An interpretable judicial case matching method and system based on the reverse optimal transport model

Abstract

The present invention implements an interpretable judicial case matching method based on the reverse optimal transport model. The method consists of three modules: the judicial element extraction module takes the input case pair, with two cases, the judicial element labels of each sentence in the cases, and the manually annotated sentence alignment labels as the module input, and outputs the prediction of the judicial element labels of all sentences in the cases; the candidate explanation generation module takes the judicial elements of the two cases extracted by the judicial element extraction module as the input and outputs the candidate explanations predicted by the model; the matching module takes the prediction result of the judicial element extraction module and the candidate explanations generated by the candidate explanation generation module as the input, and outputs the prediction of whether the cases match and the candidate explanations corresponding to the prediction as the final explanations. The method achieves the technical effect of providing explanations in two aspects for the prediction and improving the accuracy of the model prediction.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular, to an interpretable judicial case matching method based on an inverse optimal transport model. Background Art

[0002] Existing judicial case matching models can be mainly divided into three categories: models based on expert knowledge; models based on citation networks; and models based on text semantics.

[0003] The model based on expert knowledge is a traditional case matching method, which highly depends on the prior knowledge of the model designer in the judicial field. For example, [1] proposed that in the field of accident compensation, a case can be disassembled into several elements, such as the claimant supporter, the respondent supporter, and neutral factors, etc. A case can be represented as a set of these elements. Therefore, the matching of cases can be regarded as the matching between elements.

[0004] The models based on citation networks are mostly applied in common law countries. Such models construct a precedent citation network, where the nodes in the network are legal cases, and the directed edges represent the situation where the source case is cited by the target case, that is, the citation of the source case used by the target case. [2] proposed using the Jaccard similarity index between precedent citation sets to infer the similarity between two legal cases.

[0005] The models based on text semantics focus on the content of case texts and use the semantic similarity between cases to measure their matching degree. Thanks to the booming development of deep neural networks, the models based on text semantics have achieved performance beyond the first two types of models. For example, [3] proposed to split judicial cases into multiple paragraphs, encode each text paragraph using a large-scale pre-trained language model, then use max pooling and recurrent neural networks to model the semantic similarity of each paragraph pair in the two cases, and finally use the attention mechanism to output the matching score of the positive example of the case.

[0006] The existing similar judicial case matching method based on triple deep hashing learning obtains the judicial case documents to be matched, inputs the judicial case documents to be matched into a pre-trained feature extraction model to obtain the feature representation vectors of the judicial case documents to be matched, then inputs the feature representation vectors of the judicial case documents to be matched into a pre-trained triple deep hashing learning model at the same time to obtain the hash codes of the judicial case documents to be matched, and finally calculates the similarity of the judicial case documents based on the hash codes of the judicial case documents to be matched and the hash codes of the known judicial case documents.

[0007] Existing case matching models have achieved excellent performance in matching accuracy, but it is difficult for these models to provide explanations for their matching predictions. There are three reasons for this phenomenon: First, the content of judicial case documents is complex, and only some sentences reflecting judicial elements are relevant to the matching results and matching explanations. Existing models do not distinguish judicial elements from noise at the sentence level; Second, the judicial elements in cases can be divided into two categories: supporting matching and supporting non-matching. The explanation finally predicted by the model should balance these two categories of judicial elements, but existing models do not make a distinction; Third, the length of judicial case documents is long, and there are problems such as sparsity and bias in the annotation of judicial elements. Models that directly fit the labels of judicial elements for learning are prone to falling into local optima, restricting the generalization of the model. Summary of the Invention

[0008] For this reason, the present invention first proposes an interpretable judicial case matching method based on the reverse optimal transport model, which consists of three modules:

[0009] The judicial element extraction module aims at the input case pair, with two cases (X, Y), the judicial element label r of each sentence in the case X , r Y and the manually annotated sentence alignment label as the module input, and adopts the reverse optimal transport process. The model learns the "transport cost" between case sentences according to and outputs the prediction of the judicial element label of all sentences in the case where respectively represent the matching judicial elements and non-matching judicial elements in cases X and Y,

