CN114881028A - Case similarity matching method, device, computer equipment and storage medium - Google Patents

Abstract

The embodiment of the present invention discloses a case similarity matching method, device, computer equipment and storage medium. The method includes: obtaining a case judgment text in a case database; collecting stop words and special noun words from the case judgment text, and generating Stop word vocabulary database and proprietary word vocabulary database; select the first case text and the second case text that need similarity matching from the case judgment text; input the first case text and the second case text into the twin network for processing to obtain the similarity probability value of the first case text and the second case text; if the similarity probability value of the first case text and the second case text meets the set similarity threshold, then determine the first case text and the second case text for similar cases. The present invention improves the effectiveness and accuracy of case similarity matching.

Description

Case similarity matching method, device, computer equipment and storage medium

technical field

The present invention relates to the technical field of data retrieval, and more particularly to a case similarity matching method, device, computer equipment and storage medium.

Background technique

With the development of the times, the number of court trials has increased, and there are more and more case judgments in the case database. When analyzing a case, it is often necessary to find two or more similar cases from the case database for comparison. There are several ways to search.

The first one is to retrieve the cases in the case database, extract the attribute elements of people, vehicles and objects in each case object model and add them to the corresponding comparison container of the people and vehicles array, and calculate the people and vehicles to be compared. The similarity of the attributes of people, vehicles and objects in the object array, and the at least two attribute element objects with the largest similarity and the corresponding similarity values are recorded in the similarity map in the form of key-value pairs, and finally based on the similarity map. Sort and display the similarity of the attribute elements of people, vehicles and objects in the case object model of each case. This method performs similarity matching of articles based on the key-value pairs of attributes, ignoring a large amount of information on non-human vehicles and objects, serious loss of information on non-human vehicles or cases involving other fields, and has a serious impact on the effect of case matching.

The second is to include three-segment blocks that use the layout of the document and key words as constraints, and use the automatic extraction algorithm to extract the case facts, dispute focus, and judgment results of the document; based on the domain vocabulary, use the topic model to extract each document segment The subject headings of the blocks are obtained from the subject headings and non-subject headings of each document segment block, and the feature inverted index is constructed according to the subject headings and the feature words in the non-subject headings of each document segment block. The feature inverted index is mapped to the feature vector, and the similarity between the query sentence and each document in the document data set is calculated by the topic similarity model, the similarity between the query sentence and each document in the document data set is sorted, and the sorting result is output to complete the document retrieval. This method only judges the similarity based on the subject words, ignoring the content of the text itself, and it is difficult to subdivide and distinguish cases with similar subjects.

The third method is to use word2vec word vector for the legal text, express the keywords with word vectors, and use the cosine similarity to calculate the similarity between different cases. After obtaining a variety of cases related to the case, based on the keyword extraction technology to find out the judgment result, intelligently give a reasonable judgment result range for this case, and timely carry out when the actual judgment result is too different from the recommended judgment range. Smart alerts. This method is based on the word2vec technology to generate vectors for similarity comparison, the representation of words trained by word2vec, this method cannot represent the same words in different contexts and contexts, especially for professional vocabulary, it cannot be effectively expressed, so the matching effect of shape similarity Not good, not very practical.

SUMMARY OF THE INVENTION

The purpose of the present invention is to overcome the deficiencies of the prior art, and to provide a case similarity matching method, device, computer equipment and storage medium, which can improve the effectiveness and accuracy of the case similarity matching.

To achieve the above object, the present invention adopts the following technical solutions:

In the first aspect, the case similarity matching method includes:

Obtain the text of the case judgment in the case database;

Collect stop words and proper noun words from the text of case judgments, and generate stop word word database and proper word word database;

Select the first case text and the second case text that need similarity matching from the case judgment text;

Input the first case text and the second case text into the twin network for processing to obtain the similarity probability value of the first case text and the second case text;

If the similarity probability value of the first case text and the second case text meets the set similarity threshold, it is determined that the first case text and the second case text are similar cases.

Its further technical scheme is: the first case text and the second case text are input into the twin network for processing, so as to obtain similar probability values of the first case text and the second case text, and the twin network includes an ERNIE-based The network model of text vector, the network model of text vector based on WordGCN graph, and the network model of text vector based on subject word.

Its further technical solution is: the described inputting the first case text and the second case text into the twin network for processing, so as to obtain the similarity probability value of the first case text and the second case text, including:

Inputting the first case text and the second case text into the ERNIE-based text vector network model for processing to obtain the first processing features of the first case text and the second case text;

Inputting the first case text and the second case text into the network model based on the text vector of the WordGCN graph for processing to obtain the second processing features of the first case text and the second case text;

Inputting the first case text and the second case text into the network model based on the text vector of the subject word for processing, so as to obtain the third processing feature of the first case text and the second case text;

Perform concate merge processing on the first processing feature of the first case text and the second case text and the second processing feature of the first case text and the second case text to obtain the merged feature of the first case text and the second case text;

inputting the merged features of the first case text and the second case text into the fully connected layer processing to obtain the fully connected layer processing features of the first case text and the second case text;

Multiply the fully connected layer processing features of the first case text and the second case text with the third processing feature of the first case text and the second case text to obtain the text semantic representation of the first case text and the second case text feature;

Perform full connection layer and activation function processing on the text semantic representation features of the first case text and the second case text to obtain the text abstract semantic representation of the first case text and the second case text;

The text abstract semantic representation of the first case text and the second case text is processed through a matrix of a fully connected layer with a dimension of 1 and a sigmoid activation function to obtain the similarity probability value of the first case text and the second case text.

