Conference case text comparison method and system based on text clause splitting technology

Abstract

The invention provides a conference plan text comparison method and system based on a text clause splitting technology. The method comprises the steps that multiple clauses of a pre-revision text and a post-revision text of a conference plan are obtained through a natural language processing technology; inputting the plurality of terms of the text before revision and the plurality of terms of the text after revision into a pre-trained twin-tower twin neural network, and outputting similar term pairs in the text before revision and the text after revision; calculating the editing distance of the similar clause pair by using a dynamic programming algorithm to obtain the revised content of the similar clause pair; the twin-tower twin neural network is obtained based on training of a text clause labeling data set, the text clause labeling data set comprises a plurality of clause pairs with labeled semantic similarity, and text clause content modification and revision labeling of character position changes in clauses can be completed; and the comparison range can be reduced by firstly determining the similar terms, and the comparison accuracy and the comparison efficiency are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of document comparison, and in particular to a conference proposal text comparison method and system based on text clause splitting technology. Background Art

[0002] When reviewing a meeting proposal, if the auditor finds any problems during the review process, it will usually be returned to the organizer for modification. The organizer will make the modifications based on the review opinions and resubmit it for review. During the second review, the auditor usually needs to compare the contents of the two texts one by one to ensure that the problematic clauses have been modified and other clauses have not been modified.

[0003] The text comparison tools currently on the market only support mechanized comparison of simple documents, which is just a simple comparison of the same paragraphs of two documents. When auditors use common document tools for comparison, common document tools generally use mechanized comparison, for example, they only have the functions of adding and deleting comparisons, and cannot identify changes in the content of clauses and adjustments to characters in the text, which may cause auditors to miss some revised clauses and affect the audit results. In addition, when comparing, the text clauses need to be compared with the entire text before revision, which is inefficient. Summary of the invention

[0004] In order to overcome the above-mentioned deficiencies of the prior art, the present invention proposes a conference proposal text comparison method based on text clause splitting technology, comprising:

[0005] Use natural language processing technology to obtain multiple clauses of the pre-amendment text and the post-amendment text of the conference proposal;

[0006] Inputting multiple clauses of the pre-revised text and the post-revised text into a pre-trained twin-tower neural network, and outputting similar clause pairs in the pre-revised text and the post-revised text;

[0007] Using a dynamic programming algorithm to calculate the edit distance of the similar clause pairs, the revised content of the similar clause pairs is obtained;

[0008] The dual-tower twin neural network is trained based on a text clause annotation dataset, where the text clause annotation dataset includes multiple clause pairs with annotated semantic similarities.

[0009] Optionally, the obtaining the revised content of the similar clause pair by calculating the edit distance of the similar clause pair using a dynamic programming algorithm includes:

[0010] Calculating the edit distance of the similar terms pair using a dynamic programming algorithm, and obtaining an edit distance matrix generated in the process of calculating the edit distance;

[0011] Perform path backtracking based on the edit distance matrix to generate an alignment matrix of the similar clause pairs;

[0012] The difference position in the alignment matrix of the similar clause pair is marked, and the content corresponding to the difference position in the similar clause pair is determined as the revised content of the similar clause pair.

[0013] Optionally, the text clause annotation dataset is obtained in the following manner:

[0014] Collecting texts of multiple conference proposals before and after revision from a historical database, and obtaining multiple clauses in each of the texts;

[0015] Obtaining a semantic feature vector of each of the clauses, and calculating the semantic similarity between every two of the clauses based on the semantic feature vectors;

[0016] Calculating the edit distance between each two of the clauses using a dynamic programming algorithm;

[0017] Determine a number of similar clause pairs among the plurality of clauses by combining the edit distance and the semantic similarity between each two clauses;

[0018] Among the multiple clauses, the clauses belonging to several of the similar clause pairs are marked as positive samples, and the clauses not belonging to any of the similar clause pairs are marked as negative samples;

[0019] A text clause annotation dataset is constructed based on the positive samples and the negative samples with annotated semantic similarities.

[0020] Optionally, the pre-training process of the dual-tower twin neural network includes:

[0021] Dividing the text clause annotation dataset into a training set and a test set;

[0022] Taking each of the clauses in the training set as the input of the initial twin-tower neural network, outputting the semantic similarity between every two clauses, and comparing them with the annotated semantic similarity to obtain a loss value;

[0023] Iteratively updating the network parameters of the initial dual-tower twin neural network multiple times using the loss value until convergence, to obtain a dual-tower twin neural network that has completed convergence;

[0024] The test set is used to verify the converged twin-tower twin neural network to obtain a trained twin-tower twin neural network.

