Deep learning-based legal consultation intelligent response method

Abstract

The invention relates to the technical field of artificial intelligence and legal information processing, in particular to a legal consultation intelligent response method based on deep learning. The method comprises the steps that data preprocessing is conducted on legal consultation questions input by a user, and a semantic nested matrix is generated; dynamically adjusting a nesting weight, and constructing a generation matrix in combination with the semantic nesting matrix; comparing the generated matrix with a legal knowledge base matrix layer by layer to confirm the matching degree of the generated solution and the elements of the legal knowledge base matrix; taking a result with the highest matching degree as a preliminarily generated solution, and generating a feedback matrix; and optimizing and updating the generated matrix based on the feedback matrix to obtain a generated solution matrix. The problems that an existing intelligent legal consultation system is insufficient in legal semantic understanding, limited in semantic weight adjustment, poor in logicality of generated content, low in legal clause matching degree, lack of a self-adaptive optimization mechanism fed back by a user and incomplete in expression ability of natural language output are solved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and legal information processing technology, and in particular to a legal consultation intelligent response method based on deep learning. Background Art

[0002] With the development of information technology and the Internet, the demand for legal services is growing, especially in the field of online legal consultation. More and more people hope to get professional legal answers in a convenient way. However, legal language is usually highly complex and rigorous, containing a large number of professional terms, complex logical structures, and strict interpretation requirements for clauses, which makes it difficult for ordinary users to understand relevant legal clauses or obtain effective solutions when encountering legal problems. Traditional legal consulting services usually rely on manual answers from lawyers or legal professionals. This method is not only inefficient and time-consuming, but also requires high labor costs, and cannot meet the public's demand for fast and convenient legal consultation. In this context, the intelligent legal consultation system has emerged as an effective legal service means. With the help of artificial intelligence technologies such as deep learning and natural language processing, the intelligent legal consultation system attempts to achieve rapid response and professional answers to users' legal consultation questions through automation.

[0003] The core of the intelligent legal consulting system is to accurately understand the natural language legal consulting questions input by users, and generate professional and accurate answers based on the legal knowledge base or relevant regulations. To achieve this goal, deep learning technology, especially natural language processing technology (such as word vector models, semantic parsing, generative models, etc.) has been widely used in recent years. Through machine learning methods, a large amount of legal text data can be learned, which can enable the intelligent legal consulting system to have a certain semantic understanding ability. However, due to the particularity of legal language and the diversity of user needs, the challenges faced by the intelligent legal consulting system are also very significant.

[0004] The main defects of existing intelligent legal consultation systems are insufficient understanding of legal semantics, limitations in semantic weight adjustment, lack of logic in generated content, low matching degree of legal terms, lack of adaptive optimization mechanism for user feedback, and imperfect expressive ability of natural language output, which leads to a poor overall consultation experience and makes it difficult to meet users' requirements for professionalism and readability. Summary of the invention

[0005] The present invention provides a legal consultation intelligent response method based on deep learning to solve the problems existing in the existing intelligent legal consultation system, such as insufficient understanding of legal semantics, limitation of semantic weight adjustment, lack of logic in generated content, low matching degree of legal clauses, lack of adaptive optimization mechanism for user feedback, and imperfect expression ability of natural language output, which lead to poor overall consultation experience and difficulty in meeting users' requirements for professionalism and readability.

[0006] A legal consultation intelligent answering method based on deep learning, comprising the following steps:

[0007] S1. Preprocess the data of the legal consultation questions input by the user and generate a semantic nesting matrix to express the multi-layer semantic features of the legal consultation questions; optimize the semantic features at different levels by dynamically adjusting the nesting weights;

[0008] S2. Based on the adjusted weight values ​​and the semantic nesting matrix, a generation matrix is ​​constructed to generate preliminary legal consultation answers; the generation matrix is ​​compared with the legal knowledge base matrix layer by layer to confirm the matching degree between the generated answers and the elements of the legal knowledge base matrix; the result with the highest matching degree is used as the preliminary generated answer, and the generation matrix is ​​dynamically adjusted to generate a feedback matrix; based on the feedback matrix, the generation matrix is ​​optimized and updated to obtain the generated answer matrix.