Its further technical scheme is: the first case text and the second case text are input into the ERNIE-based text vector network model for processing, so as to obtain the first processing features of the first case text and the second case text, including: :

Perform sentence segmentation according to the sentence segmentation symbols of the text content in the first case text and the second case text;

Tokenize the sentence by using the word segmentation tool combined with the stop word lexicon and the proprietary word lexicon to obtain word segmentation data;

The word segmentation data is processed based on MLM through ERNIE to obtain the word vector of each word;

The word vector of each word in each sentence is summed to obtain the feature vector of the sentence vector;

The feature vectors of all sentence vectors of the text content are concatenated and fused by Bi-LSTM to obtain the first processing features of the first case text and the second case text.

Its further technical scheme is: the first case text and the second case text are input into the network model based on the text vector of the WordGCN graph for processing, so as to obtain the second processing feature of the first case text and the second case text, including: :

Encode the words of the first case text and the second case text through the relationship between words in the sentence level and corpus level in the WordGCN model to obtain word vectors;

Construct sentence vector according to word vector;

Input the sentence vector into Bi-GRU for processing to obtain the second processing feature of the first case text and the second case text.

Its further technical solution is: the first case text and the second case text are input into the network model based on the subject word text vector for processing, so as to obtain the third processing feature of the first case text and the second case text, including :

Filter the stop words in the text of the first case text and the second case text through the stop word vocabulary database;

Extract the subject words of the filtered text;

Record the position index and importance degree corresponding to the extracted subject words;

Extract proper nouns in the texts of the first case text and the second case text through the proper word vocabulary database;

Record the position index and importance degree corresponding to the extracted proper noun;

The degree of importance of the subject word and the degree of importance of proper nouns are added up to obtain the third processing feature.

The further technical solution is as follows: the subject words of the filtered text are extracted, and the subject words of the text are extracted by combining the BERTopic model with the LDA model.

In a second aspect, a case similarity matching device includes an acquisition unit, a generation unit, a selection unit, a processing unit, and a determination unit;

The obtaining unit is used to obtain the case judgment text in the case database;

The generating unit is used to collect stop words and special noun words from the text of the case judgment, and generate stop word word database and proper word word database;

The selection unit is used to select the first case text and the second case text that need to be matched for similarity from the case judgment text;

The processing unit is configured to input the first case text and the second case text into the twin network for processing, so as to obtain the similarity probability value of the first case text and the second case text;

The determining unit is configured to determine that the first case text and the second case text are similar cases if the similarity probability value of the first case text and the second case text meets the set similarity threshold.

In a third aspect, a computer device includes a memory, a processor, and a computer program stored on the memory and executable on the processor, the processor executing the computer program to achieve a similar case as above Degree matching method steps.

In a fourth aspect, a computer-readable storage medium stores a computer program, and the computer program includes program instructions that, when executed by a processor, cause the processor to execute similar cases as described above. Degree matching method steps.

Compared with the prior art, the present invention has the following beneficial effects: by adopting the ERNIE technology, the present invention can understand the meaning of the polysemy in the text by combining the context well, perform case similarity analysis based on the twin neural network, and combine the attention The mechanism analyzes key features of terms related to the specificity of the case, which increases the model's understanding of the content of the case and improves the effectiveness and accuracy of case similarity matching.

The above description is only an overview of the technical solution of the present invention. In order to understand the technical means of the present invention more clearly, it can be implemented in accordance with the content of the description, and in order to make the above and other purposes, features and advantages of the present invention more obvious and understandable, the following A preferred embodiment is given, and the detailed description is as follows.

Description of drawings

In order to explain the technical solutions of the embodiments of the present invention more clearly, the following briefly introduces the accompanying drawings used in the description of the embodiments. Obviously, the drawings in the following description are some embodiments of the present invention, which are of great significance to the art For those of ordinary skill, other drawings can also be obtained from these drawings without any creative effort.

1 is a flowchart of a case similarity matching method provided by a specific embodiment of the present invention;

2 is a schematic block diagram of a case similarity matching apparatus provided by a specific embodiment of the present invention;

FIG. 3 is a schematic block diagram of a computer device according to a specific embodiment of the present invention.

Detailed ways

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

FIG. 1 is a flowchart of a case similarity matching method provided by a specific embodiment of the present invention.

As shown in Figure 1, the case similarity matching method includes the following steps: S10-S50.

S10. Obtain the case judgment text in the case database.

In this embodiment, the case database refers to a database used for court case judgments.

S20. Collect stop words and proper noun words from the text of the case judgment, and generate a stop word word database and a proper word word database.

In this embodiment, due to factors such as revision of laws and regulations, some words will be disabled. In addition, with the development of technology, the following new special terms will appear. Therefore, stop words and proper noun words can be collected from the text of the case judgment, and a stop word word database and a proper word word database can be established and generated for them, so as to facilitate the subsequent text similarity analysis.

