A drug combination patent intelligent identification method and device based on a large language model

Abstract

This invention discloses a method and apparatus for intelligent identification of drug combination patents based on a large language model. The method includes: acquiring a standard dataset based on drug combination annotation information from patent data; expanding the standard dataset to generate an initial training set; using efficient parameter fine-tuning technology, training a base large language model using the initial training set to generate an initial recognition model; using the initial recognition model to screen difficult-to-identify samples from unannotated patent data and adding them to the initial training set to form an enhanced training set; using the enhanced training set, iteratively training the initial recognition model using a multi-task learning framework combined with a contrastive learning mechanism to obtain a drug combination patent identification model; inputting the patent data to be identified into the drug combination patent identification model to obtain the drug combination information identification result. This invention can significantly improve the accuracy, data efficiency, and model interpretability of drug combination patent identification, while reducing computational resource requirements and manual review costs.

Description

Technical Field

[0001] This application relates to the interdisciplinary field of artificial intelligence and pharmaceutical informatics, specifically to a method and device for intelligent identification of drug combination patents based on a large language model. Background Technology

[0002] Drug combination patents hold significant strategic value in the pharmaceutical industry. Statistics show that approximately 35% of new drug approvals involve combination therapies, particularly in complex disease areas such as oncology, cardiovascular disease, and anti-infectives. Therefore, accurately and efficiently identifying drug combination patents is crucial for companies to mitigate patent risks, identify R&D opportunities, and formulate competitive strategies.

[0003] Traditional patent classification and retrieval primarily rely on the International Patent Classification (IPC) or the Collaborative Patent Classification (CPC) systems. However, these systems are based on broad classifications of technical fields and lack the ability to finely differentiate between specific innovation models such as drug combination therapy. For example, classification number A61K31 / 00 covers all pharmaceutical compositions containing organic active ingredients, but it cannot distinguish between monotherapy and combination formulations.

[0004] In recent years, commercial patent analysis platforms have begun to incorporate machine learning technologies, primarily employing traditional Natural Language Processing (NLP) methods, such as TF-IDF feature extraction combined with Support Vector Machines (SVM) or Random Forests. Since 2023, some systems have attempted to use pre-trained language models like BERT, but they still face significant technical bottlenecks in the vertical task of identifying drug combination patents.

[0005] First, the model's understanding ability is insufficient: existing methods mainly rely on keyword matching and shallow semantic analysis, making it difficult to accurately understand complex pharmaceutical concepts in patent texts. Traditional methods achieve an accuracy rate of less than 60% in recognizing implicit expressions. Second, there is a shortage of training data: manual annotation of drug combination patents requires deep pharmaceutical expertise, resulting in extremely high annotation costs. Existing public datasets such as USPTO-500k lack fine-grained annotations for drug combinations, making it impossible to support the training of high-performance models. Third, there are numerous ambiguous boundary cases between drug combination patents and related drug formulation improvement patents and dosing regimen optimization patents, leading to a misclassification rate as high as 45% for existing methods.

[0006] Therefore, how to accurately identify and finely classify drug combination patents is a key technical problem that urgently needs to be solved in this field. Summary of the Invention

[0007] This application provides a method and apparatus for intelligent identification of drug combination patents based on a large language model, which can solve the technical problems of insufficient accuracy, poor generalization ability, inability to subdivide combination types, and high cost of manual examination in existing patent classification tools for drug combination identification.

[0008] In a first aspect, embodiments of this application provide a method for intelligent identification of drug combination patents based on a large language model, the method comprising:

[0009] A standard dataset is established based on the drug combination labeling information of patent data, and the standard dataset is augmented to generate an initial training set;

[0010] Using a parameter-efficient fine-tuning technique, the initial training set is used to train the base large language model to generate an initial recognition model;

[0011] The initial recognition model is used to screen out difficult-to-identify samples from unlabeled patent data and add them to the initial training set to form an enhanced training set;

[0012] Using the enhanced training set, the initial identification model is iteratively trained using a multi-task learning framework combined with a contrastive learning mechanism to obtain a drug combination patent identification model.

[0013] Input the patent data to be identified into the drug combination patent identification model to obtain the drug combination information identification result.

[0014] Secondly, embodiments of this application provide a smart recognition device for drug combination patents based on a large language model, the smart recognition device for drug combination patents based on a large language model includes:

[0015] The generation module is used to obtain a standard dataset based on drug combination labeling information from patent data, and to expand the standard dataset to generate an initial training set.

[0016] The fine-tuning module is used to train the base large language model using the initial training set by employing parameter-efficient fine-tuning techniques to generate an initial recognition model.

[0017] The filtering module is used to filter out difficult-to-identify samples from unlabeled patent data using the initial identification model and add them to the initial training set to form an enhanced training set.

[0018] The training module is used to iteratively train the initial recognition model using the enhanced training set, employing a multi-task learning framework combined with a contrastive learning mechanism, to obtain a drug combination patent recognition model.

[0019] The identification module is used to input the patent data to be identified into the drug combination patent identification model to obtain the drug combination information identification result.

[0020] The beneficial effects of the technical solutions provided in this application include:

[0021] A standard dataset is established by acquiring drug combination labeling information based on patent data, and the standard dataset is augmented to generate an initial training set. A base large language model is trained using the initial training set using efficient parameter fine-tuning techniques to generate an initial recognition model. Difficult-to-identify samples are screened from unlabeled patent data using the initial recognition model and added to the initial training set to form an enhanced training set. Using the enhanced training set, a multi-task learning framework combined with a contrastive learning mechanism is used to iteratively train the initial recognition model to obtain a drug combination patent recognition model. The patent data to be identified is input into the drug combination patent recognition model to obtain the drug combination information recognition result.

[0022] This application enables the model to efficiently and accurately identify combination drug regimens from patent data by finely annotating drug combination information, and further accurately determine the type of combination. By introducing a contrastive learning mechanism and a hard sample mining strategy, it achieves a significant improvement in the ability to distinguish boundary samples, effectively solving the problems of insufficient accuracy, poor generalization ability, inability to subdivide combination types, and high manual examination costs in existing patent classification tools for drug combination identification. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating an embodiment of the intelligent recognition method for drug combination patents based on a large language model in this application.

[0024] Figure 2 This is a schematic diagram of the parameter-efficient fine-tuning architecture based on QLoRA in this application;

[0025] Figure 3 This is a schematic diagram of the overall process of the intelligent recognition method for drug combination patents based on a large language model in this application.

[0026] Figure 4 This is a performance comparison chart for the binary classification task in this application;

[0027] Figure 5 This is a graph showing the performance data of the combined type segmentation identification in this application;

[0028] Figure 6 This is a comparison chart of the accuracy rates for identifying boundary samples and hard cases in this application;

[0029] Figure 7 The results of three rounds of iterative training for mining hard samples are shown in the figure.