[0009] Preferably, the S1 specifically includes:

[0010] The input legal consultation question is decomposed into a vocabulary sequence, and the legal vocabulary and terminology are further identified and marked; each word in the vocabulary sequence is converted into a feature point in a high-dimensional vector space, and then the semantic embedding matrix is ​​obtained through the calculation formula of the semantic embedding matrix.

[0011] Preferably, the S1 specifically includes:

[0012] The calculation formula of the semantic embedding matrix is ​​as follows:

[0013]

[0014] Among them, M i，j is an element in the semantic embedding matrix, representing the embedding semantic value of the jth word at the i-th layer of the parsing layer; N represents the number of words in the sentence; k is the word ordinal number; φ(Q k ) is the semantic feature vector of the kth word in the user input Q; ψ(T k ) is the feature vector of the kth word in the legal knowledge base matrix; σ is the activation function; γ i is the nesting level offset; θ j is the vocabulary shift coefficient.

[0015] Preferably, the S1 specifically includes:

[0016] A dynamic adjustment mechanism is introduced to calculate the nested weights in real time according to the current input content, and the calculated weight values ​​will be used as control variables for generating the matrix.

[0017] Preferably, the S1 specifically includes:

[0018] The dynamic adjustment formula of nested weight is:

[0019]

[0020] Among them, W i+1，j represents the weight value of the jth semantic feature in the i+1th layer; W i，j is the weight of the jth semantic feature in the i-th layer; α i is the basic adjustment coefficient for the i-th layer; β j is the offset parameter of the jth word; δ p is the offset adjustment term; P is the total number of offset parameters; ΔW is the weight difference; λ is the smoothing parameter; is the control parameter.

[0021] Preferably, the S2 specifically includes:

[0022] The calculation formula of the generation matrix is ​​introduced to progressively generate the semantic information of each layer. The specific formula is as follows:

[0023]

[0024] Among them, G i，j represents the generated value of the jth semantic feature in the i-th layer of the generator matrix; W i，l is the weight value of the jth semantic feature of the i-th layer after dynamic adjustment; L is the total number of nested generated levels; κ is the generated offset coefficient; |M i，l -T j，l | 2 is the semantic matching error term; T j，l represents the semantic features of the jth word in the lth layer of the nested generation process in the legal knowledge base matrix; η l Generate reconciliation items for the hierarchy.

[0025] Preferably, the S2 specifically includes:

[0026] Map user feedback data into the feedback matrix, optimize the weight values ​​of the generation matrix and the matching matrix, and correct the semantic deviation in the answer; the calculation formula of the feedback matrix is ​​as follows:

[0027]

[0028] Among them, F i，jis an element in the feedback matrix, indicating the correction value of the jth semantic feature in the i-th layer; ω 1i is the user rating weight of the i-th layer, representing user feedback data; A i，j is the element in the matching matrix A, which represents the output value of the initial solution generated according to the matching degree in the semantic feature; ω 2i is the weight coefficient of the feedback offset term; F u is the user's feedback score on the initially generated solution; δ is a positive number used to prevent the denominator from being zero.

[0029] Preferably, the S2 specifically includes:

[0030] The correction value is applied to each semantic feature in the generated matrix to form a generated answer matrix; the specific formula is:

[0031]

[0032] in, is the element of the generated matrix after feedback optimization, representing the element in the generated solution matrix; is the feedback strength coefficient.

[0033] Preferably, the S2 specifically includes:

[0034] The generated answer matrix expresses the answer content after feedback optimization, converts the answer content into a text form of natural language output, and performs intelligent response to legal consultation.

[0035] The beneficial effects of the technical solution of the present invention are:

[0036] 1. Utilize word segmentation, part-of-speech tagging, and embedding models to convert the user's natural language input into a high-dimensional semantic representation, and identify specific legal concepts based on the terminology library and corpus in the legal field, fully understand the legal semantics of the user's input, and thus more accurately reflect the semantic characteristics of legal issues; the generation of a multi-level semantic nested matrix effectively enhances the depth of understanding of the input data, ensuring that subsequent generation and matching steps are based on clear legal semantic information.