S30: Select the first case text and the second case text that need to be matched for similarity from the case judgment text.

Before selection, selection conditions can be set, for example, the similarity matching text is selected according to the content related to the description of the case and the conclusion of the case.

Of course, it should be noted that the first case text and the second case text not only refer to the similarity matching between a first case text and a second case text, but also refer to a first case text and multiple The second case text is subjected to similarity matching. If a first case text and a second case text are matched for similarity, it can be concluded whether the two are similar; if a first case text is matched with multiple second case texts, it can be obtained. The degree of similarity between the multiple second case texts and the first case texts can be sorted according to the degree of similarity, so as to facilitate case analysis by personnel.

S40. Input the first case text and the second case text into the twin network for processing, so as to obtain the similarity probability value of the first case text and the second case text.

In this embodiment, the twin network is divided into two identical and parallel left and right sub-networks, and the left and right sub-networks both include a network model based on ERNIE text vectors, a network model based on WordGCN graph text vectors, and text based on subject words. Vector network model. The first case text can be processed through the left sub-network, and the second case text can be processed through the right sub-network.

In this embodiment, step S40 specifically includes the following steps: S401-S408,.

S401. Input the first case text and the second case text into a network model based on ERNIE text vectors for processing, so as to obtain the first processing features of the first case text and the second case text.

The network model based on ERNIE's text vector is to train the ERNIE model for all cases to obtain the word vector of each word. The training method adopts the form of fine-tunning for word vector training.

In one embodiment, step S401 specifically includes the following steps: S4011-S4015.

S4011. Perform sentence segmentation according to the sentence segmentation symbols of the text content in the first case text and the second case text.

S4012. Perform word segmentation on the sentence by using a word segmentation tool in combination with a stop word vocabulary database and a proprietary word vocabulary database to obtain word segmentation data.

S4013 , process the word segmentation data based on MLM through ERNIE to obtain a word vector of each word.

S4014. Perform a summation operation on the word vector of each word in each sentence to obtain a feature vector of the sentence vector.

S4015. Concate and fuse the feature vectors of all sentence vectors of the text content through Bi-LSTM to obtain the first processing features of the first case text and the second case text.

如果词在为本中的位置为奇数的时候

position表示词在句子中的位置，i为词向量的维度范围，选用0至64维度。For steps S4011-S4015, in this embodiment, sentence segmentation is first performed based on the sentence break symbol in the text, and then the sentence is segmented by a word segmentation tool, and the sequence length in the sentence is selected to be 16 words long, and the text exceeding the word length is selected. The sequence is truncated, and padding is used to fill in the sequence shorter than the sequence length. At the same time, it should be noted that the analysis of proper nouns needs to be combined with the proper noun vocabulary for word segmentation to avoid internal segmentation of proper nouns. For the position encoding in the transformer mechanism in ERNIE, if the position of the word in the text is an even number, it is used.

if the word's position in the base is an odd number

position represents the position of the word in the sentence, i is the dimension range of the word vector, and the dimension is from 0 to 64.

The word vector of each word can be obtained by training the ERNIE model based on the MLM (Masked Language Model) method. There are multiple words in a sentence. The feature vector of the sentence vector is obtained by adding the multiple words, and all the sentence vectors are concatenated and fused to obtain the text vector. The length of the text vector is 500, that is, 500 sentences are selected, and the sentences exceeding this length are selected to be truncated, and the sentence vector embedding _{enrie-sentence} constructed based on the word vector with a length of 500 and a dimension of 64 can be obtained. Input the embedding _{enrie-sentence} into Bi-LSTM to get the text vector embedding _ernie-doc .

S402. Input the first case text and the second case text into the network model based on the text vector of the WordGCN graph for processing, so as to obtain the second processing features of the first case text and the second case text.

In one embodiment, step S402 specifically includes the following steps: S4021-S4023.

S4021. Encode the words of the first case text and the second case text through the relationship between words in the sentence level in the WordGCN model and in the corpus level to obtain word vectors.

S4022, construct a sentence vector according to the word vector.

S4023: Input the sentence vector into the Bi-GRU for processing to obtain the second processing feature of the first case text and the second case text.

For steps S4021-S4023, in this embodiment, the WordGCN model is selected to represent the sentence model. Words are encoded by word-to-word relationships at the sentence level and at the corpus level. Sentence encoding is performed by word segmentation and fixed sentence length. The sentence length is set to 16 words. If the length of the sentence exceeds the length of the sentence, it will be truncated. If the length of the sentence is not exceeded, it will be filled by pandding. The sentence vector is still obtained by adding the word vector. . Select 500 for the length of the text vector, and select truncation for sentences exceeding this length, so that the sentence vector embedding _words constructed based on the word vector with a length of 500 and a dimension of 64 can be obtained. Input the embedding _words into Bi-GRU to get the text vector embedding _graph-doc .

S403: Input the first case text and the second case text into the network model based on the text vector of the subject word for processing, so as to obtain the third processing feature of the first case text and the second case text.

In one embodiment, step S403 specifically includes the following steps: S4031-S4036.

S4031. Filter the stop words in the text of the first case text and the second case text through a stop word vocabulary database.

