Legal decision result prediction method and system

Abstract

The invention relates to a law decision result prediction method and system, and the method comprises the steps: constructing an intellectual property case keyword dictionary, and obtaining a law text of an intellectual property related case from a first public database through the intellectual property case keyword dictionary; obtaining enterprise economic data corresponding to each legal text from a second public database; constructing a structured topic extraction model, and training the structured topic extraction model by using the legal text and the enterprise economic data; constructing a legal judgment result prediction model, and training the legal judgment result prediction model by using a topic distribution result output by the trained structured topic extraction model; and obtaining a legal text and enterprise economic data of a to-be-predicted object, inputting the legal text and the enterprise economic data into the trained structured topic extraction model to obtain a topic distribution result, and inputting the topic distribution result into the trained legal judgment result prediction model to perform legal judgment result prediction. According to the method, the prediction accuracy and the interpretability of the legal judgment result of the legal text of the intellectual property case can be remarkably improved.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method and system for predicting legal judgment outcomes. Background Technology

[0002] With the rapid development of information technology, the legal field has an increasing demand for automated and intelligent judgment assistance systems. Legal judgment prediction and interpretation systems aim to predict the outcomes of cases to be judged and provide reasonable legal interpretations by analyzing historical case data, thereby assisting judges in making judicial decisions and improving judicial efficiency and consistency. Traditional legal judgments rely on judges' professional knowledge and experience, which is inefficient when dealing with a large number of cases and is susceptible to inconsistencies in judgments due to subjective factors. With the development of artificial intelligence and machine learning technologies, data-driven legal judgment prediction methods have gradually become a research hotspot. These methods mine and learn from historical case data to establish a correlation model between case characteristics and judgment outcomes, thereby enabling the prediction of judgments for new cases.

[0003] The legal system encompasses multiple professional fields, such as intellectual property, civil disputes, and criminal offenses. Among these, with the continuous advancement of legal judgment prediction technology, the prediction and interpretation of judgments in intellectual property litigation are particularly important due to its highly specialized nature and the far-reaching impact of its outcomes. For example, intellectual property litigation is becoming increasingly important for listed companies because it affects not only their current operations but also their strategic direction. In other words, besides assisting judges in judicial decision-making, the high costs associated with litigation and the outcomes of intellectual property litigation make legal judgments an economically significant activity as well. Studying litigation outcomes at the company level can provide a better understanding of how companies protect their intellectual property, how it affects their competitors, and how it sustains innovation. Intellectual property litigation also impacts the expected returns of relevant companies.

[0004] Due to the complexity of predicting legal judgments, especially those related to intellectual property litigation, and the low accuracy of existing methods for predicting legal judgments, higher demands are placed on the prediction of legal judgment outcomes.

[0005] For example, invention patent CN109308355B discloses a method and apparatus for predicting legal judgment results. The method includes: converting first factual information in the legal document to be predicted into a first word vector matrix; converting second factual information in each pre-acquired legal document sample into a second word vector matrix; encoding the first word vector matrix to obtain a first encoding result; encoding each of the second word vector matrices to obtain a second encoding result; and predicting the judgment result of each judgment task of the legal document to be predicted based on a long short-term memory neural network according to the first encoding result, each of the second encoding results, and the judgment results of each pre-acquired legal document sample. The long short-term memory neural network is constructed based on the dependencies between each judgment task. Essentially, this approach converts factual information in legal documents into mathematical vectors, learns the simple statistical correlation between "factual feature combinations" and "judgment results" in historical cases through neural networks (especially LSTMs that capture sequence dependencies), and performs pattern matching and probability inference for new cases.

[0006] For example, patent application CN111178817A discloses a method for obtaining judgment results. It first preprocesses the original case text and converts it, along with legal provisions, into word vectors. Then, it uses a judgment model that integrates bidirectional semantic understanding and attention mechanisms to extract and fuse textual and legal features. Finally, it outputs predictions of the crime, sentence, and fine. This method aims to train a model using historical judgment data to achieve rapid and objective intelligent judgment assistance. However, its essence remains statistical fitting and pattern imitation of historical judgments, lacking genuine logical deduction capabilities. Furthermore, the decision-making process of this method has low interpretability. Summary of the Invention

[0007] The purpose of this invention is to provide a method and system for predicting legal judgment results, which partially solves or alleviates the above-mentioned shortcomings in the prior art, realizes in-depth collaborative analysis of legal text data and economic data, and can significantly improve prediction accuracy and interpretability.