[0037] 2. Based on the semantic nesting matrix, a dynamic adjustment mechanism is added to calculate the nesting weights in real time and make detailed adjustments according to the needs of different levels. Through the dynamic adjustment of the nesting weights, important legal concepts can be magnified during the parsing process, while the weights of unimportant words can be reduced, thereby ensuring that the final generated answer accurately expresses the complexity of the legal issue at each level, avoiding the limitations of fixed weights in complex legal issues, and ensuring that the parsing and generation results are more in line with the professional needs of legal consultation.

[0038] 3. When generating answers, a generation matrix is ​​introduced to ensure that each semantic feature generates an answer based on multi-level analysis; by integrating information at different analysis levels, the generated answers are both accurate and coherent, meeting the requirements of legal logic; the generation matrix uses the dynamically adjusted weight values ​​and the semantic nesting matrix to generate content that is both accurate and comprehensive, ensuring the integrity and logic of the answers to legal questions.

[0039] 4. The semantic matching process in the present invention realizes a multi-level comparison between the generated answers and the legal knowledge base, and ensures the compatibility of the generated content with the legal clauses through layer-by-layer semantic matching; by comparing the generated answers and the elements of the legal knowledge base matrix at different parsing levels, high-quality answers that comply with the actual legal provisions are provided; by processing the answers with the highest matching degree, accurate answers with legal basis can be generated, thereby improving the professionalism and credibility of intelligent responses. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 This is a flow chart of the legal consultation intelligent response method based on deep learning described in the present invention. DETAILED DESCRIPTION

[0041] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the technical scheme in the embodiment of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiment of the present invention. Obviously, the described embodiment is only a part of the embodiment of the present invention, not all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0042] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0043] The following is a detailed description of a specific scheme of a legal consultation intelligent response method based on deep learning provided by the present invention in conjunction with the accompanying drawings.

[0044] Refer to the attached Figure 1 , which shows a flow chart of a legal consultation intelligent response method based on deep learning provided by an embodiment of the present invention, the method comprising the following steps:

[0045] S1. Preprocess the data of the legal consultation questions input by the user and generate a semantic nesting matrix to express the multi-layer semantic features of the legal consultation questions; optimize the semantic features at different levels by dynamically adjusting the nesting weights;

[0046] In the legal consultation intelligent answering system, when the user enters a legal consultation question, data preprocessing is first performed to ensure that the input legal consultation question meets the requirements of semantic parsing and nested matrix generation. The core of data acquisition is to perform multi-level structural processing on the natural language text entered by the user, converting complex legal language into standardized data suitable for deep learning algorithm processing.

[0047] Specifically, the input legal consultation question Q is decomposed into a vocabulary sequence through word segmentation and part-of-speech tagging technology Where N represents the number of words in the sentence, and k is the word ordinal number. Then, based on the terminology library and corpus in the legal field, specific legal vocabulary and key terms, such as contracts, intellectual property rights, dispute resolution, etc., are identified and marked, and legal concepts in specific fields are processed in a targeted manner. In order to further improve the accuracy of data acquisition, the existing embedding model (such as the pre-trained legal field word embedding model) is used to convert each word into a feature point in the high-dimensional vector space, and then the semantic embedding matrix is ​​obtained through the calculation formula of the semantic embedding matrix; each element in the semantic embedding matrix represents the nested semantic value of the jth word in the i-th layer of the parsing layer, thereby expressing the multi-layer semantic characteristics of legal consulting issues, ensuring that the subsequent generation and matching links have accurate legal semantic information.