S4032. Extract the subject words of the filtered text.

S4033 , record the position index and importance degree corresponding to the extracted subject word.

S4034. Extract proper nouns in the texts of the first case text and the second case text through the proper word vocabulary database.

S4035, record the position index and importance degree corresponding to the extracted proper noun.

S4036 , adding the importance degree of the subject word and the importance degree of the proper noun to obtain a third processing feature.

For steps S4031-S4036, in this embodiment, firstly, the stop words in the text are filtered through the stop word vocabulary database, and the subject word model is used to extract the subject words of the case based on the records after the word segmentation, to obtain the subject heading. The subject word vocabulary is BERTopic, and c-tf-idf is used internally. The keyword extraction is based on the number of categories of the case instead of the number of documents. At the same time, the LDA model is used to extract the subject words of the case content. Based on the subject headings extracted by the two methods, record the position index corresponding to the subject heading and the importance of the vocabulary. , record the position index corresponding to its proper noun, and the importance of proper noun is uniformly set to 0.5. The special word position index is generated by the union of the subject word position index and the proper noun position index. Add the importance of these keywords to get the attention weight matrix embedding _attentation , the matrix size is 16x500.

S404. Perform concate processing on the first processing feature of the first case text and the second case text and the second processing feature of the first case text and the second case text to obtain a merger of the first case text and the second case text feature.

In this embodiment, the feature embedding _merge is obtained by concatenating the embedding _ernie-doc and the embedding _graph-doc .

S405: Input the merged features of the first case text and the second case text into the fully-connected layer for processing, so as to obtain the fully-connected layer processing features of the first case text and the second case text.

In this embodiment, the embedding _merge is input to the fully connected layer to obtain the feature embedding _merge-fc .

S406: Multiply the fully connected layer processing features of the first case text and the second case text and the third processing feature of the first case text and the second case text to obtain the texts of the first case text and the second case text Semantic representation features.

In this embodiment, the final text semantic representation embedding _doc is obtained by multiplying the embedding _merge-fc matrix and the attention matrix embedding _attentation .

S407. Perform full connection layer and activation function processing on the text semantic representation features of the first case text and the second case text to obtain textual abstract semantic representations of the first case text and the second case text.

In this embodiment, the semantic representation embedding _doc is connected to the fully connected layer and the activation function is connected to a fully connected layer to obtain the text abstract semantic representation embedding _abs-doc .

S408: Process the text abstract semantic representations of the first case text and the second case text through a matrix of a fully connected layer with a dimension of 1 and a sigmoid activation function to obtain similarity probability values of the first case text and the second case text.

In this embodiment, the text abstract semantic representation embedding _abs-doc obtained through the left and right sub-networks is connected to a matrix with a dimension of 1 in the fully connected layer and then connected to a sigmoid activation function to finally output a similarity probability value of 0 to 1.

S50. If the similarity probability value of the first case text and the second case text meets the set similarity threshold, determine that the first case text and the second case text are similar cases.

In this embodiment, the similarity threshold is 0.5. If the similarity probability value between the second case text and the first case text is greater than 0.5, it is considered that the case texts of the two cases are similar, and if it is less than 0.5, the case content of the two cases is considered dissimilar. .

By adopting the ERNIE technology, the invention can understand the meaning of the polysemy in the text well in combination with the context, analyze the similarity of the case based on the twin neural network, and combine the attention mechanism to analyze the key features of the nouns in the special field of the case. It increases the model's understanding of the case content, and improves the effectiveness and accuracy of case similarity matching.

FIG. 2 is a schematic block diagram of a case similarity matching apparatus 100 provided by a specific embodiment of the present invention. Corresponding to the above-mentioned case similarity matching method, a specific embodiment of the present invention further provides a case similarity matching apparatus 100 . The case similarity matching apparatus 100 includes a unit for executing the above case similarity matching method, and the apparatus may be configured in a server.

As shown in FIG. 2 , the case similarity matching apparatus 100 includes an acquisition unit 110 , a generation unit 120 , a selection unit 130 , a processing unit 140 and a determination unit 150 .

The obtaining unit 110 is configured to obtain the case judgment text in the case database.

In this embodiment, the case database refers to a database used for court case judgments.

The generating unit 120 is configured to collect stop words and proper noun words from the text of the case judgment, and generate a stop word word database and a proper word word database.

In this embodiment, due to factors such as revision of laws and regulations, some words will be disabled. In addition, with the development of technology, the following new special terms will appear. Therefore, stop words and proper noun words can be collected from the text of the case judgment, and a stop word word database and a proper word word database can be established and generated for them, so as to facilitate the subsequent text similarity analysis.

The selecting unit 130 is configured to select the first case text and the second case text that need to be matched for similarity from the case judgment text.

Before selection, selection conditions can be set, for example, the similarity matching text is selected according to the content related to the description of the case and the conclusion of the case.

Of course, it should be noted that the first case text and the second case text not only refer to the similarity matching between a first case text and a second case text, but also refer to a first case text and multiple The second case text is subjected to similarity matching. If a first case text and a second case text are matched for similarity, it can be concluded whether the two are similar; if a first case text is matched with multiple second case texts, it can be obtained. The degree of similarity between the multiple second case texts and the first case texts can be sorted according to the degree of similarity, so as to facilitate case analysis by personnel.