[0010] The candidate explanation generation module takes the judicial elements of the two cases extracted by the judicial element extraction module as the input, and takes the candidate explanation predicted by the model as the output, where i represents the number of candidate explanations. The candidate explanation generation module outputs as the explanation for label 0 and the explanation for label 1 respectively;

[0011] The matching module takes the predicted by the judicial element extraction module and the candidate explanations generated by the candidate explanation generation module as the input, regards as the query, and the candidate explanation as the document, calculates the matching scores of respectively, and outputs the explanation with the higher score and the matching label corresponding to this explanation, that is, the prediction of whether the output case matches and the candidate explanation corresponding to this prediction as the final explanation.

[0012] In the optimal transportation process, in the scenario of judicial case matching, any two cases μ and v are respectively regarded as a uniform distribution, that is where 1 D denotes a D-dimensional vector of all 1s, and the optimal transportation plan between two cases is:

[0013]

[0014] where denotes any joint distribution, and the sentences therein satisfy the marginal distributions of μ and v; is the cost matrix, and c(x m , y n ) represents the difference between two sentences;

[0015] After that, inverse optimal transportation is adopted to try to learn a more robust cost matrix C from : Given the optimal transportation plan, that is, manual annotation, solve the cost matrix, that is, the difference situation of the sentences in the case:

[0016]

[0017] s.t. A * (C) = argmin A∈Π(μ，ν) <A, C> + γ<A, logA>

[0018] where KL represents the KL divergence between the sentence alignment label and the transportation plan A * Finally, model C as a linear combination of two sub-matrices:

[0019] C = C s + ∈C r

[0020] where C s is the semantic similarity matrix at the sentence level of two cases, and the sentence vector similarity is calculated using a neural network; C r is the element similarity matrix at the sentence level of two cases. Use a neural network to classify the sentences in the case, distinguish judicial elements and non-judicial elements, and then compare whether two sentences belong to the same type of judicial elements; ∈ is the balance coefficient;

[0021] During the training process, the judicial element extraction module learns the parameters in the cost matrix C; during the testing process, the judicial element extraction module performs forward optimal transportation according to C, determines the alignment situation of the sentences in the case according to the optimal transportation plan, and extracts the judicial element label prediction

[0022] The judicial element extraction module minimizes the loss between the model prediction and the sentence-level alignment degree of the manual annotation and the loss between the classification prediction of each sentence's judicial element by the model and the manual annotation when constructing the cost matrix C r The loss function can be written as: can be written as:

[0023]

[0024] where is the cross - entropy loss between the predicted label of each sentence's judicial element in the prediction case and the true label by the model when constructing the cost matrix C:

[0025]

[0026] where δ(r, k)=1 if r = k, otherwise 0; is the reverse optimal transport loss of the model, optimizing the cost matrix C and the corresponding optimal transport plan matrix:

[0027]

[0028] The candidate explanation generation module minimizes the difference between the natural language explanations generated by the model and the manually annotated ones, and the loss function is the loss for fine - tuning the pre - trained language model:

[0029]

[0030] The matching module minimizes the difference between the case matching labels predicted by the model and the manually annotated ones The loss function can be written as:

[0031]

[0032] where is the cross - entropy loss between the prediction of the model for the case matching relationship and the true label:

[0033]

[0034] The training process is specifically as follows: Based on the training dataset batch sizes n1, n2, n3, balance coefficients ∈, γ, learning rates η1, η2, η3 data, design the algorithm process:

[0035] For the training of the judicial element extraction module, repeat the following process: Sample mini - batch data from the training dataset Sample mini - batch data from it Predict the judicial element labels of (X i , Y i ) Calculate Structure C, C s , C r , optimize A using the Sinkhorn algorithm * , calculate Calculate Gradient update until the result converges and return

[0036] For the training of the candidate explanation generation module, repeat the following process: sample mini-batch data from the training dataset in Fine-tune three label-specific pre-trained language models until the result converges and return

[0037] For the training of the matching prediction module, repeat the following process: sample mini-batch data from the training dataset in Use and to predict the matching label Calculate Calculate Gradient update until the result converges and return

