizer commonly used in the NLP field to adaptively fine-tune the parameters of the UIE-Multilingual model;

[0123] Step 6: According to the F1 score, the fine-tuned UIE-Multilingual model is evaluated, and the UIE-Multilingual model with the best parameters is selected and stored to obtain the fine-tuned UIE-Multilingual model;

[0124] Step 7: Use the fine-tuned UIE-Multilingual model to extract accident risk influencing factors from a large number of bilingual unstructured ship accident reports;

[0125] In the inference process of the fine-tuned UIE-Multilingual model, event trigger words, event arguments and event types are output in the order of trigger word detection, argument extraction and event type prediction. The UIE-Multilingual model combines the context modeling capability of Transformer and the joint prediction of event trigger words, arguments and event types, and realizes the task of intelligently extracting accident risk influencing factors from a large number of bilingual and unstructured ship accident reports. Compared with the manual extraction method, the intelligence level is greatly improved. Compared with the conventional information extraction model in the field of natural language processing, there is no need to separately annotate the Chinese and English corpora and train the Chinese and English extraction models separately, which is more time-saving and labor-saving. In addition, the use of Chinese and English corpora for simultaneous training of the model can improve the knowledge transfer capability, thereby improving the extraction accuracy.

[0126] In order to verify the accuracy of extracting risk influencing factor information from a large number of ship accident analysis reports in Chinese and English bilingual materials using the fine-tuned UIE-Multilingual model of this embodiment, the T5 model (The Text-to-Text Transfer Transformer model) commonly used in the field of natural language processing was trained with training sets containing 5-shot, 10-shot and 15-shot labeled samples, respectively, to obtain a fine-tuned T5 model as a comparison model;

[0127] The F1 scores of the event information extraction results of the fine-tuned UIE-Multilingual model and the fine-tuned T5 model trained on the training set of 5-shot, 10-shot and 15-shot labeled samples are shown in Table 1:

[0128] Table 1 Comparison of the effects of fine-tuning UIE-Multilingual model and fine-tuning T5 model on event information extraction

[0129]

[0130] It can be seen from Table 1 that in the case of 5-shot, 10-shot and 15-shot labeled samples, the F1 score of the fine-tuned UIE-Multilingual model is always higher than the F1 score of the fine-tuned T5 model, indicating that the performance of the fine-tuned UIE-Multilingual model is better than that of the fine-tuned T5 model. The method for training the UIE-Multilingual model proposed in this embodiment is more suitable for the UIE-Multilingual model. The obtained fine-tuned UIE-Multilingual model can more accurately extract accident risk influencing factors and accident types, and has higher practicality.

[0131] Application Examples

[0132] Taking the real accident analysis report of the collision between the Changping ship and the Xinwang 138 ship on January 2, 2018 as an example, the fine-tuned UIE-Multilingual model proposed in this embodiment is used to extract event information. The results are shown in Table 2.

[0133] Table 2 Results of real event information extraction using fine-tuned UIE-Multilingual model

[0134]

[0135] As can be seen from Table 2, the fine-tuned UIE-Multilingual model extracts the required trigger word information and event argument information from the real accident analysis report, and predicts two event types: risk influencing factor events and result events based on the trigger words and event arguments.

Claims

1. A method for intelligently extracting bilingual information on ship accident risk influencing factors, characterized in that: The steps include: Step 1: Formulate data annotation rules based on the extraction task; Step 2: Use stratified sampling method to extract ship accident sample data that need to be labeled from the data set, and each sample drawn covers Chinese and English corpora; Step 3: Use annotation software to annotate the extracted ship accident sample data according to the preset annotation rules, and export it in "json" type to obtain a text sequence; Step 4: Divide the labeled ship accident sample data into a training set, a validation set, and a test set; Step 5: Use the training set to train the UIE-Multilingual model. During the training process, use the validation set to evaluate the performance of the UIE-Multilingual model and fine-tune the parameters of the UIE-Multilingual model. Step 6: According to the F1 score, evaluate the fine-tuned UIE-Multilingual model, select the UIE-Multilingual model with the best parameters and store it to obtain the fine-tuned UIE-Multilingual model; Step 7: Use the fine-tuned UIE-Multilingual model to extract accident risk influencing factors from a large number of bilingual unstructured ship accident reports.

2. The method for intelligently extracting bilingual ship accident risk influencing factor information according to claim 1 is characterized in that: The sample data of ship accidents that need to be labeled extracted in step 2 account for 20% of the total number of samples in the data set.

3. The method for intelligently extracting bilingual ship accident risk influencing factor information according to claim 1 is characterized in that: The annotation software in step 3 is doccano open source text annotation tool.

4. The method for intelligently extracting bilingual information on ship accident risk influencing factors according to claim 1 is characterized in that: In step 4, the ratio of the number of samples in the training set, validation set and test set is 8:1:

1.

5. The method for intelligently extracting bilingual information on ship accident risk influencing factors according to claim 1 is characterized in that: The number of labeled samples in the training set is 5-shot, 10-shot or 15-shot.