The processing unit 140 is configured to input the first case text and the second case text into the twin network for processing, so as to obtain the similarity probability value of the first case text and the second case text.

In this embodiment, the twin network is divided into two identical and parallel left and right sub-networks, and the left and right sub-networks both include a network model based on ERNIE text vectors, a network model based on WordGCN graph text vectors, and text based on subject words. Vector network model. The first case text can be processed through the left sub-network, and the second case text can be processed through the right sub-network.

In one embodiment, the processing unit 140 includes a first processing module, a second processing module, a third processing module, a combining module, a fourth processing module, an arithmetic module, a fifth processing module, and a sixth processing module.

The first processing module is used for inputting the first case text and the second case text into the ERNIE-based text vector network model for processing, so as to obtain the first processing features of the first case text and the second case text.

The network model based on ERNIE's text vector is to train the ERNIE model for all cases to obtain the word vector of each word. The training method adopts the form of fine-tunning for word vector training.

In one embodiment, the first processing module includes a segmentation sub-module, a word segmentation sub-module, a first processing sub-module, an operation sub-module and a fusion sub-module.

The subsection module is used for sentence segmentation according to the sentence segmentation symbols of the text content in the first case text and the second case text.

The word segmentation sub-module is used to segment the sentence by using the word segmentation tool combined with the stop word vocabulary database and the proprietary word vocabulary database to obtain word segmentation data.

The first processing sub-module is used to process the word segmentation data based on MLM through ERNIE to obtain the word vector of each word.

The operation submodule is used for summing the word vectors of each word in each sentence to obtain the feature vector of the sentence vector.

The fusion sub-module is used to concate and fuse the feature vectors of all sentence vectors of the text content through Bi-LSTM to obtain the first processing features of the first case text and the second case text.

如果词在为本中的位置为奇数的时候

position表示词在句子中的位置，i为词向量的维度范围，选用0至64维度。For the segmentation sub-module, the word segmentation sub-module, the processing sub-module, the operation sub-module and the fusion sub-module, in this embodiment, sentence segmentation is first performed based on the sentence segmentation symbols in the text. The length of the sequence of 16 words is selected as the length of 16 words. The text sequence exceeding the length of the word is truncated, and the text sequence shorter than the length of the sequence is filled by padding. At the same time, it should be noted that the analysis of proper nouns needs to be combined with the proper noun vocabulary for word segmentation to avoid internal segmentation of proper nouns. For the position encoding in the transformer mechanism in ERNIE, if the position of the word in the text is an even number, it is used.

if the word's position in the base is an odd number

position represents the position of the word in the sentence, i is the dimension range of the word vector, and the dimension is from 0 to 64.

The word vector of each word can be obtained by training the ERNIE model based on the MLM (Masked Language Model) method. There are multiple words in a sentence. The feature vector of the sentence vector is obtained by adding the multiple words, and all the sentence vectors are concatenated and fused to obtain the text vector. The length of the text vector is 500, that is, 500 sentences are selected, and the sentences exceeding this length are selected to be truncated, and the sentence vector embedding _{enrie-sentence} constructed based on the word vector with a length of 500 and a dimension of 64 can be obtained. Input the embedding _{enrie-sentence} into Bi-LSTM to get the text vector embedding _ernie-doc .

The second processing module is configured to input the first case text and the second case text into the network model based on the text vector of the WordGCN graph for processing, so as to obtain the second processing features of the first case text and the second case text.

In one embodiment, the second processing module includes an encoding sub-module, a construction sub-module, and a second processing sub-module.

The encoding sub-module is used to encode the words of the first case text and the second case text through the relationship between the words in the sentence level and the corpus level in the WordGCN model to obtain word vectors.

Build submodules for building sentence vectors from word vectors.

The second processing submodule is used to input the sentence vector into the Bi-GRU for processing, so as to obtain the second processing feature of the first case text and the second case text.

For the encoding sub-module, the construction sub-module and the second processing sub-module, in this embodiment, the WordGCN model is selected to represent the sentence model. Words are encoded by word-to-word relationships at the sentence level and at the corpus level. Sentence encoding is performed by word segmentation and fixed sentence length. The sentence length is set to 16 words. If the length of the sentence exceeds the length of the sentence, it will be truncated. If the length of the sentence is not exceeded, it will be filled by pandding. The sentence vector is still obtained by adding the word vector. . Select 500 for the length of the text vector, and select truncation for sentences exceeding this length, so that the sentence vector embedding _words constructed based on the word vector with a length of 500 and a dimension of 64 can be obtained. Input the embedding _words into Bi-GRU to get the text vector embedding _graph-doc .

The third processing module is configured to input the first case text and the second case text into the network model based on the text vector of the subject word for processing, so as to obtain the third processing feature of the first case text and the second case text.

In one embodiment, the third processing module includes a filtering sub-module, a first extraction sub-module, a first recording sub-module, a second extraction sub-module, a second recording sub-module and a calculation sub-module.

The filtering submodule is used for filtering the stop words in the text of the first case text and the second case text through the stop word vocabulary database.

The first extraction sub-module is used to extract the subject words of the filtered text.