[0025] Optionally, the obtaining of the semantic feature vector of each of the clauses, and calculating the semantic similarity between every two of the clauses based on the semantic feature vectors, includes:

[0026] Input the clauses into a long short-term memory network to obtain a word vector in each clause, and calculate a semantic feature vector of each clause based on the weight of the word vector;

[0027] The cosine similarity between the semantic feature vectors of each two of the clauses is calculated to obtain the semantic similarity between each two of the clauses.

[0028] Optionally, combining the edit distance and semantic similarity between each two of the clauses to determine similar clause pairs among the plurality of clauses comprises:

[0029] If the semantic similarity between the two clauses is greater than the semantic similarity threshold, and the edit distance is less than the edit distance threshold, the two clauses are a similar clause pair.

[0030] Optionally, the obtaining of multiple clauses in each of the texts includes: using natural language processing technology to segment and clean the texts to obtain multiple clauses in the texts.

[0031] Based on the same inventive concept, the present invention also provides a conference proposal text comparison system based on text clause splitting technology, including:

[0032] A clause splitting module is used to obtain multiple clauses of the pre-amended text and the revised text of the conference proposal by using natural language processing technology;

[0033] A similar clause pair acquisition module is used to input multiple clauses of the pre-revised text and the revised text into a pre-trained twin-tower neural network, and output similar clause pairs in the pre-revised text and the revised text;

[0034] A revision content identification module, used for obtaining the revision content of the similar clause pair by calculating the edit distance of the similar clause pair using a dynamic programming algorithm;

[0035] The dual-tower twin neural network is trained based on a text clause annotation dataset, where the text clause annotation dataset includes multiple clause pairs with annotated semantic similarities.

[0036] Optionally, the revision content identification module is specifically used to:

[0037] Calculating the edit distance of the similar terms pair using a dynamic programming algorithm, and obtaining an edit distance matrix generated in the process of calculating the edit distance;

[0038] Perform path backtracking based on the edit distance matrix to generate an alignment matrix of the similar clause pairs;

[0039] The difference position in the alignment matrix of the similar clause pair is marked, and the content corresponding to the difference position in the similar clause pair is determined as the revised content of the similar clause pair.

[0040] Optionally, the system further comprises a neural network building module, for:

[0041] Collecting texts of multiple conference proposals before and after revision from a historical database, and obtaining multiple clauses in each of the texts;

[0042] Obtaining a semantic feature vector of each of the clauses, and calculating the semantic similarity between every two of the clauses based on the semantic feature vectors;

[0043] Calculating the edit distance between each two of the clauses using a dynamic programming algorithm;

[0044] Determine a number of similar clause pairs among the plurality of clauses by combining the edit distance and the semantic similarity between each two clauses;

[0045] Among the multiple clauses, if a clause belongs to the similar clause pair, it is marked as a positive sample, otherwise, it is marked as a negative sample;

[0046] A text clause annotation dataset is constructed based on the positive samples and the negative samples with annotated semantic similarities.

[0047] Optionally, the neural network building module is further used to:

[0048] Dividing the text clause annotation dataset into a training set and a test set;

[0049] Taking each of the clauses in the training set as the input of the initial twin-tower neural network, outputting the semantic similarity between every two clauses, and comparing them with the annotated semantic similarity to obtain a loss value;

[0050] Iteratively updating the network parameters of the initial dual-tower twin neural network multiple times using the loss value until convergence, to obtain a dual-tower twin neural network that has completed convergence;

[0051] The test set is used to verify the converged twin-tower twin neural network to obtain a trained twin-tower twin neural network.

[0052] Optionally, the neural network building module is specifically used to:

[0053] Input the clauses into a long short-term memory network to obtain a word vector in each clause, and calculate a semantic feature vector of each clause based on the weight of the word vector;

[0054] The cosine similarity between the semantic feature vectors of each two of the clauses is calculated to obtain the semantic similarity between each two of the clauses.

[0055] Optionally, the neural network building module is specifically used to:

[0056] If the semantic similarity between the two clauses is greater than the semantic similarity threshold, and the edit distance is less than the edit distance threshold, the two clauses are a similar clause pair.