[0038] The technical effect to be achieved by the present invention is as follows:

[0039] First, extract the judicial elements (sentences) that support matching and non-matching from the cases; secondly, generate natural language-form, specific label explanations according to the extracted judicial elements, corresponding to the explanations for the support matching label and the support non-matching label; finally, output the matching prediction of the case pair and the corresponding explanation based on the judicial elements and the natural language explanations. The present invention provides explanations in two aspects for the model prediction and improves the accuracy of the model prediction. Brief Description of the Drawings

[0040] Figure 1 Architecture diagram of the interpretable case matching method based on reverse optimal transport; Detailed Embodiments

[0041] The following are the preferred embodiments of the present invention and in combination with the drawings, the technical solutions of the present invention are further described, but the present invention is not limited to this embodiment.

[0042] The present invention proposes an interpretable judicial case matching method based on a reverse optimal transport model.

[0043] This method mainly consists of three modules, namely, the judicial element extraction module, the candidate explanation generation module, and the matching module.

[0044] The judicial element extraction module is responsible for extracting the judicial elements (sentences) that support matching and non - matching from the input pairs of judicial cases; the candidate explanation generation module generates candidate explanations based on the extracted judicial elements, that is, generates natural - language - form explanations for each potential matching label; the matching module calculates the final matching label of the case based on the extracted judicial elements and the generated candidate explanations and outputs the corresponding candidate explanations as the final explanations for matching prediction.

[0045] The judicial element extraction module takes two cases (X, Y), the judicial element labels r of each sentence in the cases X , r Y and the manually - annotated sentence alignment labels as the module input, and outputs the prediction of the judicial element labels of all sentences in the cases ( representing the matching and non - matching judicial elements in cases X and Y respectively). Considering that judicial cases are long and complex in content, in actual annotation is often extremely sparse, and there is a high probability of false - negative problems. If the model directly fits it will lead to problems such as poor robustness and weak generalization ability. Therefore, in this method, judicial element extraction is modeled as an inverse optimal transport process, and the model learns the "transportation cost" between case sentences in order to improve the robustness of the model.

[0046] The optimal transport process defines the distance between two probability distribution distances. In the scenario of judicial case matching, the two cases can be regarded as a uniform distribution respectively, that is where 1 D represents a D - dimensional vector of all 1s. Then the optimal transport plan between the two cases is:

[0047]

[0048] where represents any joint distribution, and the sentences in it satisfy the marginal distributions of μ and v; is the cost matrix, and its element c(x m , y n ) represents the difference between two sentences. Therefore, the optimal transport can be regarded as the minimum expected value of the differences between the sentences of the two cases.

[0049] In the optimal transportation process, the cost matrix is often predefined, such as the cosine similarity of two sentence vectors, which measures the semantic similarity between two sentences. However, in the judicial case matching scenario, the basis for whether two sentences are aligned goes beyond the scope of semantic similarity. The model not only needs to consider semantics, but also whether the sentences embody certain judicial elements, such as case facts, essential facts, or disputed issues. At the same time, considering that there may be false negative problems, this method uses reverse optimal transportation to try to learn a more robust cost matrix C from .

[0050] Different from solving the transportation plan A * for the given cost matrix C, the reverse optimal transportation is more in line with the case matching scenario: given the optimal transportation plan (manually annotated), solve the cost matrix (the difference situation of sentences in the case):

[0051]

[0052] s.t. A * (C) = argmin A∈Π(μ，ν) <A, C> + γ<A, logA>

[0053] where KL represents the KL divergence between the sentence alignment label and the transportation plan A * , and here the transportation plan A * is calculated through the optimal transportation process according to the learnable cost matrix C. This method models C as a linear combination of two sub-matrices:

[0054] C = C s + ∈C r

[0055] where C s is the semantic similarity matrix at the sentence level of two cases, and the sentence vector similarity can be calculated using a neural network; C r is the element similarity matrix at the sentence level of two cases. A neural network can be used to classify the sentences in the case, distinguish judicial elements from non-judicial elements, and then compare whether the two sentences belong to the same type of judicial elements; ∈ is the balance coefficient. During the training process, the judicial element extraction module attempts to learn the parameters in the cost matrix C; during the testing process, the judicial element extraction module performs forward optimal transportation according to C, determines the alignment of sentences in the case according to the optimal transportation plan, and extracts the judicial element label prediction

[0056] The candidate explanation generation module takes the judicial elements of the two cases extracted by the judicial element extraction module as input, and takes the candidate explanations predicted by the model As the output, where \(i\) represents the number of candidate explanations. For example, in a binary classification problem, the candidate explanation generation module outputs as the explanations for label 0 and label 1 respectively. For the sake of convenience of expression, it will be denoted as Since is a set of several sentences in the original case. Therefore, in the candidate explanation generation module, this method fine-tunes the pre-trained language model to generate more compact and fluent explanations.

[0057] The matching module takes the predicted by the judicial element extraction module and the candidate explanations generated by the candidate explanation generation module as inputs, and outputs the prediction of whether the case matches and the candidate explanation corresponding to this prediction as the final explanation. This method models the matching module as a retrieval and ranking problem, that is, is regarded as the query, and the candidate explanation is regarded as the document, and calculates the and matching scores respectively, and outputs the explanation with a higher score and the matching label corresponding to this explanation.

[0058] This method includes three sub-modules, and each module includes its own training objective: the judicial element extraction module minimizes the loss between the model prediction and the sentence-level alignment degree of the manual annotation and the loss between the judicial element classification prediction of each sentence by the model and the manual annotation when constructing \(C\) r ; The candidate explanation generation module minimizes the difference between the model-generated and the manually annotated natural language explanations; the matching module minimizes the difference between the model prediction and the manually annotated case matching labels

[0059] The loss function of the judicial element extraction module can be written as:

[0060]

[0061] where is the cross-entropy loss between the judicial element labels of each sentence in the predicted case and the true labels by the model when constructing the cost matrix \(C\):

[0062]

[0063] where \(\delta(r,k)=1\) if \(r = k\) and 0 otherwise; is the model's reverse optimal transport loss, optimizing the cost matrix \(C\) and the corresponding optimal transport plan matrix:

[0064]

[0065] Loss function of the candidate explanation generation module is the loss for fine-tuning the pre-trained language model:

[0066]

[0067] Loss function of the matching module can be written as:

[0068]

[0069] where is the cross-entropy loss between the model's prediction of the case matching relationship and the true label:

[0070]

[0071] Training process of this method:

[0072] Requirements: Training dataset Batch sizes n1, n2, n3; balancing coefficients ∈, γ; learning rates η1, η2, η3.

[0073] 1 Judicial element extraction

[0074] 2: repeat:

[0075] 3: Sample a mini-batch of data from the training dataset

[0076] 4: Predict the judicial element labels of (X i , Y i )

[0077] 5: Calculate

[0078] 6: Construct C, C s , C r , and optimize A using the Sinkhorn algorithm *

[0079] 7: Calculate

[0080] 8: Calculate

[0081] 9: Gradient update

[0082] 10: until convergence

[0083] 11: return ​

[0084] 13: Candidate Explanation Generation

[0085] 14: repeat:

[0086] 15: Sample mini-batch data from the training dataset

[0087] 16: Fine-tune three label-specific pre-trained language models

[0088] 17: until convergence

[0089] 18: return

[0090] 19: Match Prediction

[0091] 20: repeat:

[0092] 21: Sample mini-batch data from the training dataset

[0093] 22: Use and to predict matching labels

[0094] 23: Calculate

[0095] 24: Calculate

[0096] 25: Gradient update

[0097] 26: until convergence

[0098] 27: return

[0099] In summary, the interpretable case matching method based on reverse optimal transport takes the training set and some hyperparameters as inputs. The three modules calculate the loss in sequence, update the model parameters using the gradient descent method, train for multiple rounds until convergence, and output the parameters of the judicial element extraction module the parameters of the candidate explanation generation module and the parameters of the match prediction module