[0030] Figure 8 This is a schematic diagram of the functional modules of an embodiment of the intelligent recognition device for drug combination based on a large language model according to this application. Detailed Implementation

[0031] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0032] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0033] In one aspect, embodiments of this application provide a method for intelligent identification of drug combination patents based on a large language model.

[0034] In one embodiment, reference is made to Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the intelligent recognition method for drug combination patents based on a large language model, as described in this application. Figure 1 As shown, the intelligent identification method for drug combination patents based on a large language model includes:

[0035] Step S101: Obtain a standard dataset based on the drug combination labeling information of patent data, and expand the standard dataset to generate an initial training set.

[0036] In one embodiment, establishing a standard dataset based on drug combination annotation information from patent data includes: obtaining corresponding drug combination annotation information by having evaluators perform fine-grained annotation on multiple patent data under a target classification number, wherein the drug combination annotation information includes: drug combination patent binary classification label, combination type label, key sentence fragments, drug name entities, and annotation confidence; performing secondary annotation on samples whose annotation confidence is lower than a preset annotation confidence threshold; establishing the standard dataset based on multiple patent data and corresponding drug combination annotation information, and controlling the quality of the standard dataset to ensure that the consistency coefficient among evaluators is greater than a preset consistency threshold and the data signal-to-noise ratio is greater than a preset signal-to-noise ratio threshold.

[0037] In this embodiment, patent texts are first obtained from a publicly available patent database, with a preference for patent texts from the last 10 years to ensure the timeliness and relevance of the data. During the acquisition of patent texts, special attention is paid to patent texts under the A61K (pharmaceutical compositions) classification, as this classification covers a large amount of patent information related to drug combinations, providing a rich data foundation for subsequent annotation and analysis. Patent data, including fields such as patent number, application date, abstract, and claims, is extracted from the patent texts. These fields contain the core information of the patent and are of great significance for subsequent analysis and judgment.

[0038] In this embodiment, four experts with backgrounds in pharmacy or pharmacology and extensive experience in patent analysis were selected as evaluators to provide accurate and detailed fine-grained annotations for each patent. The specific annotation content is as follows:

[0039] Drug combination patent classification label: used to determine whether a patent is a drug combination patent.

[0040] Combination Type Label: For all samples identified as combination patents, their specific type is labeled. Combination types include fixed-dose combination (FDC), separate-dosage combination, sequential-dosage combination, and combined treatment regimen. This label further refines the characteristics of combination patents, providing a more accurate basis for subsequent classification analysis and application.

[0041] Key Sentence Fragments: Mark key sentence fragments in the patent abstract that support the judgment. These key sentence fragments are the direct basis for determining whether a patent is a combined patent and its specific type. By marking these key sentences, the basis for the judgment can be clearly shown, improving the transparency and traceability of the marking.

[0042] Drug name entities: These include the start and end positions of the drug name in the text, drug entity type labels, and BIO annotation format (B-drug indicates the beginning of the drug name, I-drug indicates the inside of the drug name, and O indicates a non-drug entity). These annotations provide training signals for the NER subtask of drug entity recognition, enabling the model to accurately identify drug entities in the patent abstract.

[0043] Confidence level: Used to indicate the degree of confidence an expert has in their judgment, and is divided into 5 levels. The confidence level label reflects the expert's confidence in the labeled results, providing a reference for subsequent data quality control and results analysis.

[0044] This embodiment employs a combination of double-blind annotation and a consultation mechanism to ensure the accuracy and reliability of the annotation results. Each patent text is independently annotated by two experts. For samples where the annotation results from the two experts are consistent (i.e., both experts annotate the same sample with a confidence level ≥ 3), their annotation results are directly adopted as samples in the standard dataset. For samples where the annotation results from the two experts are inconsistent, or samples where the annotation confidence level is lower than the preset annotation confidence threshold (level 3), the sample is submitted to a third expert for review, or a team discussion is organized to reach a consensus and determine the final annotation result. This process aims to resolve annotation disagreements through collective wisdom and ensure the accuracy of the annotation results.

[0045] The final standard dataset contains 1,000 high-quality labeled data points, with a positive sample (drug combination patent) to negative sample (non-drug combination patent) ratio of approximately 5:5. The balanced sample distribution can effectively reduce the interference of class imbalance on model training.

[0046] Regarding data quality, this embodiment sets dual quantitative control indicators: the first is inter-expert consistency, which is evaluated using the Fleiss' Kappa coefficient, requiring the inter-expert consistency coefficient to be greater than the consistency threshold of 0.85 to ensure that the annotation results have high repeatability and credibility; the second is the signal-to-noise ratio of the annotated data, requiring the signal-to-noise ratio of the final annotated data to be greater than the signal-to-noise ratio threshold of 0.95 to effectively filter noisy labels and ensure the proportion of effective information in the dataset.

[0047] Furthermore, the standard dataset is augmented to generate an initial training set, including: selecting multiple representative samples from the standard dataset and constructing augmented cue words based on the representative samples; using a target large language model to generate synthetic training data based on the augmented cue words; performing quality filtering on the synthetic training data and merging it with the standard dataset to generate the initial training set.

[0048] It is worth noting that in this embodiment, GPT-4o can be selected as the target large language model. By utilizing GPT-4o's few-shot learning capability and based on the principle of Few-shot Learning, the model is shown several examples to demonstrate the task pattern, allowing the model to understand the task to be performed without additional training, thereby generating large-scale synthetic training data.

[0049] For example, in this embodiment, data augmentation prompts are constructed using a pre-designed prompt template. The prompt template includes a task definition (e.g., 200 words), representative sample examples (6 representative samples selected from a standard dataset), and generation instructions (e.g., generating a 200-word patent abstract and corresponding drug combination annotation information).

[0050] The selection criteria for representative samples in this embodiment include: 1. Category coverage: ensuring that there are typical samples for all five types of combination therapy labels (fixed-dose combination, separate-dose combination, sequential-dose combination, combination therapy regimen, and non-combination therapy); 2. Expression diversity: covering explicit expressions (such as "This invention provides a fixed-dose combination of A and B") and implicit expressions (such as "The combined use of A and B shows a synergistic effect"); 3. Difficulty gradient: including simple and clear cases as well as cases with ambiguous boundaries; 4. Quality assurance: only selecting samples with high consistency in expert annotation (confidence level ≥ 4); 5. Moderate text length: selecting patent abstracts of about 200 words, which contain sufficient information without being too long.