[0008] To solve the aforementioned technical problems, the present invention specifically adopts the following technical solution: A first aspect of the present invention is to provide a method for predicting legal judgment outcomes, comprising the following steps: Construct a keyword dictionary for intellectual property cases, and use the keyword dictionary to obtain legal texts of intellectual property-related cases from a first public database, wherein the legal texts of the intellectual property-related cases include legal judgments; Obtain the enterprise economic data corresponding to the legal text from the second publicly available database; A structured topic extraction model is constructed, and the structured topic extraction model is trained using the legal text and the enterprise economic data to obtain the trained structured topic extraction model. A legal judgment outcome prediction model is constructed, and the topic distribution results output by the trained structured topic extraction model are used to train the legal judgment outcome prediction model to obtain the trained legal judgment outcome prediction model. The process involves acquiring legal texts related to intellectual property cases of the entity to be predicted, as well as the entity's economic data. These legal texts and economic data are then input into a trained structured topic extraction model to extract the topic distribution results of the legal texts. Finally, the topic distribution results are input into a trained legal judgment prediction model to output the predicted judgment result. Note that the legal texts of the entity to be predicted do not include legal judgment results.

[0009] Furthermore, the construction of the intellectual property case keyword dictionary includes: Construct intellectual property seed words, calculate the cosine similarity between words in legal texts and the intellectual property seed words as semantic similarity, and select the top N words with the highest semantic similarity to the intellectual property seed words to form the keyword dictionary for intellectual property cases.

[0010] Furthermore, the intellectual property seed terms include infringement of trade secrets, infringement of patent rights, patent disputes, patent ownership, patent application right ownership, intellectual property, trademark infringement, patent disputes, copyright disputes, and copyright.

[0011] Furthermore, during the training process of the structured topic extraction model, the enterprise economic data and legal judgment results corresponding to each legal text are used as dual supervision signals to jointly constrain the model, so that the structured topic extraction model outputs the corresponding topic distribution results.

[0012] Furthermore, the topic distribution of each legal text is generated through a logistic normal distribution, and the position parameters of the logistic normal distribution are calculated based on the enterprise economic data and legal judgment results corresponding to the legal text. The covariance matrix of the logistic normal distribution is obtained based on the number of topics.

[0013] Furthermore, the enterprise economic data includes at least one or more of the following: leverage ratio, company market capitalization to asset ratio, return on assets, company size, number of years since company establishment, and board size.

[0014] Furthermore, before constructing the structured topic extraction model, the legal text is preprocessed, including: The legal text was segmented using a Chinese word segmentation tool to obtain a word sequence; The word sequence is filtered according to a preset stop word list to remove stop words without substantial semantic meaning, resulting in a preprocessed text word set. The stop word list includes general Chinese stop words and stop words specific to legal texts.

[0015] Furthermore, it also includes the following steps: Obtain the posterior distribution of the topic distribution results of legal texts from the trained structured topic extraction model; Multiple samples were drawn from the posterior distribution of the topic distribution results of legal texts using a combined sampling method. For both winning and losing judgments, calculate the conditional expectation value of the proportion of each topic in each sample group; By comparing the average proportion of the same topic in the winning and losing samples, the correlation strength between the corresponding topic and the judgment result can be obtained; The topics are risk-labeled based on the strength of their correlation with the judgment results.

[0016] Secondly, this application also discloses a legal judgment outcome prediction system, the system comprising: The dataset acquisition module is configured to construct a keyword dictionary for intellectual property cases, and uses the keyword dictionary to obtain legal texts of intellectual property-related cases from a first public database. The legal texts of the intellectual property-related cases include: legal judgment results. The dataset filtering module is configured to retrieve enterprise economic data corresponding to the legal text from a second public database. A structured topic extraction model construction module is configured to construct a structured topic extraction model, and to train the structured topic extraction model using the legal text and the enterprise economic data to obtain the trained structured topic extraction model; preferably, during the training process of the structured topic extraction model, the enterprise economic data corresponding to each legal text and the legal judgment result are used as dual supervision signals to jointly constrain the generation of topic distribution; The legal judgment result prediction model construction module is configured to construct a legal judgment result prediction model, and uses the topic distribution results output by the trained structured topic extraction model to train the legal judgment result prediction model to obtain the trained legal judgment result prediction model. The prediction module is configured to acquire legal texts and corporate economic data related to intellectual property cases of the object to be predicted. The legal texts and corporate economic data are input into a trained structured topic extraction model to extract the topic distribution results of the legal texts. The topic distribution results of the legal texts are then input into a trained legal judgment result prediction model to output the judgment prediction results.

[0017] Furthermore, the dataset acquisition module's construction of the intellectual property case keyword dictionary specifically includes constructing intellectual property seed words, calculating the cosine similarity between words in the legal text and the intellectual property seed words as semantic similarity, and selecting the top N words with the highest semantic similarity to the intellectual property seed words to constitute the intellectual property case keyword dictionary. The intellectual property seed words include infringement of trade secrets, infringement of patent rights, patent disputes, patent ownership, patent application right ownership, intellectual property, trademark infringement, patent disputes, copyright disputes, and copyright.

[0018] Furthermore, the system also includes a preprocessing module, configured to use a Chinese word segmentation tool to segment the legal text to obtain a word sequence; and to filter the word sequence according to a preset stop word list to remove stop words without substantive meaning, thereby obtaining a preprocessed text word set.