[0057] Optionally, the neural network building module is specifically used to:

[0058] The text is segmented and cleaned by using natural language processing technology to obtain multiple clauses in the text.

[0059] In another aspect, the present application further provides an electronic device, comprising: at least one processor and a memory; the memory and the processor are connected via a bus;

[0060] The memory is used to store one or more programs;

[0061] When the one or more programs are executed by the at least one processor, the conference proposal text comparison method based on the text clause splitting technology as described above is implemented.

[0062] On the other hand, the present application also provides a computer-readable storage medium having an execution program stored thereon, which, when executed, implements the conference proposal text comparison method based on the text clause splitting technology as described above.

[0063] Compared with the closest prior art, the present invention has the following beneficial effects:

[0064] The conference proposal text comparison method and system based on text clause splitting technology provided by the present invention include: using natural language processing technology to respectively obtain multiple clauses of the pre-revised text and the revised text of the conference proposal; inputting the multiple clauses of the pre-revised text and the revised text into a pre-trained twin-tower neural network, and outputting similar clause pairs in the pre-revised text and the revised text; using a dynamic programming algorithm to calculate the edit distance of the similar clause pairs to obtain the revised content of the similar clause pairs; the twin-tower neural network is trained based on a text clause annotation data set, and the text clause annotation data set includes annotated semantically similar clauses. The invention uses natural language processing technology to split clauses, which can realize automatic processing of texts; uses a dual-tower twin neural network to identify similar clause pairs, which reduces the comparison range and improves the comparison efficiency and accuracy compared to the existing method of comparing any two clauses; calculates the edit distance of similar clause pairs through a dynamic programming algorithm, and can obtain the difference position in the relative clause pair from the calculation path, that is, the revision position, and then completes the modification of the text clause content and the revision annotation of the character position change in the clause; the final revised content is used in the meeting proposal review process, and the abnormal information is automatically prompted to the reviewer. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 A schematic diagram of the flow of a conference proposal text comparison method based on text clause splitting technology provided by the present invention;

[0066] Figure 2 A schematic diagram of the structure of a conference proposal text comparison system based on text clause splitting technology provided by the present invention;

[0067] Figure 3 The present invention provides a schematic structural diagram of an electronic device. DETAILED DESCRIPTION

[0068] The specific implementation modes of the present invention are further described in detail below with reference to the accompanying drawings.

[0069] Example 1

[0070] The conference proposal text comparison method based on the text clause splitting technology provided by the present invention is as follows: Figure 1 As shown, including:

[0071] S1. Use natural language processing technology to obtain multiple clauses of the pre-revised text and the revised text of the meeting proposal;

[0072] S2, inputting multiple clauses of the text before revision and the text after revision into a pre-trained twin-tower neural network, and outputting similar clause pairs in the text before revision and the text after revision;

[0073] S3, obtaining revised contents of the similar clause pairs by calculating the edit distance of the similar clause pairs using a dynamic programming algorithm;

[0074] The dual-tower twin neural network is trained based on a text clause annotation dataset, where the text clause annotation dataset includes multiple clause pairs with annotated semantic similarities.

[0075] In step S1, natural language processing technology is used to obtain multiple clauses of the pre-revised text and the revised text of the conference proposal.

[0076] Specifically, for the pre-revised and post-revised texts of the conference proposals, natural language processing (NLP) technology is used to pre-process the proposal content, including text analysis, word segmentation, stop word removal, syntactic analysis and semantic understanding, to automatically identify and split the clauses in the pre-revised and post-revised texts.

[0077] Text clause splitting mainly involves the automated processing of the text content of conference proposals in order to split complex text clauses into more understandable or manageable parts. It usually involves natural language processing technology, including text analysis, word segmentation, syntactic analysis, semantic understanding, etc.

[0078] The following is a detailed discussion of text clause splitting technology. With the advent of the information age, text data has exploded, and how to efficiently process and analyze this text data has become an important issue. The emergence of text clause splitting technology aims to provide an automated method to split complex text clauses into smaller units for subsequent retrieval, analysis, understanding or application, which is of great significance to many fields such as law, finance, and scientific research. Text clause splitting technology has the following advantages:

[0079] Automated processing: Text clause splitting usually adopts automated processing technology, which can automatically identify and split the clauses in the text, reduce manual intervention, and improve processing efficiency; High precision: Through advanced NLP technology, the patent can accurately identify the clause boundaries in the text to ensure the accuracy of splitting; Flexibility: It usually takes into account the differences in different text types and clause structures, and can flexibly respond to various complex text splitting needs; Scalability: With the continuous development of NLP technology, text clause splitting technology can also be continuously upgraded and optimized to adapt to new text processing needs.