The first recording sub-module is used to record the position index and importance degree corresponding to the extracted subject words.

The second extraction submodule is used for extracting proper nouns in the texts of the first case text and the second case text through the proper word vocabulary database.

The second recording sub-module is used to record the position index and importance degree corresponding to the extracted proper noun.

The calculation sub-module is used for adding the importance degree of the subject word and the importance degree of the proper noun to obtain the third processing feature.

For the filtering sub-module, the first extraction sub-module, the first recording sub-module, the second extraction sub-module, the second recording sub-module and the calculation sub-module, in this embodiment, the stop words in the text are The word vocabulary database is filtered, and the subject heading model is used to extract the subject heading of the case based on the records after the above word segmentation, so as to obtain the subject heading in the case. The subject word vocabulary is BERTopic, and c-tf-idf is used internally. The keyword extraction is based on the number of categories of the case instead of the number of documents. At the same time, the LDA model is used to extract the subject words of the case content. Based on the subject words extracted by the two methods, record the position index corresponding to the subject words and the importance degree of the vocabulary. , record the position index corresponding to its proper noun, and the importance of proper noun is uniformly set to 0.5. The special word position index is generated by the union of the subject word position index and the proper noun position index. Add the importance of these keywords to get the attention weight matrix embedding _attentation , which is 16x500 in size.

The merging module is used to concate the first processing feature of the first case text and the second case text and the second processing feature of the first case text and the second case text to obtain the first case text and the second case text. Merged features of text.

In this embodiment, the feature embedding _merge is obtained by concatenating the embedding _ernie-doc and the embedding _graph-doc .

The fourth processing module is configured to input the merged features of the first case text and the second case text into the fully connected layer for processing, so as to obtain the fully connected layer processing features of the first case text and the second case text.

In this embodiment, the embedding _merge is input to the fully connected layer to obtain the feature embedding _merge-fc .

The operation module is used for multiplying the fully connected layer processing features of the first case text and the second case text and the third processing feature of the first case text and the second case text to obtain the first case text and the second case text. Textual semantic representation features of text.

In this embodiment, the final text semantic representation embedding _doc is obtained by multiplying the embedding _merge-fc matrix and the attention matrix embedding _attentation .

The fifth processing module is configured to perform full connection layer and activation function processing on the text semantic representation features of the first case text and the second case text, so as to obtain the text abstract semantic representation of the first case text and the second case text.

In this embodiment, the semantic representation embedding _doc is connected to the fully connected layer and the activation function is connected to a fully connected layer to obtain the text abstract semantic representation embedding _abs-doc .

The sixth processing module is used to process the text abstract semantic representation of the first case text and the second case text through the matrix of the fully connected layer with dimension 1 and the sigmoid activation function, so as to obtain the first case text and the second case text. Similar probability value.

In this embodiment, the text abstract semantic representation embedding _abs-doc obtained through the left and right sub-networks is connected to a matrix with a dimension of 1 in the fully connected layer and then connected to a sigmoid activation function to finally output a similarity probability value of 0 to 1.

The determining unit 150 is configured to determine that the first case text and the second case text are similar cases if the similarity probability value of the first case text and the second case text meets the set similarity threshold.

In this embodiment, the similarity threshold is 0.5. If the similarity probability value between the second case text and the first case text is greater than 0.5, it is considered that the case texts of the two cases are similar, and if it is less than 0.5, the case content of the two cases is considered dissimilar. .

By adopting the ERNIE technology, the invention can understand the meaning of the polysemy in the text well in combination with the context, analyze the similarity of the case based on the twin neural network, and combine the attention mechanism to analyze the key features of the nouns in the special field of the case. It increases the model's understanding of the case content, and improves the effectiveness and accuracy of case similarity matching.

The above-mentioned case similarity matching apparatus 100 can be implemented in the form of a computer program, and the computer program can be executed on a computer device as shown in FIG. 3 .

Please refer to FIG. 3 , which is a schematic block diagram of a computer device provided by an embodiment of the present application. The computer device 500 may be a server, where the server may be an independent server or a server cluster composed of multiple servers.

As shown in FIG. 3 , the computer device includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, the above steps of the case similarity matching method are implemented.

The computer device 700 may be a terminal or a server. The computer device 700 includes a processor 720 connected by a system bus 710 , memory, and a network interface 750 , where the memory may include a non-volatile storage medium 730 and an internal memory 740 .

The nonvolatile storage medium 730 may store an operating system 731 and a computer program 732 . The computer program 732, when executed, can cause the processor 720 to execute any one of the case similarity matching methods.

The processor 720 is used to provide computing and control capabilities to support the operation of the entire computer device 700 .

The internal memory 740 provides an environment for running the computer program 732 in the non-volatile storage medium 730. When the computer program 732 is executed by the processor 720, the processor 720 can execute any one of the case similarity matching methods.