[0051] The specific format of the prompt words is "patent abstract + tag": Task Description: Determine whether the following patent abstracts are drug combination patents. If so, please indicate their specific type. Example 1: Patent Abstract: This invention provides a pharmaceutical composition comprising compound A and compound B, wherein A and B exist in a fixed ratio (1:2) in a single formulation for the treatment of type 2 diabetes. Tag: {"Combination": "Yes", "Combination Type": "Fixed-Dose Combination (FDC)", "Confidence Level": 5} Example 2: Patent Abstract: This invention relates to a method of treating cancer, comprising administering antibody C to a patient, followed by administering chemotherapy drug D, wherein C and D are packaged and administered separately. Tag: {"Combination": "Yes", "Combination Type": "Sequential Dosing Combination", "Confidence Level": 5} Example 3: Patent Abstract: This invention provides a novel crystal form of drug E with improved solubility and bioavailability, which can be used to treat hypertension. Tags: {"Whether it is used together": "No", "Type of use": "None", "Confidence level": 5}... (6 examples in total) Now please judge the following patent: [Patent abstract to be judged].

[0052] This embodiment selects 6 representative samples to ensure balanced category coverage. Five main types (four combined types + one non-combined type) require at least 5 examples; we choose 6 to ensure it includes an additional boundary case or complex case. Furthermore, the GPT model needs to understand the task within a limited context window. The 6 examples (approximately 200 words each) plus the task description total approximately 1500-2000 tokens, which is sufficient without being overly resource-intensive within the GPT-4o context window. Experiments have shown that 6 examples achieve a balance between performance and efficiency; too few (e.g., 2-3) will lead to insufficient model understanding, while too many (e.g., 10+) will increase cost with limited performance improvement. This embodiment also allows for a dynamic example selection mechanism. Each time 6 examples are used, the "dynamic example selection mechanism" dynamically selects 6 examples from the standard dataset based on the features of the samples to be generated, ensuring the relevance of the examples.

[0053] In the process of generating synthetic training data by using GPT-4o to expand prompts based on data, techniques such as dynamic example selection mechanism, mutation prompt injection, setting the temperature parameter to 1.0, and bootstrap iterative generation can be adopted to achieve diversity control of synthetic training data.

[0054] The generated synthetic training data undergoes four quality checks: format check (length 100-500 characters), consistency check (confidence difference <0.3), diversity check (similarity <0.95), and expert sampling (8% of data are sampled per batch, accuracy requirement >90%). After passing these checks, the synthetic training data is merged with the standard dataset to generate the initial training set.

[0055] In this embodiment, training data was generated using GPT-4o, expanding the sample size from 1,000 data points in the standard dataset to 12,000 data points, a 12-fold increase. Experimental results show that the synthetic training data achieved an accuracy of 93% in manual sampling.

[0056] Step S102: Using the parameter efficient fine-tuning technique, the base large language model is trained using the initial training set to generate an initial recognition model.

[0057] It is worth noting that, as an example, Figure 2 As shown, this application selects Qwen3-32B-Instruct as the base model, which has 32 billion parameters and possesses powerful bilingual (Chinese and English) understanding and reasoning capabilities, laying a solid linguistic foundation for subsequent fine-tuning. In this embodiment, during the training of the base model, efficient parameter fine-tuning based on QLoRA (Quantized Low-Rank Adaptation) is adopted. By combining model quantization and low-rank adaptation, the memory requirement is reduced by more than 80%, to 48GB, which can be completed on a single A100-80GB GPU.

[0058] Specifically, a parameter-efficient fine-tuning technique is employed to train the base large language model using the initial training set to generate an initial recognition model. This includes: performing 4-bit quantization on the weight parameters of the base large language model and freezing its original weights; decomposing the weight increment changes of the base large language model into two low-rank matrices; setting the hyperparameters of the base large language model; and then training the low-rank matrices using the initial training set.

[0059] For example, the QLoRA in this embodiment combines three key technologies: 4-bit quantization, LoRA low-rank adaptation, and hyperparameter configuration.

[0060] 4-bit quantization processing involves converting the base model weights from the standard 16-bit floating-point (FP16) precision to a 4-bit (NF4) data type. This NF4 quantization scheme is an information-theoretic optimal quantization method that assigns optimal 4-bit representations to non-uniform quantization intervals based on the characteristic that neural network weights typically follow a zero-centered normal distribution. After this quantization, the VRAM occupancy of the model weights is significantly reduced by approximately 75% compared to the FP16 precision.

[0061] Suppose the original pre-trained weight matrix is W (dimension d×k), and the change in weights during fine-tuning is ΔW. The core idea of LoRA low-rank adaptation is that the change in model weights ΔW can be expressed as the product of two low-rank matrices ΔW = B × A. Among them, B has a dimension of r×k and represents the "down projection matrix" or "dimensionality reduction matrix", which reduces the input features from d dimensions to r dimensions (r << d). Its role is to convert the low-dimensional representation back to the original output space to achieve an incremental update of the original weights. A has a dimension of d×r and represents the "up projection matrix" or "dimensionality increase matrix", which maps the r-dimensional intermediate representation back to the k-dimensional output space. Its role is to extract the low-dimensional essential representation of the input features and capture the key information patterns related to the task.

[0062] Among them, r is much smaller than d and k. These three parameters r, d, and k define the matrix dimensions of LoRA. d represents the input dimension (feature dimension) of the original weight matrix W, k represents the output dimension of the original weight matrix W, and r represents the rank of LoRA, which is also the dimension of low-rank decomposition.

[0063] If the number of parameters to be trained for the original full-parameter fine-tuning of the base large language model = d × k, the number of parameters to be trained through LoRA fine-tuning = d × r + r × k = r × (d + k). By setting the size relationship of the three as r << d and r << k, parameter compression is achieved. For example, assume d = k = 4096 and r = 32. Then, for full-parameter fine-tuning, 4096 × 4096 = 16,777,216 parameters need to be trained, and through LoRA fine-tuning, 32 × (4096 + 4096) = 262,144 parameters need to be trained, with a compression ratio of approximately 64 times.

[0064] The hyperparameter configuration of the model is that the LoRA rank r is set to 32, the scaling coefficient alpha is set to 64, the learning rate is set to 2e-4, the batch size is set to the actual batch size of 6, and the effective batch size is made to reach 128 through gradient accumulation. The number of training epochs is 4, and an early stopping strategy is adopted. The optimizer uses AdamW, the weight decay coefficient is set to 0.01, and the Dropout rate is set to 0.1.