[0080] Text clause splitting is an important application of NLP technology in the field of text processing. Through the technical characteristics of automation, high precision, flexibility and scalability, it can effectively solve the problem of text clause splitting and provide strong support for multiple fields such as law, finance, and scientific research.

[0081] In step S2, multiple clauses of the pre-revised text and the revised text are input into a pre-trained dual-tower twin neural network, and similar clause pairs in the pre-revised text and the revised text are output.

[0082] S21. Obtain a text clause annotation dataset.

[0083] S21-1. Collect the texts of multiple conference proposals before and after revision from the historical database, and obtain multiple clauses in each of the texts.

[0084] Text data for training and testing neural networks are collected from historical databases. The text data can come from different sources, such as Internet texts, academic papers, conference proposal documents, etc. Natural language processing technology is used to clean and segment the pre-revised text and the post-revised text in the text data to obtain multiple clauses in the pre-revised text and the post-revised text. The text cleaning includes removing noise data, processing missing values, removing stop words, etc., to improve data quality.

[0085] S21-2, obtaining a semantic feature vector of each of the clauses, and calculating the semantic similarity between every two of the clauses based on the semantic feature vectors.

[0086] Specifically, the clauses are input into a long short-term memory network to obtain a word vector in each clause, and a semantic feature vector of each clause is calculated based on the weight of the word vector;

[0087] The cosine similarity between the semantic feature vectors of each two of the clauses is calculated to obtain the semantic similarity between each two of the clauses.

[0088] Optionally, in this solution, the semantic feature vector of each of the clauses may be obtained through methods such as bag-of-words model, TF-IDF (Term Frequency-Inverse Document Frequency), word embedding (such as Word2Vec (a model for generating word embeddings), GloVe (a word embedding model based on global word co-occurrence statistics), BERT (Bidirectional Encoder Representation Model)), etc.

[0089] S21-3. Calculate the edit distance between every two of the clauses using a dynamic programming algorithm.

[0090] S21-4. Determine several similar clause pairs among the plurality of clauses by combining the edit distance and semantic similarity between each two of the clauses.

[0091] Specifically, if the semantic similarity between the two clauses is greater than a semantic similarity threshold, and the edit distance is less than an edit distance threshold, then the two clauses are a similar clause pair.

[0092] S21-5. Among the multiple clauses, the clauses belonging to several of the similar clause pairs are marked as positive samples, and the clauses that do not belong to any of the similar clause pairs are marked as negative samples; based on the positive samples and the negative samples with marked semantic similarities, a text clause annotation dataset is constructed.

[0093] Combining the semantic similarity and edit distance of the terms to jointly determine similar term pairs can make the results more accurate and more efficient than manual labeling.

[0094] S22: training a twin-tower twin neural network based on the text clause annotation dataset, wherein the pre-training process of the twin-tower twin neural network includes:

[0095] Dividing the text clause annotation dataset into a training set and a test set;

[0096] Set the corresponding parameters in the network training process, such as learning rate, number of iterations, and batch size;

[0097] Taking each of the clauses in the training set as the input of the initial twin-tower neural network, outputting the semantic similarity between every two clauses, and comparing them with the annotated semantic similarity to obtain a loss value;

[0098] Iteratively updating the network parameters of the initial dual-tower twin neural network multiple times using the loss value until convergence, to obtain a dual-tower twin neural network that has completed convergence;

[0099] The test set is used to verify the converged twin-tower twin neural network to obtain a trained twin-tower twin neural network.

[0100] When using the test set to verify the dual-tower twin neural network, select evaluation indicators to evaluate the performance of the neural network, such as accuracy, recall rate, and F1 score; according to the test evaluation results, tune the network parameters of the neural network to improve the performance of the neural network.

[0101] S23. Input multiple clauses of the text before revision and the text after revision into a pre-trained dual-tower twin neural network, and output similar clause pairs in the text before revision and the text after revision.