[0019] Furthermore, the system also includes: a risk labeling module, configured to obtain the posterior distribution of legal text topic distribution from the trained structured topic extraction model; extract multiple sets of samples from the posterior distribution of legal text topic distribution using a combined sampling method; calculate the conditional expectation value of the proportion of each topic in each set of samples for both high-risk and low-risk judgment conditions; compare the average proportion of the same topic in high-risk and low-risk samples to obtain the correlation strength between the corresponding topic and the judgment result; and label the topic with risk based on the correlation strength between the corresponding topic and the judgment result.

[0020] Beneficial technical effects: 1. Traditional judgment prediction primarily relies on natural language processing of legal texts (including court judgments, indictments, etc.), neglecting the implicit correlation between corporate economic data (or financial data) and legal judgment outcomes. For example, two patent infringement cases with similar facts may have drastically different legal judgments due to differences in the defendant companies' financial situations (such as cash flow difficulties potentially leading to settlements). Therefore, to improve the accuracy of judgment prediction, this application constructs a cross-modal joint modeling framework for legal texts and corporate financial data (or economic data). Through the supervisory constraint mechanism in the structured topic model, unstructured legal texts such as court judgments (including legal data) and structured economic data (including financial indicators) are uniformly represented at the semantic level, achieving deep collaborative analysis of judicial data and economic data. Therefore, compared with traditional judgment prediction, it can significantly improve prediction accuracy and interpretability, and provide reliable quantitative decision support data for intellectual property litigation.

[0021] 2. The model in this application learns both textual semantic relationships and financial-judgment statistical patterns during training. That is, by capturing the more fundamental "topic-economic / financial-judgment" relationship, rather than superficial statistical correlations, each topic ultimately contains a dual description of legal attributes and related financial attributes. This gives the machine learning features business readability, and compared to traditional methods, the model has a certain reasoning ability for novel cases. On the other hand, by constructing a structured topic extraction model constrained by economic data, economic data is introduced as a priori constraint into the topic generation process, solving the semantic drift problem of traditional topic models in legal-financial cross-domain scenarios.

[0022] 3. This application does not perform feature concatenation at the end of the model, but rather integrates features at the source of feature generation (at the topic model level). It unifies unstructured text and high-dimensional structured data (i.e., economic data) into a low-dimensional, dense, and semantically rich topic space. This space naturally becomes a bridge connecting text and finance, enabling machine learning models to learn more efficiently. For example, compared to simply concatenating textual and financial features and directly inputting them into a machine learning model, the existence of a "semantic gap"—that is, the high-dimensional sparse features of text and the low-dimensional dense features of finance—make it difficult for the model to effectively coordinate and learn the deep-seated relationship between the two.

[0023] 4. The "topic-economic / financial-judgment" relationship learned by the model in this application is a universal knowledge. It can be applied to predict new cases and to create intellectual property risk profiles for companies. By combining a company's financial data, it is possible to infer which types of topics will stand out in potential litigation, thereby helping companies assess their risk exposure. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. The elements or parts in the drawings are not necessarily drawn to scale. Obviously, the drawings described below are some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without any creative effort.

[0025] Figure 1 This is a flowchart of a method for predicting legal judgment outcomes according to this application.

[0026] Figure 2 This is a schematic diagram of an STM topic model in one embodiment of this application.

[0027] Figure 3This is a schematic diagram comparing the performance of different text analysis methods on various machine learning models in one embodiment of this application.

[0028] Figure 4 This is a schematic diagram comparing the accuracy of different text analysis methods on various machine learning models in one embodiment of this application.

[0029] Figure 5 This is a schematic diagram of the module structure of a legal judgment result prediction system in one embodiment of this application.

[0030] Figure 6 This is a schematic diagram showing different themes and their corresponding keyword groups obtained by processing several legal texts using the STM theme model of this application.

[0031] Figure 7 This is a schematic diagram illustrating topics that are easier to win and easier to lose, obtained by processing several legal texts using the STM topic model of this application. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0033] In this document, suffixes such as “module,” “part,” or “unit” used to denote elements are used only for illustrative purposes and have no specific meaning in themselves. Therefore, “module,” “part,” or “unit” can be used interchangeably.

[0034] In this document, the terms "upper," "lower," "inner," "outer," "front," "rear," "one end," and "the other end," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0035] In this document, unless otherwise explicitly specified and limited, the terms "installed," "equipped with," and "connected," etc., should be interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection, a direct connection, or an indirect connection through an intermediate medium; it can be a connection within two components. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0036] In this document, "and / or" includes any and all combinations of one or more of the listed related items.

[0037] In this article, "multiple" means two or more, that is, it includes two, three, four, five, etc.