[0048] The calculation formula of the semantic embedding matrix is ​​as follows:

[0049]

[0050] Among them, M i，j is an element in the semantic nesting matrix, indicating the nested semantic value of the jth word in the i-th layer of the parsing layer, which is used to reflect the legal semantic weight of the word in multi-layer parsing; φ(Q k ) is the semantic feature vector of the kth word in the user input Q, which represents the characteristics of the vocabulary in the user input Q in the semantic space; ψ(T k ) is the feature vector of the kth word in the legal knowledge base matrix, representing the specific semantics of the word in the legal field; σ is an activation function used to perform nonlinear smoothing on the elements of the semantic embedding matrix to make the output result more semantically discerning; γ i is the nested level offset, indicating the semantic differences between different layers; θ j is the vocabulary shift coefficient, which is used to adjust the weight difference of each word at different layers, so as to capture the multi-layer characteristics of legal language in the nested parsing process.

[0051] After completing the calculation of the semantic nesting matrix, the semantic features at different levels are optimized by adjusting the nesting weights so that they can better express the complexity of legal issues in the generation process; a dynamic adjustment mechanism is introduced to calculate the nesting weights in real time based on the current input content to reflect the differences in the importance of vocabulary and levels; the calculated weight values ​​will be used as control variables for the generation matrix to ensure that the semantic information in the generation process is accurately distributed and expressed at different levels.

[0052] The dynamic adjustment formula of nested weight is:

[0053]

[0054] Among them, W i+1，j represents the weight value of the jth semantic feature in the i+1th layer, which is used to control the generation strength of the corresponding semantic feature in the generation matrix; W i，j is the weight of the jth semantic feature in the i-th layer, indicating the semantic importance of the vocabulary in the current layer; α i is the basic adjustment coefficient for the i-th layer, which is used to define the basic weight value of each layer of analysis, and is calculated by the frequency of occurrence of each word in the legal terms at different levels; β j is the offset parameter of the jth word, which is used to adjust the importance of the vocabulary in the current level, and is obtained by calculating the frequency of each word as a keyword in the knowledge base; δ p is the offset adjustment term, which reflects the semantic differences of different parsing layers by setting different offsets in the multi-layer structure, and is obtained by calculating the variance of each word vector in each layer; P is the total number of offset parameters, which represents the number of offset parameters used to calculate the semantic differences between layers in the multi-layer nested structure; ΔW is the weight difference; λ is a smoothing parameter used to control the size of the offset; is a control parameter used to ensure smoothness.

[0055] S2. Based on the adjusted weight values ​​and the semantic nesting matrix, a generation matrix is ​​constructed to generate preliminary legal consultation answers; the generation matrix is ​​compared with the legal knowledge base matrix layer by layer to confirm the matching degree between the generated answers and the elements of the legal knowledge base matrix; the result with the highest matching degree is used as the preliminary generated answer, and the generation matrix is ​​dynamically adjusted to generate a feedback matrix; based on the feedback matrix, the generation matrix is ​​optimized and updated to obtain the generated answer matrix.

[0056] After completing the dynamic adjustment of the nested weights, the adjusted weight values ​​work together with the semantic nesting matrix to construct a generation matrix to generate preliminary legal consultation answers; the generation value of each semantic feature is generated through the calculation formula of the generation matrix, thereby expressing the complete legal semantic information of the user's question; the calculation formula of the generation matrix is ​​introduced to progressively generate the semantic information of each layer, so that the generated answer is not only accurate but also has legal level coherence.

[0057] The calculation formula for the generated matrix is ​​as follows:

[0058]

[0059] Among them, G i，j represents the generated value of the jth semantic feature in the i-th layer of the generation matrix, which is used to express the semantic content of the generated answer; W i，l is the weight value of the jth semantic feature of the i-th layer after dynamic adjustment, which indicates the weight value of the l-th layer of nested generation of the i-th layer parsing during the generation process; L is the total number of nested generation levels; κ is the generation offset coefficient, which prevents semantic level deviation during the generation process and ensures the accuracy of the generated solution; |M i，l -T j，l | 2 is the semantic matching error term, which is used to measure the difference between the input semantics and the legal terms; T j，l represents the semantic features of the jth word in the lth layer of the nested generation process in the legal knowledge base matrix; η l Generate adjustments for the levels so that the generated answer can reflect the semantic uniqueness of each level.