6. The method for intelligently extracting bilingual ship accident risk influencing factor information according to claim 1 is characterized in that: In step 5, the specific process of training the UIE-Multilingual model is as follows: Step 5.1: Input representation Take SSI and the text sequence X in the training set i As the input of UIE-Multilingual model, the text sequence X i Converted to a continuous vector representation h0, expressed as: h0=Token Embedding(x)+Positional Embedding(x)+Segment Embedding(x) In the formula, Token Embedding(x) represents mapping the words in the input sequence to the vector space, Positional Embedding(x) and Segment Embedding(x) provide position encoding and paragraph encoding for the Transformer encoder respectively, and x is the text sequence X i A collection of; Step 5.2: Feature extraction The pre-trained ERNIE encoder extracts the contextual semantic features of the vector representation h0. The ERNIE encoder uses a multi-layer Transformer encoder as the basic encoder. The Transformer encoder contains multiple self-attention mechanisms and feedforward neural network layers. The process of extracting the contextual semantic features of the vector representation h0 by the Transformer encoder is expressed as follows: h = Transformer(h0;θ) Where θ represents the linear transformation weights W1 and W2, the query matrix Q, the key matrix K, the value matrix V, and the bias vectors b1 and b2 in the Transformer, and h is the output of the Transformer encoder; Step 5.3: Detect event trigger words and extract event arguments The pooled output of the ERNIE encoder is used as the text sequence X i The global extraction result of ERNIE encoder is used as the sensitive representation result of each token. The pooling output and encoding output are used as the input of the extraction task-specific head to detect event trigger words and extract event arguments. The extraction task-specific header includes an event extraction header and an entity extraction header, wherein: the event extraction header uses the pooled output of the ERNIE encoder as input to detect event trigger words; the entity extraction head uses the encoded output of the ERNIE encoder as input to extract event arguments; The process of detecting event trigger words by the event extraction head is as follows: The event extraction head detects whether each input position i is an event trigger word. If the position i is an event trigger word, the category of the trigger word is identified, which is expressed as: p(t i |x)=Softmax(W t ·h i +b t ) In the formula, p(t i |x) represents the probability distribution of the event trigger word category at position i, t i Indicates event trigger words, h i represents the Transformer feature representation of position i, W t and b t Represent the weight and bias of the linear transformation, C t is the number of event trigger word categories, d is the dimension of the attention head; In order to reduce the difference between the predicted output of the event extraction head and the true label, the cross entropy loss function is used as the trigger word detection loss, which is expressed as: The process of extracting event arguments from the entity extraction head is as follows: The entity extraction head predicts the argument of each position i, expressed as: p(a i |x,t)=Softmax(W a ·h i +b a ) In the formula, p(a i |x, t) is the probability distribution of the argument category at position i; a i represents the argument at position i, t is the trigger word position information, which is used as the conditional input; W a and b a Represent the weight and bias of the linear transformation, C a is the number of argument categories, d is the dimension of the attention head; In order to reduce the difference between the predicted output of the entity extraction head and the true label, the cross entropy loss function is used as the argument extraction loss, which is expressed as: Step 5.4: Event type prediction First, the event trigger word t i The expression h t and argument a i The expression h a Aggregate and get the overall representation of the event, expressed as: h c =f(h t +h a ) In the formula, h t The event trigger word t i The expression of h a For the argument a i , f() is an aggregation function, which means weighted average or concatenation; Then predict the probability distribution of event type, expressed as: p(c|x,t,a)=Softmax(W c ·h c +b c ) Where p(c|x,t,a) is the probability distribution of event type, W c and b c Respectively represent the weight and bias of the linear transformation, h c represents the overall representation of the event, and c represents the event type; In order to reduce the difference between the predicted output event type and the true label value, the cross entropy loss function is used as the event type prediction loss, which is expressed as: L type =-log p(c|x,t,a) Step 5.5: Joint Optimization The loss L for trigger word detection trigger , argument extraction loss L argument and event type prediction loss L type The weighted summation gives the total loss, expressed as: L total =λ1L trigger +λ2L argument +λ3L type Where λ1, λ2 and λ3 all represent weight coefficients.

7. The method for intelligently extracting bilingual information on ship accident risk influencing factors according to claim 6 is characterized in that: In step 5.2, the processing process of the self-attention mechanism on the input text sequence x is expressed as: Where Q, K, and V represent query, key, and value matrices, respectively. Q, K, and V are all obtained by linearly transforming the input vector representation h0. k is the dimension of the attention head.

8. The method for intelligently extracting bilingual information on ship accident risk influencing factors according to claim 6 is characterized in that: In step 5.2, the processing process of the feedforward neural network layer on the input text sequence x is expressed as: FFN(x)=ReLU(xW1+b1)W2+b2 Where W1 and W2 are linear transformation weights, and b1 and b2 are bias vectors.

9. The method for intelligently extracting bilingual information on ship accident risk influencing factors according to claim 1 is characterized in that: In step 5, the AdamW optimizer is used to adaptively fine-tune the parameters of the UIE-Multilingual model.

Images