The network interface 750 is used for network communication, such as sending assigned tasks and the like. Those skilled in the art can understand that the structure shown in FIG. 3 is only a block diagram of a partial structure related to the solution of the present application, and does not constitute a limitation on the computer device 700 to which the solution of the present application is applied. The specific computer device 700 may include more or fewer components than shown, or combine certain components, or have a different arrangement of components. Wherein, the processor 720 is used for running the program code stored in the memory to realize the following steps:

Obtain the text of the case judgment in the case database;

Collect stop words and proper noun words from the text of case judgments, and generate stop word word database and proper word word database;

Select the first case text and the second case text that need similarity matching from the case judgment text;

Input the first case text and the second case text into the twin network for processing to obtain the similarity probability value of the first case text and the second case text;

If the similarity probability value of the first case text and the second case text meets the set similarity threshold, it is determined that the first case text and the second case text are similar cases.

In one embodiment: the first case text and the second case text are input into a twin network for processing to obtain similar probability values of the first case text and the second case text, and the twin network includes an ERNIE-based The network model of text vector, the network model of text vector based on WordGCN graph, and the network model of text vector based on subject word.

In one embodiment: the inputting the first case text and the second case text into the twin network for processing to obtain similar probability values of the first case text and the second case text, including:

Inputting the first case text and the second case text into the ERNIE-based text vector network model for processing to obtain the first processing features of the first case text and the second case text;

Inputting the first case text and the second case text into the network model based on the text vector of the WordGCN graph for processing to obtain the second processing features of the first case text and the second case text;

Inputting the first case text and the second case text into the network model based on the text vector of the subject word for processing, so as to obtain the third processing feature of the first case text and the second case text;

Perform concate merge processing on the first processing feature of the first case text and the second case text and the second processing feature of the first case text and the second case text to obtain the merged feature of the first case text and the second case text;

inputting the merged features of the first case text and the second case text into the fully connected layer processing to obtain the fully connected layer processing features of the first case text and the second case text;

Multiply the fully connected layer processing features of the first case text and the second case text with the third processing feature of the first case text and the second case text to obtain the text semantic representation of the first case text and the second case text feature;

Perform full connection layer and activation function processing on the text semantic representation features of the first case text and the second case text to obtain the text abstract semantic representation of the first case text and the second case text;

The text abstract semantic representation of the first case text and the second case text is processed through a matrix of a fully connected layer with a dimension of 1 and a sigmoid activation function to obtain the similarity probability value of the first case text and the second case text.

In one embodiment: the first case text and the second case text are input into the ERNIE-based text vector network model for processing, so as to obtain the first processing feature of the first case text and the second case text, including :

Perform sentence segmentation according to the sentence segmentation symbols of the text content in the first case text and the second case text;

Tokenize the sentence by using the word segmentation tool combined with the stop word lexicon and the proprietary word lexicon to obtain word segmentation data;

The word segmentation data is processed based on MLM through ERNIE to obtain the word vector of each word;

The word vector of each word in each sentence is summed to obtain the feature vector of the sentence vector;

The feature vectors of all sentence vectors of the text content are concatenated and fused by Bi-LSTM to obtain the first processing features of the first case text and the second case text.

In one embodiment: described inputting the first case text and the second case text to a network model based on the text vector of the WordGCN graph for processing, to obtain the second processing feature of the first case text and the second case text, including :

Encode the words of the first case text and the second case text through the relationship between words in the sentence level and corpus level in the WordGCN model to obtain word vectors;

Construct sentence vector according to word vector;

Input the sentence vector into Bi-GRU for processing to obtain the second processing feature of the first case text and the second case text.

In one embodiment: the first case text and the second case text are input into the network model based on the text vector of the subject word for processing, so as to obtain the third processing feature of the first case text and the second case text, including :

Filter the stop words in the text of the first case text and the second case text through the stop word vocabulary database;

Extract the subject words of the filtered text;

Record the position index and importance degree corresponding to the extracted subject words;

Extract proper nouns in the texts of the first case text and the second case text through the proper word vocabulary database;

Record the position index and importance degree corresponding to the extracted proper noun;

The degree of importance of the subject word and the degree of importance of proper nouns are added up to obtain the third processing feature.

In one embodiment, the subject words of the filtered text are extracted by using the BERTopic model combined with the LDA model to extract the subject words of the text.

It should be understood that, in this embodiment of the present application, the processor 720 may be a central processing unit (Central Processing Unit, CPU), and the processor 720 may also be other general-purpose processors, digital signal processors (Digital Signal Processor, DSP), dedicated integrated Circuit (Application Specific Integrated Circuit, ASIC), off-the-shelf Programmable Gate Array (Field-Programmable Gate Array, FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, and the like. Wherein, the general-purpose processor can be a microprocessor or the processor can also be any conventional processor or the like.

Those skilled in the art can understand that the structure of the computer device 700 shown in FIG. 3 does not constitute a limitation on the computer device 700, and may include more or less components than the one shown, or combine some components, or different components layout.

In another embodiment of the present invention, a computer-readable storage medium is provided. The computer-readable storage medium may be a non-volatile computer-readable storage medium. The computer-readable storage medium stores a computer program, where the computer program implements the case similarity matching method disclosed in the embodiment of the present invention when the computer program is executed by the processor.

Those skilled in the art can clearly understand that, for the convenience and brevity of description, for the specific working process of the above-described devices, devices and units, reference may be made to the corresponding processes in the foregoing method embodiments, which will not be repeated here. Those of ordinary skill in the art can realize that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. Interchangeability, the above description has generally described the components and steps of each example in terms of function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Skilled artisans may implement the described functionality using different methods for each particular application, but such implementations should not be considered beyond the scope of the present invention.