[0060] The generated matrix is ​​compared with the legal knowledge base matrix layer by layer to confirm the matching degree between the generated answer and the elements of the legal knowledge base matrix. Through layer-by-layer semantic matching, it is ensured that the generated answer can reflect the legal consultation content at all levels, and finally provide accurate legal consultation answers.

[0061] The formula for nested semantic matching is:

[0062]

[0063]

[0064] Among them, Match i，j represents the matching degree of the jth semantic feature of the i-th layer, which is used to measure the similarity between the generated answer and the elements of the legal knowledge base matrix; K i，j is an element of the legal knowledge base matrix; ε represents the sensitivity parameter of the match, which is used to control the accuracy of the match and is obtained by calculating the mean square error between the generated solution and the elements of the legal knowledge base matrix; In order to smooth the deviation term, avoid the matching value being too large or too small, and ensure the consistency of the matching process, it is obtained by calculating the weighted average of the elements of the matrix of the generated solution and the legal knowledge base; ∈ To fine-tune the parameters, we can further optimize the matching value and make it closer to the actual answer; δ is a positive number used to prevent the denominator from being zero.

[0065] After completing the nested semantic matching, the result with the highest matching degree is used as the preliminary generated answer. The generated value of the generation matrix is ​​dynamically adjusted using the user feedback data. Through adaptive feedback adjustment, the final generated answer is more in line with user needs. The feedback matrix is ​​generated and the user feedback data is mapped to the feedback matrix to optimize the weight values ​​of the generation matrix and the matching matrix, thereby correcting the semantic deviation in the answer. The calculation formula of the feedback matrix is ​​as follows:

[0066]

[0067] Among them, F i，j is an element in the feedback matrix, indicating the correction value of the jth semantic feature of the i-th layer during the feedback adjustment process; ω 1i is the user rating weight of the i-th layer, which represents the user's feedback on the system's answer satisfaction, that is, the user feedback data; i，j is the element in the matching matrix A, which represents the output value of the initial solution generated according to the matching degree in the semantic feature; ω 2i is the weight coefficient of the feedback offset term; F u It is the user's feedback score on the initially generated answer. The generated result of the feedback matrix will be combined with the generation matrix and the matching matrix to perform multiple rounds of feedback optimization to continuously correct the deviation of the generated answer and make the final answer more consistent with the user's needs.

[0068] After calculating the correction value of the feedback adjustment, the correction value in the feedback matrix is ​​used to perform the final optimization update on the generation matrix to ensure that the generated answer better meets the user's needs and improves the accuracy and professionalism of the answer. i，j Applied to each semantic feature in the generated matrix to adjust the details of the generated content and form the final generated answer matrix. The specific formula is:

[0069]

[0070] in, is the element of the generation matrix after feedback optimization, that is, the element in the final generation solution matrix; is the feedback strength coefficient, which is used to control the degree of adjustment of the correction value in the generated matrix.

[0071] The final generated answer matrix expresses the answer content after feedback optimization, converting it into a text form of natural language output to provide to users. The final generated answer will cover legal terms, specific article citations and professional legal interpretations, so that it meets the user's consulting needs in terms of accuracy and professionalism, and realizes high-quality legal consultation intelligent response.

[0072] In summary, a deep learning-based intelligent response method for legal consultation has been completed.

[0073] The order of the embodiments of the invention is for description only and does not represent the advantages and disadvantages of the embodiments. The processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0074] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from other embodiments.

[0075] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions described in the above embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention.

Claims

1. A legal consultation intelligent response method based on deep learning, characterized in that: The following steps are involved: S1. Preprocess the data of the legal consultation questions input by the user and generate a semantic nesting matrix to express the multi-layer semantic features of the legal consultation questions; Optimize semantic features at different levels by dynamically adjusting nesting weights; S2. Based on the adjusted weight values ​​and the semantic embedding matrix, a generation matrix is ​​constructed to generate preliminary legal consultation answers; Compare the generated matrix with the legal knowledge base matrix layer by layer to confirm the matching degree between the generated solution and the elements of the legal knowledge base matrix; The result with the highest matching degree is taken as the initially generated solution, and the generated matrix is ​​dynamically adjusted to generate a feedback matrix; The generation matrix is ​​optimized and updated based on the feedback matrix to obtain the generation solution matrix.