[0038] Figure 1 A flowchart of a legal judgment outcome prediction method according to this application is shown, with reference to... Figure 1 The method specifically includes the following steps: S1: Construct a keyword dictionary for intellectual property cases, and use the keyword dictionary to obtain the legal texts of all intellectual property-related cases from the first public database.

[0039] In some embodiments, the legal text of the intellectual property-related case includes: legal facts, reasoning for the judgment, and legal rulings, etc. For example, the legal text includes publicly available court judgments, as well as procedural documents (such as procedural documents and relevant evidence), etc.

[0040] Using tools such as Python, legal texts, such as court judgments, are filtered. A keyword dictionary for intellectual property cases is constructed using word vectors (Word2Vec). All texts are then filtered to collect legal texts related to intellectual property (such as court judgments related to intellectual property).

[0041] Word vector models are deep learning models in the field of natural language processing trained using word embedding techniques combined with massive amounts of text data. They embed high-dimensional text data into a low-dimensional vector space by statistically analyzing the occurrence of other words in the context of a word, resulting in each word having a corresponding 200 or 300-dimensional vector to represent its semantics. Constructing a keyword dictionary for intellectual property cases first requires obtaining a series of seed words. This application extracted 10 seed words, including infringement of trade secrets, infringement of patent rights, patent disputes, patent ownership, patent application right ownership, intellectual property, trademark infringement, patent disputes, copyright disputes, and copyright.

[0042] Furthermore, this paper uses the cosine similarity algorithm to calculate the word vector distance between all words and the seed word in legal texts such as court judgments, in order to determine the similarity of each word to the seed word. Based on the similarity data, this paper retains the top 100 words that are most similar to the seed word.

[0043] in, and represent and The first and second dimensional vectors of each n-dimensional vector Each component. This paper uses the above formula to calculate the similarity between word vectors of two words, thereby obtaining the semantic similarity between the two words.

[0044] S2: Obtain the enterprise economic data corresponding to each legal text from the second public database.

[0045] In some embodiments, since each legal text contains information about the parties involved in the legal event (e.g., infringers and patent holders, and the economic data of these parties can be obtained from a second public database), the corresponding corporate economic data can be obtained from the second database based on the information of these parties, such as company names and listed company codes (which can also be obtained from the internet based on company names). Specifically, corporate economic data includes publicly available financial information (such as various indicators in profit and loss statements, balance sheets, and cash flow statements), market valuation data (such as Tobin's Q, price-to-earnings ratio, price-to-book ratio, and other market pricing indicators), corporate governance structure data (such as board size, proportion of independent directors, and equity concentration), R&D and innovation investment data (such as R&D expenditure ratio, number of patents held, and proportion of technical personnel), relative industry performance data (such as industry ranking, market share, and industry growth rate comparison), and macroeconomic and industry environment data (such as the strength of policy support for the industry, tax incentives, and industry prosperity index). These multi-dimensional economic data collectively constitute a comprehensive economic profile of the enterprise, providing rich non-textual feature information for predicting intellectual property litigation judgments.

[0046] For example, corporate economic data mainly includes one or more of the following: leverage ratio, market capitalization to asset ratio (Tobin's Q), return on assets, company size, years since company establishment, and board size.

[0047] In some embodiments, the online court provides judgment documents from various fields. The judgment documents related to intellectual property are filtered out using a keyword dictionary for intellectual property cases. Python is then used to identify all intellectual property cases involving listed companies (i.e., parties to legal events) or companies with publicly available financial information (i.e., parties to legal events). The judgment documents corresponding to these intellectual property cases with existing judgments are then used as training samples for subsequent training processes.

[0048] In some embodiments, publicly available financial information can be matched from a second public database (e.g., the CSMAR database) based on information such as the listed company code.

[0049] In some embodiments, legal texts are preprocessed. Specifically, preprocessing is performed using jieba word segmentation, which segments the Chinese text in conjunction with context, removing stop words to eliminate useless information. For example, firstly, a Chinese word segmentation tool is used in conjunction with a legal professional dictionary and a business economic terminology dictionary to perform multi-granular word segmentation on the legal text, ensuring that compound terms such as company names, financial indicators (e.g., "asset-liability ratio"), and legal concepts (e.g., "unfair competition") are correctly identified as complete units, generating a structured word sequence. Subsequently, this word sequence is intelligently filtered according to a pre-set multi-level stop word list. The word list obtained after preprocessing the legal text not only includes general meaningless function words, but also specifically identifies and removes procedural expressions (e.g., "this court believes" and "the plaintiff claims"), format symbols, and redundant information unrelated to judgment prediction in the legal text, ultimately obtaining a clean and semantically rich text word set, laying a high-quality data foundation for subsequent topic modeling and multimodal feature fusion.

[0050] In some embodiments, after segmenting and removing stop words from the document (i.e., legal text), m is obtained by statistically analyzing the frequency of word v in the entire corpus, followed by Laplace smoothing and logarithmic transformation. v , as the basic frequency prior of words.