[0102] In the judgment of clause similarity, the dual-tower twin network model can quickly and accurately judge the similarity between clauses by calculating the distance between clauses in a low-dimensional semantic space, thereby achieving efficient clause similarity retrieval and matching.

[0103] In step S3, a dynamic programming algorithm is used to calculate the edit distance of the similar clause pairs to obtain revised contents of the similar clause pairs.

[0104] Calculating the edit distance of the similar terms pair using a dynamic programming algorithm, and obtaining an edit distance matrix generated in the process of calculating the edit distance;

[0105] Perform path backtracking based on the edit distance matrix to generate an alignment matrix of the similar clause pairs;

[0106] The difference position in the alignment matrix of the similar clause pair is marked, and the content corresponding to the difference position in the similar clause pair is determined as the revised content of the similar clause pair.

[0107] By using the path backtracking in the dynamic programming algorithm to mark the difference position, the revision content in similar clause pairs can be determined, and the revision content can be automatically identified. The revision content is marked with different colors.

[0108] The present invention provides a function of comparing the versions before and after revision. On the one hand, the auditor can upload different versions of the meeting proposal text for comparison. By comparing the versions, the auditor has a clear understanding of the work; on the other hand, the organizer adjusts the terms of the meeting proposal and analyzes the risk probability and severity based on the adjusted terms. During the review stage, the system automatically prompts the auditor to compare the original terms of the relevant terms. The auditor can analyze the corresponding legal risks, their probability of occurrence and the severity of the consequences for the revised terms, and identify and determine the relevant risks.

[0109] In view of the shortcomings in the prior art, the present invention provides a conference proposal text comparison method, which can complete the revision annotation of the text clause content modification and the character position change in the clause, and automatically prompt the reviewer with abnormal information by marking the revised content. The purpose of the present invention is to provide a text comparison tool to support the reviewer to view the differences between different versions of conference proposals during the conference proposal review process, and to more clearly and intuitively display which clauses and contents have been modified in different versions of conference proposal texts, and to prompt the relevant reviewers with abnormal information. The present invention does not simply compare whether two texts have new or deleted paragraphs, but uses the NLP clause splitting technology to first split the complete conference proposal text into multiple clauses, input them into a pre-trained dual-tower twin neural network, output the similar clause pairs in the text before and after the revision, and use the dynamic programming algorithm to determine the revised content of the similar clause pairs. By marking the revised content with different colors, the revised edited content between clauses can be displayed more intuitively.

[0110] Example 2

[0111] Based on the same inventive concept, the present invention also provides a conference proposal text comparison system based on text clause splitting technology, such as Figure 2 As shown, including:

[0112] A clause splitting module is used to obtain multiple clauses of the pre-amended text and the revised text of the conference proposal by using natural language processing technology;

[0113] A similar clause pair acquisition module is used to input multiple clauses of the pre-revised text and the revised text into a pre-trained twin-tower neural network, and output similar clause pairs in the pre-revised text and the revised text;

[0114] A revision content identification module, used for obtaining the revision content of the similar clause pair by calculating the edit distance of the similar clause pair using a dynamic programming algorithm;

[0115] The dual-tower twin neural network is trained based on a text clause annotation dataset, where the text clause annotation dataset includes multiple clause pairs with annotated semantic similarities.

[0116] In a possible implementation, the revised content identification module is specifically used to:

[0117] Calculating the edit distance of the similar terms pair using a dynamic programming algorithm, and obtaining an edit distance matrix generated in the process of calculating the edit distance;

[0118] Perform path backtracking based on the edit distance matrix to generate an alignment matrix of the similar clause pairs;

[0119] The difference position in the alignment matrix of the similar clause pair is marked, and the content corresponding to the difference position in the similar clause pair is determined as the revised content of the similar clause pair.

[0120] In a possible implementation, the system further includes a neural network building module, which is used to:

[0121] Collecting texts of multiple conference proposals before and after revision from a historical database, and obtaining multiple clauses in each of the texts;

[0122] Obtaining a semantic feature vector of each of the clauses, and calculating the semantic similarity between every two of the clauses based on the semantic feature vectors;

[0123] Calculating the edit distance between each two of the clauses using a dynamic programming algorithm;

[0124] Determine a number of similar clause pairs among the plurality of clauses by combining the edit distance and the semantic similarity between each two clauses;

[0125] Among the multiple clauses, if a clause belongs to the similar clause pair, it is marked as a positive sample, otherwise, it is marked as a negative sample;

[0126] A text clause annotation dataset is constructed based on the positive samples and the negative samples with annotated semantic similarities.