[0065] In this embodiment, a Dropout rate of 0.1 means randomly dropping 10% of the neurons (or 10% of the connections or activation values) during training. This embodiment forces the model to learn more robust feature representations by randomly dropping 10% of the neuron connections, thereby preventing overfitting. Dropout is mainly applied to the low-rank matrices A and B of LoRA, randomly setting 10% of the activation values ​​to zero during training. During model training, 10% of the neurons are randomly dropped, and the remaining 90% of the activation values ​​are multiplied by 1 / 0.9 for scaling compensation. During model inference, Dropout is turned off, and the complete network structure is used.

[0066] In this embodiment, the dropout rate is set to 0.1 because oLoRA itself has already achieved regularization through low-rank constraints. An excessively high dropout rate may impair the learning effect of small samples. 0.1 strikes a balance between preventing overfitting and maintaining model capacity.

[0067] With QLoRA technology, the fine-tuning memory requirement of the Qwen3-32B is reduced to 48GB, and fine-tuning can be completed on a single A100-80GB GPU.

[0068] Step S103: Use the initial recognition model to screen out difficult-to-identify samples in the unlabeled patent data and add them to the initial training set to form an enhanced training set.

[0069] It is worth noting that the difficult samples identified in this embodiment include two main types: boundary samples and confusion cases. To improve the model's ability to identify boundary samples and confusion cases, the system introduces active learning and difficult sample mining strategies. After training a preliminary model using the initial training set built in the first two stages, batch predictions are performed in a large-scale unlabeled patent database, focusing on identifying samples with low model confidence (prediction probability in the range of 0.4-0.6).

[0070] Boundary samples refer to samples that fall into the ambiguous zone between combined and non-combined use, or different types of combined use. These mainly include: drug-excipient combinations: such as "a combination of compound A and stabilizer B," requiring determination of whether B is an excipient or an active ingredient; optimized dosing regimens: such as "administering probiotic D before treatment with compound C," requiring determination of whether this is a combined use or a simple dosing regimen; novel combined use regimens: such as "drug E encapsulated in a nanocarrier and then combined with drug F," which have complex structures and are difficult to classify; implicit combined use statements: such as "this invention discovered a synergistic effect between G and H," without explicitly stating whether a combination has been formed.

[0071] Confusion-prone cases refer to samples that are easily misclassified as the wrong category. These mainly include: confusion between combination types: such as "give A first and then give B", which may be misclassified as sequential dosing, but may actually be a FDC dosing guideline; complex multi-drug combinations: such as "A+B+C triple therapy", which requires determination of whether it is a single FDC or a combination therapy of multiple drugs; and samples with ambiguous wording: unclear wording in the patent abstract may simultaneously have multiple combination features.

[0072] The main difference between the two is that boundary samples emphasize that the "location" is on the classification boundary and is inherently difficult to classify clearly, while easily confused cases emphasize that the "risk of misjudgment" has a clear correct category, but the model is prone to misjudging.

[0073] Specifically, the process of using the initial identification model to screen for difficult-to-identify samples in unlabeled patent data and adding them to the initial training set to form an enhanced training set includes: using the initial identification model to identify multiple unlabeled patent data and selecting samples whose model confidence is within a preset confidence interval as difficult-to-identify samples; submitting the difficult-to-identify samples to the evaluator for re-labeling; performing data augmentation on the re-labeled difficult-to-identify samples using oversampling and loss weighting; and adding the data-augmented difficult-to-identify samples to the initial training set to form an enhanced training set.

[0074] In this embodiment, during the difficult sample identification and data augmentation stage, the model undergoes three iterations. In each iteration, 400 difficult samples are identified and submitted to experts for annotation. After annotation, oversampling (repeated 3 times) and loss weighting (weight is 3 times) strategies are adopted, so 400×3=1200 training instances are added in each iteration. After 3 iterations, a total of 3×400=1200 original samples are added, corresponding to 3600 training instances.

[0075] After three rounds of iterations in the difficult sample mining and enhancement phases, the complete training set includes: standard dataset: 1,000 high-quality manually labeled data (always retained), 11,000 synthetic training data generated by GPT-4o (i.e., 12,000 initial training data - 1,000 standard dataset data), and 3,600 training instances corresponding to 1,200 difficult samples, for a total of approximately 15,600 training instances.

[0076] In the hard sample mining stage, samples with a prediction confidence level between 0.4 and 0.6 are identified by the model, while two types of hard samples, namely boundary samples and confusing cases, are captured. Then, the model's ability to identify these samples is enhanced by expert annotation and oversampling strategy (repeated 3 times) and loss weighting strategy (weight 3 times). This hierarchical data construction strategy ensures that the model learns both general joint recognition patterns (through large-scale synthetic data) and can handle difficult boundary cases (through hard samples mined by active learning).

[0077] Step S104: Using the enhanced training set, the initial recognition model is iteratively trained using a multi-task learning framework combined with a contrastive learning mechanism to obtain a drug combination patent recognition model.

[0078] Specifically, this application constructs a multi-task learning architecture that simultaneously optimizes multiple related tasks. These related tasks include: a binary classification task to determine whether a patent is a drug combination patent, which serves as the primary task. Auxiliary tasks include: a combination type multi-classification task, used to further subdivide samples judged as drug combinations into fixed-dose combinations (FDC), separate-dosage combinations, sequential-dosing combinations, or combined treatment regimens, enhancing the model's fine-grained understanding of medication patterns; a key sentence fragment extraction task, used to locate the most relevant sentences in the abstract supporting the combination judgment, improving model interpretability and assisting subsequent manual review; a drug name entity recognition (NER) task, using BIO annotation to extract all drug entities, enhancing the model's sensitivity to active ingredients; and a patent similarity prediction task, using paired abstracts as input to predict the semantic similarity between the two in terms of combination intent.

[0079] The multi-task learning framework consists of 5 sub-tasks, each with a corresponding loss function. The overall loss function of the multi-task learning framework is a weighted linear combination of the loss functions of each associated task: L_total = 0.4×L_classification+0.2×L_type+0.2×L_extraction+0.1×L_ner+0.1×L_similarity.

[0080] Here, L_classification represents the binary classification task loss (the main task), with a weight of 0.4. It represents the cross-entropy loss of the binary classification task, determining whether a patent is a drug combination patent, and is calculated as the standard binary cross-entropy loss. L_type represents the combination type classification loss, with a weight of 0.2. It represents the multi-class classification loss for samples determined to be combination patents, further classifying their specific types (FDC / separate administration / sequential administration / combination treatment regimen), and is calculated as the 4-class cross-entropy loss, the second layer of hierarchical classification. L_extraction represents the key fragment extraction loss, with a weight of 0.2. It represents the loss of training the model to identify key sentences in the patent abstract that support the combination judgment, and is calculated as the sequence labeling task loss, similar to NER, labeling whether each token is a key basis. This task provides interpretability, helping users understand the model's decision basis. L_ner represents the drug entity recognition loss (weight 0.1), which is the loss of the named entity recognition task for identifying drug names in patent abstracts. It is calculated as the conditional random field (CRF) loss for sequence labeling or the token-level cross-entropy loss, and is used to help the model better understand drug-related concepts. L_similarity represents the patent similarity prediction loss (weight 0.1), which is the regression or classification loss for predicting the similarity of two patent abstracts in terms of synergistic intent, and is calculated as the cosine similarity loss or contrastive loss. This task helps the model learn a better patent representation space.