[0051] S3: Construct a structured topic extraction model and train the structured topic extraction model using legal texts and corporate economic data to obtain the trained structured topic extraction model.

[0052] Typically, multiple themes can be extracted from a single legal text, and different legal texts may contain the same themes or different themes. Therefore, each theme extracted from all legal texts may correspond to the same one or more economic data points of the same company, or it may point to economic data of different companies or different categories.

[0053] In some embodiments, a STM (Structural Topic Model) is used for topic extraction. Preferably, the optimal number of topics can be determined using the residual method. When using STM for topic extraction, constraints are imposed by the financial indicators (X) and legal judgment results (y) of listed companies, making each topic an interpretable "legal-economic" basic unit. For example, a topic is essentially a statistical prototype of "how a company with certain financial characteristics (X) will be expressed (vocabulary) in legal texts and what kind of judgment result (y) it tends to lead to," and the topic distribution results are calculated.

[0054] Figure 2 An example of an STM topic model: D: A series of judgment documents input, including multiple texts d, each text d containing several words; X: Financial metrics, including leverage ratio, Tobin's Q, return on assets, company size, years since company founding, and board size; Γ: A parameter matrix related to financial indicators (i.e., economic data); y: Legal judgment result: win or lose; Define a ( K 1)×( K 1) The covariance matrix, used for Logistic Normal distribution, where K is the number of topics; For each text d, generate its topic distribution. This distribution follows a Logistic Normal distribution, and is composed of the parameter matrix Γ related to the financial indicators and the financial indicator X. d The sum and covariance matrix Σ determine this; .

[0055] That is, for a judgment document d, the model outputs It is a probability vector of length K (number of topics), where each element represents the proportion by which a particular judgment document d was "generated" or "interpreted" by a certain topic. That is... It is a low-dimensional dense vector that has already incorporated economic data. Therefore, when it is used as input to a subsequent prediction model, the prediction effect will be much better than splicing the original features.

[0056] For example, assuming K=5, a judgment concerning "patent infringement by a high-tech company" might have the following thematic distribution: =[0.45,0.35,0.15,0.05,0.00], which means that 45% of the content of this document can be explained by topic A, 35% of the content is related to topic B, 15% is related to topic C, and so on.

[0057] Other parameters include: For each word n ​​in each text d, based on the topic distribution Generate topic tags for words The label follows a multinomial distribution; .

[0058] β is a tensor (D*K*V) representing the probability distribution of words v in a text d on topic k (e.g., certain legal or economic terms are more likely to appear under topic k), where each element... The calculation formula is as follows: .

[0059] Where V represents the total number of words in the vocabulary, v represents the word index in the vocabulary, and v=1,…,V; K represents the total number of topics, k represents the topic index, and k=1,…,K; m v This represents the prior probability of word v, that is, the probability of word v appearing in all documents; , , Run the variational EM algorithm on the STM model until convergence to obtain the model parameters; This represents the topic-word association parameter between topic k and word v; Indicates the legal judgment result. d The parameters relating to the word v; Indicates the legal judgment result. d The correlation parameters between topic k and word v; Based on the topic tags of the words and topic-word distribution Generate the word that actually appears at the nth position in the text d. It follows a multinomial distribution; in, This represents a vector of length V with document d and topic k fixed for β.

[0060] In some embodiments, further results are presented based on the STM model, including the distribution of topic keywords and low-risk topics (such as topics that are easier to win) or high-risk topics (such as topics that are easier to lose), increasing interpretability.

[0061] The distribution of thematic keywords is as follows: After the STM model is trained, the covariates (financial indicators, judgment results) are fixed at the reference level (using empirical values), and substituted into the original... The calculation formula, after normalization, yields a global topic-word baseline distribution independent of the legal text. This distribution, stripped of the economic and judgment context of specific cases, represents the core semantic structure of each topic and is used to generate topic keywords and conduct interpretability analysis.

[0062] The topics identified as low-risk (e.g., easier to win) or high-risk (e.g., easier to lose) are obtained through the following steps: After the STM model is trained, the estimated topic distribution will be obtained. and its posterior distribution, and then a combined sampling method was used to select from... A large sample is drawn from the posterior distribution, and the conditional expected value of the proportion of each topic under both low-risk (e.g., "winning") and high-risk (e.g., "losing") conditions is calculated. By comparing the average proportion of the same topic in winning and losing samples, the correlation strength between the topic and the judgment outcome is quantified, thereby identifying low-risk topics (e.g., topics prone to winning) and high-risk topics (e.g., topics prone to losing). For example, if the average proportion of a topic in losing cases is significantly higher than in winning cases, then that topic will be marked as a high-risk topic, as well as a topic more likely to lose. See [link to relevant documentation]. Figure 7 This provides users with statistically based decision-making insights.

[0063] S4: Construct a legal judgment outcome prediction model, and use the topic distribution results output by the trained structured topic extraction model to train the legal judgment outcome prediction model to obtain the trained legal judgment outcome prediction model.