[0100] This technical solution is a model framework, and different configurations of the judicial element extraction module, candidate explanation generation module, and match prediction module will have different specific implementations. This section introduces a sample implementation applied to this model.​​

[0101] 1 Judicial element extraction module

[0102] Given two input judicial cases X and Y, this example first uses the BERT model [4] pre-trained in a large-scale judicial case corpus to map all sentences in the input cases into fixed-length sentence embedding vectors. Specifically, the output of the BERT model on the "[CLS]" character is used as the embedding of the entire sentence, denoted as and where x m , y n represent the sentence embeddings of the m-th sentence of case X and the n-th sentence of case Y, respectively. Secondly, this example projects the sentence embeddings through a feed-forward neural network and calculates the Euclidean distance between x m and y n to construct a semantic similarity matrix at the sentence level of the two cases Classify the judicial elements of X and Y through a gated dilated convolutional neural network [5], and construct a similarity matrix C at the sentence level of the elements of the cases according to whether the two sentences belong to the same judicial element r ∈ {0, 1} M×N , C s and C r The linear combination of is used as the transportation cost of optimal transportation. Finally, this example uses the Sinkhorn [6] algorithm to solve the global alignment scheme of X and Y at the sentence level according to the transportation cost, denoted as where this example sets a threshold to take the sentences corresponding to the elements greater than the threshold in the alignment scheme as the sentences supporting the matching decision, otherwise, they are the sentences not supporting the matching decision.

[0103] 2 Candidate explanation generation module

[0104] Given two input judicial cases X, Y, the prediction of sentence-level judicial element labels and the manually annotated explanation e, this example first divides the data into several parts according to the matching label z (assuming the task is binary classification, so there are two types of labels, and the data is divided into two parts. The following takes binary classification as an example), and performs the following operations on each part of the data: In the training stage, use and the corresponding sentences as input, e as the label, and fine-tune the T5-PEGASUS [7] model pre-trained in a large-scale Chinese corpus to make the model learn to generate judicial interpretations guided by manual annotations. In the test stage, two trained generation models are used to generate the corresponding explanations

[0105] 3. Matching module

[0106] The matching module of this example uses the sentences representing judicial elements and candidate explanations As input, output the prediction of case matching relationships and the corresponding labels. Similar to the judicial element extraction module, the matching module first uses the BERT model pre-trained in a large-scale judicial case corpus to map all sentences in the input case into fixed-length sentence embedding vectors. Specifically, the output of the BERT model on the "[CLS]" character is used as the embedding of the entire sentence. Secondly, in this example, the judicial elements and candidate explanations of the two cases are respectively concatenated ( where [·] represents the vector concatenation operation), and a two-layer feed-forward neural network is used to calculate the similarity score between them. The explanation with a high score and its corresponding matching label are used as the final output of the model.

Claims

1. An interpretable judicial case matching method based on the reverse optimal transportation model, characterized in that: It consists of three modules: The judicial element extraction module aims at the input case pair, with two cases (X, Y), the judicial element labels r of each sentence in the cases X , r Y and the sentence alignment labels manually marked as the module input, and adopts the reverse optimal transportation process. The model learns the "transportation cost" between the case sentences and outputs the prediction of the judicial element labels of all sentences in the cases where respectively represent the matching and non-matching judicial elements in cases X and Y The candidate explanation generation module uses the judicial elements of the two cases extracted by the judicial element extraction module as input, and the candidate explanations predicted by the model as output, where i represents the number of candidate explanations, and the candidate explanation generation module outputs as the explanation for label 0 and the explanation for label 1 respectively; The matching module uses the and the candidate explanations generated by the candidate explanation generation module as inputs, and regards them as queries, and the candidate explanations as documents, and calculates the and matching scores respectively, and outputs the explanation with the higher score and the matching label corresponding to this explanation. If has a higher score, the model outputs that the two cases match, and the candidate explanation as the reason for the case match If has a higher score, the model outputs that the two cases do not match, and the candidate reason as the reason for the case non - match That is, it outputs the prediction of whether the cases match or not and the candidate explanation corresponding to this prediction as the final explanation.