[0081] For difficult samples, the loss for all tasks is multiplied by a weight factor of 3. This weighted combination allows the model to learn relevant auxiliary knowledge while completing the main task, improving overall performance and interpretability.

[0082] In one embodiment, a contrastive learning enhancement strategy is set to improve the accuracy of the model. Specifically, it includes: constructing training sample pairs for contrastive learning based on the enhanced training set, wherein samples with the same combination type constitute positive sample pairs, the original sample and multiple enhanced samples generated by data augmentation constitute positive sample pairs, and samples with different combination types constitute negative sample pairs; calculating supervised contrastive loss based on the constructed positive and negative sample pairs, and applying additional weights to difficult negative sample pairs whose feature similarity with the positive sample pairs is higher than a preset threshold.

[0083] It is worth noting that the core idea of ​​contrastive learning is to learn discriminativeness in the feature space. By bringing the representations of similar samples closer together and pushing the representations of dissimilar samples further apart, samples of the same class are placed in adjacent regions of the feature space, while samples of different classes are placed in non-adjacent regions, thereby learning a feature space with discriminative power.

[0084] After processing by the Qwen3-32B encoder, each patent abstract is mapped to a high-dimensional vector space (feature space). Discriminative power refers to the following properties of this feature space: intra-class compactness: samples of the same category (such as joint patents of the FDC type) are close in the feature space, forming a tight cluster; inter-class separability: samples of different categories (such as FDC vs. sequential drug administration) are far apart in the feature space, making them easy to distinguish; semantic preservation: the distance in the feature space reflects the semantic similarity of the samples—patents that are more similar are closer in the feature space.

[0085] For example, the construction of positive sample pairs in the contrastive learning framework includes three strategies:

[0086] The first strategy is to pair similar samples as positive samples, for example, pairing two patent abstracts that both belong to the FDC type:

[0087] Sample 1: This invention provides a fixed-dose combination tablet comprising aspirin and a statin. Sample 2: This invention relates to a fixed-ratio composition of glimepiride and metformin. Both samples are of the FDC type and therefore constitute a positive sample pair.

[0088] The second strategy involves data augmentation to generate positive sample pairs. Three augmented versions are generated for each training sample, and the original sample is paired with each augmented version to form a positive sample pair. Data augmentation methods include: back-translation: Chinese → English → Chinese, introducing changes in expression while preserving semantics; synonym replacement: replacing "composition" with "preparation," and "containing" with "containing"; sentence rewriting: adjusting sentence structure, such as changing active sentences to passive sentences; and deleting non-critical information: randomly deleting modifiers that have little impact on the judgment.

[0089] For example, the original sample is: "This invention provides a novel fixed-dosage combination comprising compounds A and B for the treatment of diabetes." The corresponding first enhanced sample is: "This invention relates to a fixed-ratio formulation of compounds A and B, which can be used for the treatment of diabetes." The corresponding second enhanced sample is: "A pharmaceutical composition for treating diabetes, comprising fixed doses of A and B." The corresponding third enhanced sample is: "The fixed-dosage combination of this invention contains compounds A and B and is suitable for diabetic patients." The original sample and these three enhanced samples constitute three positive sample pairs.

[0090] The third strategy is actually a negative sample strategy, where samples of different classes are used as negative sample pairs, such as pairing FDC with sequential drug administration combinations.

[0091] For example, a negative sample pair could be: Sample 1 (FDC): This invention provides a fixed-dose combination tablet comprising compounds A and B; Sample 2 (separate administration): This invention relates to a combination therapy of compounds C and D, wherein C and D are packaged and administered separately. These two samples belong to different types of combination therapy and therefore constitute a negative sample pair.

[0092] The scope of negative sample pair construction includes: FDC vs separate dosing combination, FDC vs sequential dosing combination, FDC vs combination therapy, separate dosing vs sequential dosing, and any combination type vs non-combination patent.

[0093] In this embodiment, the difficult negative sample identification method includes: first, feature extraction, using the currently trained Qwen3-32B encoder to map each patent abstract to a feature vector (e.g., 768-dimensional or higher). Next, cosine similarity calculation is performed; for a given anchor sample (positive sample), its cosine similarity with all samples of different categories is calculated. Among all negative sample pairs, if the cosine similarity exceeds a threshold (e.g., 0.7), it is marked as a difficult negative sample. In this embodiment, supervised contrastive loss is used as the contrastive loss function, and the temperature parameter τ is set to 0.07. Difficult negative samples (those with high similarity to positive samples but of different categories) are given additional weights (3 times the weight).

[0094] This embodiment adds extra weights to hard negative samples, forcing the model to learn more fine-grained discriminative features. As training progresses, the model's feature representation continuously improves, and the set of hard negative samples dynamically changes. Some initially hard negative samples may become easily distinguishable, while new hard cases are identified. This dynamic mechanism ensures that the model always focuses on the most challenging samples.

[0095] After training through contrastive learning, the model learns a feature space that allows it to accurately locate a new patent to be judged. Based on its distance from known class samples, the model can accurately determine its class. In particular, for boundary samples, the model can extract more discriminative features for differentiation.

[0096] Experiments show that by incorporating contrastive learning and difficult negative sample mining, the model's recognition accuracy on boundary samples improves by 10 percentage points, and its F1 score improves by 7 percentage points.

[0097] Step S105: Input the patent data to be identified into the drug combination patent identification model to obtain the drug combination information identification result.

[0098] It is worth noting that the input patent data to be identified in this embodiment includes patent abstract text, claim text, etc., and the output of the drug combination patent identification model includes patent binary classification results, combination type, key sentence fragments, drug name entities, and confidence level.

[0099] Optionally, this embodiment also includes confidence level stratification and human-machine collaboration. The prediction results are divided into three levels based on confidence level: if the confidence level is greater than or equal to the first confidence threshold (confidence level ≥ 0.9), it is determined to be in the high confidence level and automatically passes without manual review; if the confidence level is less than the preset first confidence threshold but greater than the preset second confidence threshold (0.7 < confidence level < 0.9), it is determined to be in the medium confidence level, and the drug combination information identification result is marked as pending review; if the confidence level is less than or equal to the second confidence threshold (confidence level ≤ 0.7), it is determined to be in the low confidence level, and the drug combination information identification result is marked as a difficult case, with detailed analysis prompts generated.