[0127] In a possible implementation, the neural network building module is further used to:

[0128] Dividing the text clause annotation dataset into a training set and a test set;

[0129] Taking each of the clauses in the training set as the input of the initial twin-tower neural network, outputting the semantic similarity between every two clauses, and comparing them with the annotated semantic similarity to obtain a loss value;

[0130] Iteratively updating the network parameters of the initial dual-tower twin neural network multiple times using the loss value until convergence, to obtain a dual-tower twin neural network that has completed convergence;

[0131] The test set is used to verify the converged twin-tower twin neural network to obtain a trained twin-tower twin neural network.

[0132] In a possible implementation, the neural network building module is specifically used to:

[0133] Input the clauses into a long short-term memory network to obtain a word vector in each clause, and calculate a semantic feature vector of each clause based on the weight of the word vector;

[0134] The cosine similarity between the semantic feature vectors of each two of the clauses is calculated to obtain the semantic similarity between each two of the clauses.

[0135] In a possible implementation, the neural network building module is specifically used to:

[0136] If the semantic similarity between the two clauses is greater than the semantic similarity threshold, and the edit distance is less than the edit distance threshold, the two clauses are a similar clause pair.

[0137] In a possible implementation, the neural network building module is specifically used to:

[0138] The text is segmented and cleaned by using natural language processing technology to obtain multiple clauses in the text.

[0139] Example 3

[0140] like Figure 3As shown, the present invention also provides an electronic device, which may be a computer device, a single-chip device, an intelligent mobile device, etc. The electronic device in this embodiment may include a processor, a memory, a transceiver component, etc. The memory, the processor, and the transceiver component are connected via a bus; the memory may be used to store an execution program, and an exemplary execution program may include instructions; the processor is used to execute the instructions stored in the memory. The memory may also be used to store data, which may be called and / or modified when the instructions are executed.

[0141] The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, which is suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions in a storage medium to implement the corresponding method flow or corresponding functions, so as to realize the steps of the conference proposal text comparison method based on text clause splitting technology in the above-mentioned embodiment.

[0142] Example 4

[0143] Based on the same inventive concept, the present invention also provides a readable storage medium, specifically an electronic device readable storage medium (Memory), which is a memory device in an electronic device for storing programs and data. It is understandable that the storage medium here may include a built-in storage medium in an electronic device, and of course may also include an extended storage medium supported by the electronic device. The storage medium provides a storage space, which stores the operating system of the terminal. In addition, one or more instructions suitable for being loaded and executed by a processor are also stored in the storage space, and these instructions may be one or more execution programs (including program codes). It should be noted that the storage medium here may be a high-speed RAM memory, or a non-volatile memory (non-volatile memory), such as at least one disk storage. The processor loads and executes one or more instructions stored in the storage medium, which can implement the steps of the conference proposal text comparison method based on the text clause splitting technology in the above embodiment.

[0144] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0145] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0146] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0147] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0148] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit its protection scope. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that after reading the present invention, those skilled in the art can still make various changes, modifications or equivalent substitutions to the specific implementation methods of the application, but these changes, modifications or equivalent substitutions are all within the protection scope of the claims to be approved.

Claims

1. A conference proposal text comparison method based on text clause splitting technology, characterized in that: include: Use natural language processing technology to obtain multiple clauses of the pre-amendment text and the post-amendment text of the conference proposal; Inputting multiple clauses of the pre-revised text and the post-revised text into a pre-trained twin-tower neural network, and outputting similar clause pairs in the pre-revised text and the post-revised text; Using a dynamic programming algorithm to calculate the edit distance of the similar clause pairs, the revised content of the similar clause pairs is obtained; The dual-tower twin neural network is trained based on a text clause annotation dataset, where the text clause annotation dataset includes multiple clause pairs with annotated semantic similarities.

2. The method according to claim 1, characterized in that The method of obtaining revised contents of the similar clause pairs by calculating the edit distance of the similar clause pairs using a dynamic programming algorithm includes: Calculating the edit distance of the similar terms pair using a dynamic programming algorithm, and obtaining an edit distance matrix generated in the process of calculating the edit distance; Perform path backtracking based on the edit distance matrix to generate an alignment matrix of the similar clause pairs; The difference position in the alignment matrix of the similar clause pair is marked, and the content corresponding to the difference position in the similar clause pair is determined as the revised content of the similar clause pair.