Claims (10)

1. The case similarity matching method is characterized by comprising the following steps:

acquiring case judgment text in a case database;

collecting stop words and special noun words from case judgment text and generating a stop word library and a special word library;

selecting a first case text and a second case text which need to be subjected to similarity matching from case judgment text;

inputting the first case text and the second case text into a twin network for processing to obtain a similar probability value of the first case text and the second case text;

and if the similarity probability values of the first case text and the second case text meet the set similarity threshold, judging that the first case text and the second case text are similar cases.

2. The case similarity matching method according to claim 1, wherein the first case text and the second case text are input into a twin network for processing, so as to obtain the similarity probability values of the first case text and the second case text, and the twin network comprises a network model of an ERNIE-based text vector, a network model of a WordGCN-based text vector and a network model of a subject word-based text vector.

3. The case similarity matching method according to claim 2, wherein the inputting the first case text and the second case text into the twin network for processing to obtain the similarity probability value of the first case text and the second case text comprises:

inputting the first case text and the second case text into a network model of a text vector based on ERNIE for processing to obtain first processing characteristics of the first case text and the second case text;

inputting the first case text and the second case text into a network model based on a text vector of a WordGCN image for processing to obtain second processing characteristics of the first case text and the second case text;

inputting the first case text and the second case text into a network model based on text vectors of subject words for processing to obtain third processing characteristics of the first case text and the second case text;

performing concatee merging processing on the first processing characteristics of the first case text and the second processing characteristics of the first case text and the second case text to obtain merging characteristics of the first case text and the second case text;

inputting the combined features of the first case text and the second case text into a full-link layer for processing to obtain full-link layer processing features of the first case text and the second case text;

carrying out multiplication operation on the fully connected layer processing characteristics of the first case text and the second case text and the third processing characteristics of the first case text and the second case text to obtain text semantic representation characteristics of the first case text and the second case text;

carrying out full-connection layer and activation function processing on text semantic representation characteristics of the first case text and the second case text to obtain text abstract semantic representations of the first case text and the second case text;

and processing the text abstract semantic representations of the first case text and the second case text through a matrix of a full connection layer with the dimension of 1 and a sigmoid activation function to obtain the similar probability values of the first case text and the second case text.

4. The case similarity matching method according to claim 3, wherein the step of inputting the first case text and the second case text into a network model based on ERNIE text vectors for processing to obtain the first processing features of the first case text and the second case text comprises:

sentence segmentation is carried out according to sentence breaking symbols of text contents in the first case text and the second case text;

performing word segmentation on the sentence by a word segmentation tool in combination with a stop word vocabulary base and a special word vocabulary base to obtain word segmentation data;

processing the word data based on the MLM through ERNIE to obtain a word vector of each word;

summing the word vectors of each word in each sentence to obtain the feature vectors of the sentence vectors;

and carrying out concatee fusion on the feature vectors of all sentence vectors of the text content through Bi-LSTM to obtain the first processing features of the first case text and the second case text.

5. The case similarity matching method according to claim 3, wherein the step of inputting the first case text and the second case text into a network model based on text vectors of WordGCN images for processing to obtain second processing features of the first case text and the second case text comprises:

coding words of the first case text and the second case text through the relation between words in a sentence level and a corpus level in a WordGCN model to obtain word vectors;

constructing a statement vector according to the word vector;

and inputting the sentence vector into the Bi-GRU for processing to obtain a first case text and a second case text second processing characteristic.

6. The case similarity matching method according to claim 3, wherein the step of inputting the first case text and the second case text into a network model based on text vectors of subject words for processing to obtain third processing features of the first case text and the second case text comprises:

filtering stop words in the texts of the first case text and the second case text through a stop word vocabulary library;

extracting the subject terms of the filtered text;

recording the position index and the importance degree corresponding to the extracted subject term;

extracting proper nouns in the texts of the first case text and the second case text through a proper word vocabulary library;

recording the position index and the importance degree corresponding to the extracted proper nouns;

and adding the importance degree of the subject term and the importance degree of the proper noun to obtain a third processing characteristic.

7. The case similarity matching method according to claim 6, wherein the subject words of the filtered text are extracted by combining a BERTOPIC model with an LDA model.

8. The case similarity matching device is characterized by comprising an acquisition unit, a generation unit, a selection unit, a processing unit and a judgment unit;

the acquisition unit is used for acquiring case judgment text in a case database;

the generating unit is used for collecting the stop words and special noun words from the case judgment text and generating a stop word library and a special word library;

the selecting unit is used for selecting a first case text and a second case text which need to be subjected to similarity matching from case judgment text;

the processing unit is used for inputting the first case text and the second case text into the twin network for processing to obtain the similar probability values of the first case text and the second case text;

and the judging unit is used for judging that the first case text and the second case text are similar cases if the similarity probability value of the first case text and the second case text meets the set similarity threshold.

9. A computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the processor implementing the case similarity matching method steps according to any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium, characterized in that the storage medium stores a computer program comprising program instructions which, when executed by a processor, cause the processor to carry out the case similarity matching method steps according to any one of claims 1 to 7.

Images