[0100] As a preferred implementation, the method also incorporates interpretability and human-machine collaboration mechanisms. An attention weight matrix is ​​extracted from the multi-head attention mechanism of the drug combination patent identification model. A visual heatmap is generated based on this matrix, displaying words and phrases with high attention scores, thus visualizing attention and allowing users to intuitively see the keywords and phrases the model focuses on. Based on the segment extraction task in multi-task learning, key statements supporting classification decisions are automatically labeled. Each key basis is accompanied by a confidence score and basis type label, enabling automatic extraction of key segments with an accuracy of 88%. Using SHAP (SHapley Additive exPlanations), based on game theory's Shapley value, the marginal contribution of each feature to the model output is calculated, achieving feature attribution and providing a theoretically rigorous explanation.

[0101] In practical deployments, key fragment extraction is typically used to provide rapid interpretation for high-confidence predictions, while SHAP is used to provide more in-depth analysis for low-confidence or controversial predictions. This hierarchical interpretation strategy balances computational efficiency with user needs.

[0102] In a specific embodiment, the intelligent recognition method for drug combination patents based on a large language model of this application can be based on, for example: Figure 3The system architecture shown is a modular architecture comprising four core parts: a data processing module, a model training module, a model inference module, and a human-computer interaction module. The data processing module is responsible for the collection, cleaning, annotation, and enhancement of raw patent data. This module includes a manual annotation subsystem, a GPT-4o synthetic data generation subsystem, and a hard sample mining subsystem. The model training module is responsible for training and optimizing the model on the prepared dataset. This module is implemented based on the HuggingFace Transformers and PEFT libraries, supporting techniques such as QLoRA, gradient checkpointing, and mixed-precision training. The model inference module is responsible for classifying and predicting newly input patents. The input is the patent abstract text, and the output includes classification results, confidence scores, key evidence fragments, attention visualizations, and other information. The human-computer interaction module provides a web interface or API interface, supporting users to submit patents, view classification results, review and correct predictions, and provide feedback.

[0103] The specific implementation process of the intelligent identification method for drug combination patents based on large language models includes:

[0104] The data preparation phase includes the following steps:

[0105] Step 201: Patent Data Acquisition: Obtain patent data from publicly available patent databases (such as USPTO, CNIPA, EPO). Focus on patents under the A61K classification (pharmaceutical compositions), prioritizing patents from the last 10 years. Extract fields such as patent number, application date, abstract, and claims.

[0106] Step 202: Standard Dataset Construction: Recruit four domain experts with a Master's degree or above in Pharmacy or Pharmacology and over five years of experience in patent analysis to form an annotation team. Build an annotation platform using professional annotation tools (such as Label Studio). The annotation process is designed as a double-blind annotation + consultation mechanism. Each patent is independently annotated by two experts. For consistent samples (both experts annotate the same sample with a confidence level ≥3), the results are directly adopted. For inconsistent samples, the results are submitted to a third expert or discussed by the team. Ultimately, 1000 high-quality annotated data points are constructed, with a positive to negative sample ratio of approximately 5:5.

[0107] Step 203: GPT-4o Synthetic Training Data Generation: Large-scale synthetic training data generation was performed using OpenAI's GPT-4o API (version 2025-03-26). The prompt word template consists of three parts: task definition (200 words), Few-shot examples (6 samples randomly sampled from the gold standard dataset), and generation instructions (requiring the generation of a 200-word patent abstract). Diversity control was achieved through dynamic example selection, mutation prompt injection, setting the temperature parameter to 1.0, and four rounds of bootstrapping iterative generation. Quality filtering of the synthetic training data was performed using four quality checkpoints (format check, consistency check, diversity check, similarity threshold of 0.95, and expert sampling of 8%). The number of high-quality data points was expanded from 1000 to 12000, achieving a synthetic data accuracy of 93%.

[0108] Step 204, Hard Sample Mining: A preliminary model is trained using a standard dataset and batch-synthesized training data. This model is then used to make batch predictions in a large-scale unlabeled patent database, identifying hard samples with prediction probabilities between 0.4 and 0.6. From the identified hard samples, 400 are selected and submitted to experts for annotation. After annotation, oversampling (repeated 3 times) and loss weighting (3 times the weight) strategies are used to augment the training set. This process is repeated 3 times, with 400 samples mined each time, resulting in a total of 1200 high-quality hard samples. The accuracy of hard sample identification increases from 58% to 82%.

[0109] The model training phase includes the following steps:

[0110] Step 301, Environment Configuration: Set up the training environment. You can use Ubuntu 22.04 and an NVIDIA A100 GPU. Install dependencies such as Python 3.10, PyTorch 2.3.0, Transformers 4.40.0, PEFT 0.10.0, and Bitsandbytes 0.43.0.

[0111] Step 302, Base Model Loading and Quantization: Download the Qwen3-32B-Instruct model from HuggingFace. Configure 4-bit quantization parameters: quantization type nf4, computation data type bfloat16, and enable double quantization. Apply the quantization configuration when loading the model; the quantized model size is approximately 8GB.

[0112] Step 303, LoRA Configuration and Injection: Configure the LoRA hyperparameters including: rank r=32, scaling factor alpha=64, target module as Q, K, V, O projection matrix, and dropout rate 0.1. Inject the LoRA adapter into the base model, which can train approximately 120 million parameters.

[0113] Step 304, Data Preparation and Preprocessing: Organize the training data into a uniform format, and use a Tokenizer to convert the text into a sequence of token IDs, with a maximum sequence length of 512 tokens. Construct a DataLoader with a batch size of 6, and enable gradient accumulation to achieve an effective batch size of 128.

[0114] Step 305, Multi-task Loss Function Design: Implement the multi-task loss function: L_total = 0.4×L_classification+0.2×L_type+0.2×L_extraction+0.1×L_ner+0.1×L_similarity. Multiply the loss of difficult samples across all tasks by a weight factor of 3.

[0115] Step 306, Contrastive Learning Implementation: For each training batch, perform data augmentation on the samples (generate 3 augmented versions for each sample), construct positive and negative sample pairs, calculate supervised contrastive loss, temperature parameter τ=0.07, and weight the hard negative samples by 3 times.

[0116] Step 307, Training Loop and Optimization: Use the AdamW optimizer with a learning rate of 2e-4 and weight decay of 0.01. Learning rate scheduling: warmup for the first 10% of steps, cosine annealing for the remaining 90%. Gradient clipping: maximum gradient norm of 1.0. Mixed-precision training uses BFloat16. Training for 4 epochs, employing an early stopping strategy.