[0064] In some embodiments, STM-based topic distribution results As input, the model is trained using a legal judgment outcome prediction model (such as random forest, support vector machine, deep neural network, XGBoost, etc.) to predict the legal judgment outcome of intellectual property cases.

[0065] In some embodiments, the loss function used in the legal judgment outcome prediction model is: ; .

[0066] Where N is the total number of samples; y i z represents the true judgment result of the i-th sample (value 1 for a win, 0 for a loss); i σ is the original predicted value calculated by the judgment result prediction model for the i-th sample (similarly, if the prediction is a win (e.g., the probability of losing is less than the preset threshold), the value is 1; if the prediction is a loss, the value is 0); σ is the activation function.

[0067] The essence of STM is the EM algorithm, which lacks a loss function. This application introduces a financially and judicially constrained STM structured topic model and deeply couples it with a downstream machine learning classifier, achieving accurate and interpretable prediction of intellectual property litigation risks.

[0068] S5: Obtain the legal texts of all intellectual property-related cases of the object to be predicted (i.e., legal texts excluding judgment results, such as legal texts for which judgment results have not yet been issued), as well as the enterprise economic data of the object to be predicted. Input the legal texts and the enterprise economic data into the trained structured topic extraction model to extract the topic distribution results of each legal text. Then, input the topic distribution results of each legal text into the trained legal judgment result prediction model to output the judgment prediction results.

[0069] For example, the predicted judgment result includes the probability of losing a certain intellectual property case. Preferably, if the probability of losing is greater than a preset probability threshold, it means that the object to be predicted is more likely to lose in intellectual property cases under the corresponding legal text (furthermore, relevant topics that make the case more likely to be lost can also be output); if the probability of losing is less than the preset probability threshold, it means that the object to be predicted is more likely to win in intellectual property cases under the corresponding legal text (furthermore, relevant topics that make the case more likely to be won can also be output).

[0070] In other embodiments, in step S5 above, in addition to the topic distribution results mentioned above, in order to further improve the prediction accuracy, the corresponding enterprise economic data can also be input into the legal judgment result prediction model for prediction (correspondingly, in the training process of step S4, the topic distribution results and the corresponding enterprise economic data are also input into the legal judgment result prediction model for training).

[0071] To verify that the model in this application has better predictive performance, this application conducts model performance tests based on different machine learning models and text analysis methods, such as... Figure 3 and Figure 4 As shown, Figure 3 This paper presents a comparison of the prediction accuracy and AUC values ​​of different text analysis methods on various machine learning models. Figure 4This paper presents a comparison of the prediction accuracy of different text analysis methods across various machine learning models. The text analysis methods include Structured Topic Modeling (STM), Latent Dirichlet Allocation (LDA), and Word Freq. The machine learning models compared include Logistic Regression (LR), Random Forest (RF), Support Vector Machine (SVM), XGBoost, and Deep Neural Networks (DNN). In terms of accuracy, the STM method performs best across most models, especially achieving a maximum accuracy of 84.60% in the DNN model. The accuracy of the LDA method fluctuates significantly across different models, while the accuracy of the Word Freq method is relatively low. Regarding AUC, STM also performs well across most models, achieving an AUC of 0.838 in the DNN model, while LDA has a relatively low AUC across all models, with an AUC of 0.657 in the DNN model. These results indicate that the STM method is more effective at capturing the structured information of text data, thus improving prediction accuracy and AUC.

[0072] This application employs a structured topic model, using economic data and legal judgments as training constraints. This forces the model to learn topic distributions that are correlated with specific financial characteristics (such as high R&D investment and low profit margins) and final legal judgments (such as victory, defeat, and damages). The resulting "topics" are no longer simply unsupervised text clusters, but rather contain a hybrid of "legal semantics" and "financial attributes." For example, it might automatically learn a topic highly relevant to "patent infringement evidence collection by companies with high R&D investment."

[0073] The advantages of this application are: 1) Higher prediction accuracy: The topic distribution obtained through constrained learning is a feature tailored to the final prediction task. It filters out noise in the text that is irrelevant to finance and judgment, and strengthens the discriminative signal. Therefore, the machine learning model using this as input feature can more accurately capture the key factors affecting the outcome of legal judgments, thus achieving higher prediction accuracy than traditional features (e.g., single text features irrelevant to financial characteristics). 2) Highly interpretable features: Themes are interpretable: Each theme consists of a set of related theme keywords (see...). Figure 6Furthermore, due to the constraints of financial data, this application can clearly explain the commercial and legal implications of each topic in predictive interpretability. When predicting a case, this application can analyze which "topics" contribute the most to the prediction results. For example, if the model predicts that a company is likely to lose a case (e.g., the probability of losing is greater than a preset probability threshold), it is mainly because it scores highly on the topic of "difficulty in providing evidence due to insufficient R&D investment." This provides unprecedented decision-making insights for legal professionals, investors, and managers. 3) It solves the semantic gap problem and achieves deep information fusion: This application does not perform feature splicing at the end of the model, but rather performs fusion at the source of feature generation (at the topic model level). It unifies unstructured text and high-dimensional structured data into a low-dimensional, dense, and semantically rich topic space. This space naturally becomes a bridge connecting text and finance, enabling machine learning models to learn more efficiently. 4) It has good generalization ability and the ability to discover new knowledge: The "topic-finance-judgment" relationship learned by the model is a universal knowledge. It can be applied to predict new cases and to create intellectual property risk profiles for companies. For example, by combining a company’s financial data, it is possible to infer which types of themes will stand out in all potential intellectual property litigation (such as which themes are more likely to be lost and which themes are more likely to be won), thereby assessing its risk exposure.