3. The method according to claim 1, characterized in that The text clause annotation dataset is obtained in the following manner: Collecting texts of multiple conference proposals before and after revision from a historical database, and obtaining multiple clauses in each of the texts; Obtaining a semantic feature vector of each of the clauses, and calculating the semantic similarity between every two of the clauses based on the semantic feature vectors; Calculating the edit distance between each two of the clauses using a dynamic programming algorithm; Determine a number of similar clause pairs among the plurality of clauses by combining the edit distance and the semantic similarity between each two clauses; Among the multiple clauses, the clauses belonging to several of the similar clause pairs are marked as positive samples, and the clauses not belonging to any of the similar clause pairs are marked as negative samples; A text clause annotation dataset is constructed based on the positive samples and the negative samples with annotated semantic similarities.

4. The method according to claim 3, characterized in that The pre-training process of the dual-tower twin neural network includes: Dividing the text clause annotation dataset into a training set and a test set; Taking each of the clauses in the training set as the input of the initial twin-tower neural network, outputting the semantic similarity between every two clauses, and comparing them with the annotated semantic similarity to obtain a loss value; Iteratively updating the network parameters of the initial dual-tower twin neural network multiple times using the loss value until convergence, to obtain a dual-tower twin neural network that has completed convergence; The test set is used to verify the converged twin-tower twin neural network to obtain a trained twin-tower twin neural network.

5. The method according to claim 3, characterized in that The step of obtaining a semantic feature vector of each of the clauses and calculating the semantic similarity between each two of the clauses based on the semantic feature vectors includes: Input the clauses into a long short-term memory network to obtain a word vector in each clause, and calculate a semantic feature vector of each clause based on the weight of the word vector; The cosine similarity between the semantic feature vectors of each two of the clauses is calculated to obtain the semantic similarity between each two of the clauses.

6. The method according to claim 3, characterized in that Combining the edit distance and the semantic similarity between each two of the clauses to determine similar clause pairs among the plurality of clauses comprises: If the semantic similarity between the two clauses is greater than the semantic similarity threshold, and the edit distance is less than the edit distance threshold, the two clauses are a similar clause pair.

7. The method according to claim 3, characterized in that The obtaining of the multiple clauses in each of the texts includes: using natural language processing technology to segment and clean the texts to obtain the multiple clauses in the texts.

8. The conference proposal text comparison system based on text clause splitting technology is characterized by: include; A clause splitting module is used to obtain multiple clauses of the pre-amended text and the revised text of the conference proposal by using natural language processing technology; A similar clause pair acquisition module is used to input multiple clauses of the pre-revised text and the revised text into a pre-trained twin-tower neural network, and output similar clause pairs in the pre-revised text and the revised text; A revision content identification module, used for obtaining the revision content of the similar clause pair by calculating the edit distance of the similar clause pair using a dynamic programming algorithm; The dual-tower twin neural network is trained based on a text clause annotation dataset, where the text clause annotation dataset includes multiple clause pairs with annotated semantic similarities.

9. The system according to claim 8, characterized in that The revision content identification module is specifically used for: Calculating the edit distance of the similar terms pair using a dynamic programming algorithm, and obtaining an edit distance matrix generated in the process of calculating the edit distance; Perform path backtracking based on the edit distance matrix to generate an alignment matrix of the similar clause pairs; The difference position in the alignment matrix of the similar clause pair is marked, and the content corresponding to the difference position in the similar clause pair is determined as the revised content of the similar clause pair.

10. The system according to claim 8, characterized in that The system also includes a neural network building module for: Collecting texts of multiple conference proposals before and after revision from a historical database, and obtaining multiple clauses in each of the texts; Obtaining a semantic feature vector of each of the clauses, and calculating the semantic similarity between every two of the clauses based on the semantic feature vectors; Calculating the edit distance between each two of the clauses using a dynamic programming algorithm; Determine a number of similar clause pairs among the plurality of clauses by combining the edit distance and the semantic similarity between each two clauses; Among the multiple clauses, the clauses belonging to several of the similar clause pairs are marked as positive samples, and the clauses not belonging to any of the similar clause pairs are marked as negative samples; A text clause annotation dataset is constructed based on the positive samples and the negative samples with annotated semantic similarities.

Images