[0117] Step 308, Model Evaluation: Evaluate the model performance on an independent test set (2000 samples). Evaluation metrics include accuracy, precision, recall, F1 score, ROC-AUC, and confusion matrix. Select the model that performs best on both the validation and test sets.

[0118] The model deployment and application phase includes the following steps:

[0119] Step 401: Inference Service Setup: Deploy the trained model as an inference service using frameworks such as FastAPI or vLLM. Service functions include an input interface (receiving patent abstract text), an output interface (returning classification results, confidence scores, and key evidence), and batch processing optimization. The inference time for a single patent is less than 0.5 seconds.

[0120] Step 402, User Interface Development: Develop a web front-end interface to provide a patent submission page, a results display page (showing judgment conclusions, confidence levels, key evidence, and attention visualization), a batch analysis page, and an examination correction page.

[0121] Step 403, Human-Machine Collaborative Workflow: Implement a confidence-based stratified workflow: High confidence (≥0.9): Automatic processing, no manual review required; Medium confidence (0.7-0.9): Quick expert review; Low confidence (≤0.7): In-depth expert analysis.

[0122] like Figures 4 to 7 As shown, the beneficial effects achieved by the embodiments of this application are as follows: (1) The recognition accuracy is significantly improved. The evaluation on the constructed test dataset (containing 2000 patent samples) shows that the accuracy in classification tasks (whether it is a joint patent) has increased from 80% of the traditional BERT method to 94% of the present invention; the F1 score has increased from 78% of the traditional method to 92%; the recall rate has increased from 74% to 91%; and the precision rate has increased from 82% to 94%. In the multi-class classification task of joint types: the overall accuracy is 90%; the macro average F1 score is 88%; the recognition accuracy of each category is FDC 93%, the separate dosing combination is 88%, the sequential dosing combination is 86%, and the combination treatment regimen is 85%. In the boundary sample recognition: the drug-excipient combination has increased from 58% to 82%, an increase of 24 percentage points; the implicit combination expression has increased from 62% to 85%, an increase of 23 percentage points; the novel combination regimen has increased from 55% to 78%, an increase of 23 percentage points; and the complex multi-drug combination has increased from 64% to 83%, an increase of 19 percentage points.

[0123] (2) Data efficiency is greatly improved through GPT-4o synthetic data generation and hard sample mining strategies: manual annotation cost is reduced by 90% to achieve the same model performance. Traditional methods require 10,000 data annotations, while this invention only requires 1,000 high-quality annotations; the training data scale is expanded 12 times, from 1,000 manual annotations to 12,000 high-quality training data; the quality of synthetic data is close to that of manual annotation, with an accuracy of 93%; the efficiency of new domain transfer learning is high, requiring only 80 annotation data for domain adaptation to achieve an accuracy of 87%.

[0124] (3) Reduced computing resource requirements: QLoRA technology significantly lowers the hardware threshold for large-scale model fine-tuning: memory requirements are reduced by 82%, from 260GB to 48GB; hardware costs are reduced, from requiring an 8×A100 cluster to a single A100 with 80GB of memory for fine-tuning.

[0125] (4) Enhanced model interpretability: The multi-level interpretability mechanism of this invention significantly improves user trust: attention visualization allows users to intuitively see the keywords and phrases that the model focuses on; key evidence is automatically extracted with an accuracy of 88%; SHAP feature attribution provides a theoretically rigorous explanation.

[0126] Secondly, embodiments of this application also provide a patent intelligent recognition device for drug combination based on a large language model.

[0127] In one embodiment, reference is made to Figure 8 , Figure 8 This is a functional module diagram of an embodiment of the intelligent recognition device for drug combination based on a large language model, as described in this application. Figure 8 As shown, the intelligent recognition device for drug combination patents based on a large language model includes:

[0128] The generation module is used to obtain a standard dataset based on drug combination labeling information from patent data, and to expand the standard dataset to generate an initial training set.

[0129] The fine-tuning module is used to train the base large language model using the initial training set by employing parameter-efficient fine-tuning techniques to generate an initial recognition model.

[0130] The filtering module is used to filter out difficult-to-identify samples from unlabeled patent data using the initial identification model and add them to the initial training set to form an enhanced training set.

[0131] The training module is used to iteratively train the initial recognition model using the enhanced training set, employing a multi-task learning framework combined with a contrastive learning mechanism, to obtain a drug combination patent recognition model.

[0132] The identification module is used to input the patent data to be identified into the drug combination patent identification model to obtain the drug combination information identification result.

[0133] Furthermore, in one embodiment, the generation module is further configured to:

[0134] By having the evaluator perform fine-grained annotation on multiple patent data under the target classification number, the corresponding drug combination annotation information is obtained. The drug combination annotation information includes: drug combination patent binary classification label, combination type label, key sentence fragments, drug name entity, and annotation confidence level.

[0135] For samples whose label confidence level is lower than a preset label confidence level threshold, secondary labeling is performed;

[0136] The standard dataset is established based on multiple patent data and corresponding drug combination labeling information, and the quality of the standard dataset is controlled to ensure that the inter-evaluator consistency coefficient is greater than a preset consistency threshold and the data signal-to-noise ratio is greater than a preset signal-to-noise ratio threshold.

[0137] Furthermore, in one embodiment, the generation module is further configured to:

[0138] Multiple representative samples are selected from the standard dataset, and data augmentation prompts are constructed based on the representative samples;

[0139] Using the target large language model, prompt words are expanded based on the data to generate synthetic training data;

[0140] The synthetic training data is quality filtered and merged with the standard dataset to generate the initial training set.

[0141] Furthermore, in one embodiment, the fine-tuning module is also used for:

[0142] The weight parameters of the base large language model are quantized to 4 bits, and their original weights are frozen.

[0143] The weight increment change of the base large language model is decomposed into two low-rank matrices. After setting the hyperparameters of the base large language model, the low-rank matrices are trained using the initial training set.

[0144] Furthermore, in one embodiment, the filtering module is also used for:

[0145] The initial identification model was used to identify multiple unlabeled patent data, and samples with model confidence scores within a preset confidence interval were selected as difficult-to-identify samples.

[0146] The difficult-to-identify samples were submitted to the evaluators for further labeling.

[0147] For the re-labeled, difficult-to-identify samples, oversampling and loss-weighted processing are used for data augmentation;

[0148] The data-enhanced samples that are difficult to identify are added to the initial training set to form an enhanced training set.

[0149] Furthermore, in one embodiment, the training module is also used for:

[0150] A multi-task learning architecture is constructed to simultaneously optimize multiple associated tasks, including: a binary classification task to determine whether a patent is a drug combination patent, a multi-classification task of combination type, a key sentence fragment extraction task, a drug name entity recognition task, and a patent similarity prediction task.