[0074] Further reference Figure 5 As an implementation of the above-described method, this application provides an embodiment of a legal judgment outcome prediction system, which is similar to... Figure 1 Corresponding to the method embodiments shown, the system can be specifically applied to various electronic devices.

[0075] refer to Figure 5 A legal judgment outcome prediction system, comprising: The dataset acquisition module 101 is configured to construct a keyword dictionary for intellectual property cases. It uses this keyword dictionary to acquire legal texts of intellectual property-related cases from a first public database (such as China Judgments Online), and these legal texts include legal judgments. Preferably, the dataset acquisition module 101 is specifically configured to pre-construct intellectual property seed words, calculate the cosine similarity between words in the legal text and the intellectual property seed words as semantic similarity, and select the top N words with the highest semantic similarity to the intellectual property seed words to form the keyword dictionary for intellectual property cases. The intellectual property seed words include infringement of trade secrets, infringement of patent rights, patent disputes, patent ownership, patent application right ownership, intellectual property, trademark infringement, patent disputes, copyright disputes, and copyright. The dataset filtering module 102 is configured to obtain corporate economic data corresponding to each legal text from the second public database; wherein, the corporate economic data includes at least one or more of the following: leverage ratio, company market value to asset ratio, return on assets, company size, company establishment years and board size; these corporate economic data are all publicly disclosed economic data of listed companies or companies with publicly available financial information. The structured topic extraction model construction module 103 is configured to construct a structured topic extraction model. The model is trained using the legal text and enterprise economic data to obtain the trained structured topic extraction model. Preferably, during the training process of the structured topic extraction model, the enterprise economic data and legal judgment results corresponding to each legal text are used as dual supervision signals to jointly constrain the generation of the topic distribution. Further, the topic distribution result of each legal text is generated through a logistic normal distribution, and the position parameters of the logistic normal distribution are calculated based on the enterprise economic data and legal judgment results corresponding to the legal text. The covariance matrix of the logistic normal distribution is obtained based on the number of topics. The legal judgment result prediction model construction module 104 is configured to construct a legal judgment result prediction model and use the topic distribution results output by the trained structured topic extraction model to train the legal judgment result prediction model to obtain the trained legal judgment result prediction model. The prediction module 105 is configured to obtain legal texts related to intellectual property cases of the object to be predicted, as well as the enterprise economic data of the object to be predicted. The legal texts and enterprise economic data are input into the trained structured topic extraction model to extract the topic distribution results of the legal texts, and the topic distribution results are input into the trained legal judgment result prediction model to output the judgment prediction results.

[0076] The system in this embodiment also includes a preprocessing module, configured to preprocess the legal text before constructing the structured topic extraction model. Specifically, the legal text is segmented using a Chinese word segmentation tool to obtain a word sequence; the word sequence is then filtered according to a preset stop word list to remove stop words without substantive meaning, resulting in a preprocessed text word set; wherein, the stop word list includes general Chinese stop words and legal text-specific stop words.

[0077] Furthermore, the system also includes: a risk labeling module, configured to obtain the posterior distribution of the topic distribution results of legal text from the trained structured topic extraction model; extract multiple sets of samples from the posterior distribution of the topic distribution results of legal text using a combined sampling method; calculate the conditional expectation value of the proportion of each topic in each set of samples for both winning and losing judgment conditions; compare the average proportion of the same topic in winning and losing samples to obtain the correlation strength between the corresponding topic and the judgment result; and label the topic with risk based on the correlation strength between the corresponding topic and the judgment result.

[0078] In another aspect, this application also provides a computer-readable storage medium, which may be included in the electronic device described in the above embodiments; or it may exist independently and not assembled into the electronic device. The aforementioned computer-readable storage medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to perform the following... Figure 1 The method shown.

[0079] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0080] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes several instructions to cause a computer terminal (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0081] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims. All of these forms are within the protection scope of this application.