2. The interpretable judicial case matching method based on the reverse optimal transportation model according to claim 1, wherein: In the optimal transportation process, in the scenario of judicial case matching, any two cases μ and v are respectively regarded as a uniform distribution, that is where 1 D represents a D-dimensional vector of all 1s, and the optimal transportation plan between the two cases is as follows: Among them represents an arbitrary joint distribution, where the sentences satisfy marginal distributions of μ and ν; is the cost matrix, and c(x m , y n ) represents the difference between two sentences; After that, reverse optimal transport is used to try to learn a more robust cost matrix C from : Given the optimal transport plan, i.e., the manual annotation, solve for the cost matrix, i.e., the difference situation of the sentences in the case: s.t.A * (C) = agrmin A∈Π(μ,v) <A, C> + γ<A, logA> where KL represents the sentence alignment label and transportation plan A * The KL divergence between them. Finally, C is modeled as a linear combination of two submatrices: C=C s +∈C r Among them, C s is the semantic similarity matrix at the sentence level of two cases, calculating the sentence vector similarity using a neural network; C r is the element similarity matrix at the sentence level of two cases, classifying the sentences in the cases using a neural network to distinguish judicial elements and non-judicial elements, and then comparing whether the two sentences belong to the same type of judicial elements; ∈ is the balance coefficient; During the training process, the judicial element extraction module learns the parameters in the cost matrix C; during the testing process, the judicial element extraction module performs forward optimal transport according to C, determines the alignment of sentences in the case based on the optimal transport plan, and extracts the prediction of the judicial element label 3. The interpretable judicial case matching method based on the reverse optimal transportation model according to claim 2, wherein: The judicial element extraction module serves as minimizing the loss between the prediction of the minimized model and the sentence-level alignment of the manually annotated ones and the loss between the prediction of the judicial elements of each sentence by the model and the manual annotation when constructing C r The loss function can be written as: can be written as: Among them, is the cross-entropy loss between the judicial element labels and the true labels of each sentence in the prediction case when the model constructs the cost matrix C: where if r = k, then δ(r,k) = 1, otherwise δ(r,k) = 0; is the model reverse optimal transport loss, optimizing the cost matrix C and the corresponding optimal transport plan matrix:

4. The interpretable judicial case matching method based on the reverse optimal transportation model according to claim 3, characterized in that: The candidate explanation generation module minimizes the difference between the natural language explanations generated by the model and the manually annotated ones, and the loss function is the loss for fine-tuning the pre-trained language model:

5. The interpretable judicial case matching method based on the reverse optimal transportation model according to claim 4, characterized in that: The matching module minimizes the difference between the case matching labels predicted by the model and the manually annotated ones. Loss function It can be written as: Among them is the cross-entropy loss between the model's prediction of the case matching relationship and the true label, represents the training data set:

6. The interpretable judicial case matching method based on the reverse optimal transportation model according to claim 5, characterized in that: The specific training process is as follows: Based on the training dataset Batch sizes n1, n2, n3, balance coefficients ∈, γ, learning rates η1, η2, η3 data, design the algorithm process: For the training of the judicial element extraction module, repeat the following process: Sample a small batch of data from the training dataset and predict the judicial element labels of (X , Y i , Y i ) Calculate Construct C, C s , C r , and optimize A using the Sinkhorn algorithm * , calculate Calculate Update the gradient until the result converges and return The training of the candidate explanation generation module repeats the following process: sampling mini-batch data from the training dataset and fine-tuning three label-specific pre-trained language models until the results converge and return The training of the matching prediction module repeats the following process: Sampling mini-batch data from the training dataset and using and to predict the matching labels Calculating Calculating Gradient update until the result converges and returns 7. An interpretable judicial case matching system based on the reverse optimal transportation model, characterized in that: An input module that includes reading cases, annotating the judicial elements of each sentence in the case, and the sentence alignment labels manually annotated, a calculation module that applies an interpretable judicial case matching method based on the reverse optimal transport model as described in any one of claims 1-6, and an output module that outputs a prediction conclusion on whether the cases match and the candidate explanations corresponding to the prediction.

Images