2. The deep learning-based legal consultation intelligent response method according to claim 1 is characterized in that: The S1 specifically includes: The input legal consultation question is decomposed into a vocabulary sequence, and the legal vocabulary and terminology are further identified and marked; each word in the vocabulary sequence is converted into a feature point in a high-dimensional vector space, and then the semantic embedding matrix is ​​obtained through the calculation formula of the semantic embedding matrix.

3. The deep learning-based legal consultation intelligent response method according to claim 2 is characterized in that: The S1 specifically includes: The calculation formula of the semantic embedding matrix is ​​as follows: Among them, M i，j is an element in the semantic embedding matrix, representing the embedding semantic value of the jth word in the i-th layer of the parsing layer; N represents the number of words in the sentence; k is the word ordinal number; φ(Q k ) is the semantic feature vector of the kth word in the user input Q; ψ(T k ) is the feature vector of the kth word in the legal knowledge base matrix; σ is the activation function; γ i is the nesting level offset; θ j is the vocabulary shift coefficient.

4. The deep learning-based legal consultation intelligent response method according to claim 1 is characterized in that: The S1 specifically includes: A dynamic adjustment mechanism is introduced to calculate the nested weights in real time according to the current input content, and the calculated weight values ​​will be used as control variables for generating the matrix.

5. The deep learning-based legal consultation intelligent response method according to claim 4 is characterized in that: The S1 specifically includes: The adjustment formula is: Among them, W i+1，j represents the weight value of the jth semantic feature in the i+1th layer; W i，j is the weight of the jth semantic feature in the i-th layer; α i is the basic adjustment coefficient for the i-th layer; β j is the offset parameter of the jth word; δ p is the offset adjustment term; P is the total number of offset parameters; ΔW is the weight difference; λ is the smoothing parameter; is the control parameter.

6. The deep learning-based legal consultation intelligent response method according to claim 1 is characterized in that: The S2 specifically includes: The calculation formula of the generation matrix is ​​introduced to progressively generate the semantic information of each layer. The specific formula is as follows: Among them, G i，j represents the generated value of the jth semantic feature in the i-th layer of the generator matrix; W i，l is the weight value of the ith semantic feature of the ith layer after dynamic adjustment; L is the total number of nested generated levels; κ is the generated offset coefficient; |M i，l -T j，l | 2 is the semantic matching error term; T j，l represents the semantic features of the jth word in the lth layer of the nested generation process in the legal knowledge base matrix; η l Generate reconciliation items for the hierarchy.

7. The deep learning-based legal consultation intelligent response method according to claim 6 is characterized in that: The S2 specifically includes: Map user feedback data into the feedback matrix, optimize the weight values ​​of the generation matrix and the matching matrix, and correct the semantic deviation in the answer; the calculation formula of the feedback matrix is ​​as follows: Among them, F i，j is an element in the feedback matrix, indicating the correction value of the jth semantic feature in the i-th layer; ω 1i is the user rating weight of the i-th layer, representing user feedback data; A i，j is the element in the matching matrix A, which represents the output value of the initial solution generated according to the matching degree in the semantic feature; ω 2i is the weight coefficient of the feedback offset term; F u is the user's feedback score on the initially generated solution; δ is a positive number used to prevent the denominator from being zero.

8. The deep learning-based legal consultation intelligent response method according to claim 7 is characterized in that: The S2 specifically includes: The correction value is applied to each semantic feature in the generated matrix to form a generated answer matrix; the specific formula is: in, is the element of the generated matrix after feedback optimization, representing the element in the generated solution matrix; is the feedback strength coefficient.

9. The deep learning-based legal consultation intelligent response method according to claim 8 is characterized in that: The S2 specifically includes: The generated answer matrix expresses the answer content after feedback optimization, converts the answer content into a text form of natural language output, and performs intelligent response to legal consultation.

Images