Claims

1. A method for predicting legal judgment outcomes, characterized in that, Includes the following steps: Construct a keyword dictionary for intellectual property cases, and use the keyword dictionary to obtain legal texts of intellectual property-related cases from a first public database, wherein the legal texts of the intellectual property-related cases include legal judgments; Obtain the enterprise economic data corresponding to the legal text from the second publicly available database; A structured topic extraction model is constructed, and the structured topic extraction model is trained using the legal text and the enterprise economic data to obtain the trained structured topic extraction model. During the training process, the enterprise economic data corresponding to each legal text and the legal judgment result are used as dual supervision signals to jointly constrain the structured topic extraction model, so that the structured topic extraction model outputs the corresponding topic distribution results. A legal judgment outcome prediction model is constructed, and the topic distribution results output by the trained structured topic extraction model are used to train the legal judgment outcome prediction model to obtain the trained legal judgment outcome prediction model. Obtain legal texts and corporate economic data related to intellectual property cases of the object to be predicted. Input the legal texts and corporate economic data into a trained structured topic extraction model to extract the topic distribution results for each legal text. Input the topic distribution results into a trained legal judgment result prediction model to output the judgment prediction results.

2. The method for predicting legal judgment outcomes according to claim 1, characterized in that: The constructed keyword dictionary for intellectual property cases includes: Construct intellectual property seed words, calculate the cosine similarity between words in legal texts and the intellectual property seed words as semantic similarity, and select the top N words with the highest semantic similarity to the intellectual property seed words to form the keyword dictionary for intellectual property cases.

3. The method for predicting legal judgment outcomes according to claim 2, characterized in that: The intellectual property seed terms include infringement of trade secrets, infringement of patent rights, patent disputes, patent ownership, patent application right ownership, intellectual property, trademark infringement, patent disputes, copyright disputes, and copyright.

4. The method for predicting legal judgment outcomes according to claim 1, characterized in that: The topic distribution results for each legal text are generated using a logistic normal distribution. The location parameters of the logistic normal distribution are calculated based on the enterprise economic data and legal judgment results corresponding to the legal text. The covariance matrix of the logistic normal distribution is obtained based on the number of topics.

5. A method for predicting legal judgment outcomes according to any one of claims 1-4, characterized in that: The enterprise economic data mentioned above includes at least one or more of the following: leverage ratio, company market capitalization to asset ratio, return on assets, company size, number of years since company establishment, and board size.

6. A method for predicting legal judgment outcomes according to any one of claims 1-4, characterized in that: Before constructing the structured topic extraction model, the legal text is preprocessed, including: The legal text was segmented using a Chinese word segmentation tool to obtain a word sequence; The word sequence is filtered according to a preset stop word list to remove stop words without substantial semantic meaning, resulting in a preprocessed text word set.

7. A method for predicting legal judgment outcomes according to any one of claims 1-4, characterized in that, The companies mentioned include listed companies or companies whose financial information is publicly available.

8. A method for predicting legal judgment outcomes according to any one of claims 1-4, characterized in that, It also includes the following steps: Obtain the posterior distribution of document topic distribution from the trained structured topic extraction model; Multiple samples are drawn from the posterior distribution of the document topic distribution using a combined sampling method. For both winning and losing judgments, calculate the conditional expectation value of the proportion of each topic in each sample group; By comparing the average proportion of the same topic in the winning and losing samples, the correlation strength between the corresponding topic and the judgment result can be obtained; The topics are risk-labeled based on the strength of their correlation with the judgment results.

9. A legal judgment outcome prediction system, characterized in that, The system includes: The dataset acquisition module is configured to construct a keyword dictionary for intellectual property cases, and uses the keyword dictionary to obtain legal texts of intellectual property-related cases from a first public database, wherein the legal texts of the intellectual property-related cases include legal judgment results; The dataset filtering module is configured to retrieve enterprise economic data corresponding to the legal text from a second public database. A structured topic extraction model construction module is configured to build a structured topic extraction model. The structured topic extraction model is trained using the legal text to obtain the trained structured topic extraction model. During the training process, the enterprise economic data and legal judgment results corresponding to each legal text are used as dual supervision signals to jointly constrain the generation of topic distribution. The legal judgment result prediction model construction module is configured to construct a legal judgment result prediction model, and uses the topic distribution results output by the trained structured topic extraction model to train the legal judgment result prediction model to obtain the trained legal judgment result prediction model. The prediction module is configured to acquire legal texts related to intellectual property cases of the object to be predicted, as well as the enterprise economic data of the object to be predicted. The legal texts and the enterprise economic data are input into a trained structured topic extraction model to extract the topic distribution results of each legal text. The topic distribution results of the legal texts are then input into a trained legal judgment result prediction model to output the judgment prediction results.

10. A legal judgment outcome prediction system according to claim 9, characterized in that, The topic distribution results for each legal text are generated using a logistic normal distribution. The location parameters of the logistic normal distribution are calculated based on the enterprise economic data and legal judgment results corresponding to the legal text. The covariance matrix of the logistic normal distribution is obtained based on the number of topics.

Images