[0151] The overall loss function of the multi-task learning framework is a weighted linear combination of the loss functions of each associated task.

[0152] Furthermore, in one embodiment, the training module is also used for:

[0153] Training sample pairs for contrastive learning are constructed based on the augmented training set, wherein samples of the same type are paired to form positive sample pairs, the original sample and multiple augmented samples generated by data augmentation are paired to form positive sample pairs, and samples of different types are paired to form negative sample pairs.

[0154] The supervised contrast loss is calculated based on the constructed positive and negative sample pairs, and additional weights are applied to difficult negative sample pairs whose feature similarity to the positive sample pairs is higher than a preset threshold.

[0155] Furthermore, in one embodiment, the device further includes a coordination module, which is used for:

[0156] An attention weight matrix is ​​extracted from the multi-head attention mechanism of the drug combination patent identification model, and a visual heatmap is generated based on the attention weight matrix.

[0157] Furthermore, in one embodiment, the collaboration module is also used for:

[0158] If the confidence level of the result is greater than or equal to the preset first confidence level threshold, the drug combination information identification result will automatically pass the review.

[0159] If the confidence level of the result is less than the preset first confidence level threshold but greater than the preset second confidence level threshold, the drug combination information identification result will be marked as pending review.

[0160] If the confidence level of the result is less than or equal to the second confidence level threshold, the drug combination information identification result is marked as a difficult case and a detailed analysis prompt is generated;

[0161] The first confidence threshold is greater than the second confidence threshold.

[0162] The functions of each module in the above-mentioned intelligent identification device for drug combination patents based on a large language model correspond to the steps in the above-mentioned intelligent identification method for drug combination patents based on a large language model. Their functions and implementation processes will not be described in detail here.

[0163] It should be noted that the sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0164] The terms "comprising" and "having," and any variations thereof, in the specification, claims, and accompanying drawings of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such process, method, product, or apparatus. The terms "first," "second," and "third," etc., are used to distinguish different objects, etc., and do not indicate a sequence, nor do they limit "first," "second," and "third" to different types.

[0165] In the description of the embodiments of this application, terms such as "exemplary," "for example," or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as "exemplary," "for example," or "for instance" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary," "for example," or "for instance" is intended to present the relevant concepts in a concrete manner.

[0166] In the description of the embodiments of this application, unless otherwise stated, " / " means "or". For example, A / B can mean A or B. The "and / or" in the text is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist simultaneously, and B exists alone. In addition, in the description of the embodiments of this application, "multiple" means two or more.

[0167] In some processes described in the embodiments of this application, multiple operations or steps are included in a specific order. However, it should be understood that these operations or steps may not be executed in the order they appear in the embodiments of this application, or they may be executed in parallel. The sequence number of the operation is only used to distinguish different operations, and the sequence number itself does not represent any execution order. In addition, these processes may include more or fewer operations, and these operations or steps may be executed sequentially or in parallel, and these operations or steps may be combined.

[0168] 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) as described above, and includes several instructions to cause a terminal device to execute the methods described in the various embodiments of this application.

[0169] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for identifying drug combination patents based on a large language model, characterized in that, The intelligent identification method for drug combination patents based on a large language model includes: A standard dataset is established based on the drug combination labeling information of patent data, and the standard dataset is augmented to generate an initial training set; Using a parameter-efficient fine-tuning technique, the initial training set is used to train the base large language model to generate an initial recognition model; The initial recognition model is used to screen out difficult-to-identify samples from unlabeled patent data and add them to the initial training set to form an enhanced training set; Using the enhanced training set, the initial identification model is iteratively trained using a multi-task learning framework combined with a contrastive learning mechanism to obtain a drug combination patent identification model. Input the patent data to be identified into the drug combination patent identification model to obtain the drug combination information identification result; The drug combination labeling information includes: drug combination patent binary classification label, combination type label, key sentence fragments, drug name entities, and labeling confidence level; The method further includes iteratively training the initial recognition model using the enhanced training set and a multi-task learning framework combined with a contrastive learning mechanism: Training sample pairs for contrastive learning are constructed based on the augmented training set, wherein samples of the same type are paired to form positive sample pairs, the original sample and multiple augmented samples generated by data augmentation are paired to form positive sample pairs, and samples of different types are paired to form negative sample pairs. The supervised contrast loss is calculated based on the constructed positive and negative sample pairs, and additional weights are applied to difficult negative sample pairs whose feature similarity to the positive sample pairs is higher than a preset threshold.

2. The intelligent recognition method for drug combination patents based on a large language model as described in claim 1, characterized in that, A standard dataset was established based on drug combination labeling information from patent data, including: By having the evaluator perform fine-grained annotation on multiple patent data under the target classification number, the corresponding drug combination annotation information can be obtained; For samples whose label confidence level is lower than a preset label confidence level threshold, secondary labeling is performed; The standard dataset is established based on multiple patent data and corresponding drug combination labeling information, and the quality of the standard dataset is controlled to ensure that the inter-evaluator consistency coefficient is greater than a preset consistency threshold and the data signal-to-noise ratio is greater than a preset signal-to-noise ratio threshold.

3. The intelligent identification method for drug combination patents based on a large language model as described in claim 1, characterized in that, The standard dataset is augmented to generate an initial training set, including: Multiple representative samples are selected from the standard dataset, and data augmentation prompts are constructed based on the representative samples; Using the target large language model, prompt words are expanded based on the data to generate synthetic training data; The synthetic training data is quality filtered and merged with the standard dataset to generate the initial training set.

4. The intelligent recognition method for drug combination patents based on a large language model as described in claim 1, characterized in that, Employing efficient parameter fine-tuning techniques, the initial training set is used to train the base large language model, generating an initial recognition model, including: The weight parameters of the base large language model are quantized to 4 bits, and their original weights are frozen. The weight increment change of the base large language model is decomposed into two low-rank matrices. After setting the hyperparameters of the base large language model, the low-rank matrices are trained using the initial training set.

5. The intelligent identification method for drug combination patents based on a large language model as described in claim 2, characterized in that, The initial recognition model is used to screen for difficult-to-identify samples in unlabeled patent data and added to the initial training set to form an enhanced training set, including: The initial identification model was used to identify multiple unlabeled patent data, and samples with model confidence scores within a preset confidence interval were selected as difficult-to-identify samples. The difficult-to-identify samples were submitted to the evaluators for further labeling. For the re-labeled, difficult-to-identify samples, oversampling and loss-weighted processing are used for data augmentation; The data-enhanced samples that are difficult to identify are added to the initial training set to form an enhanced training set.

6. The intelligent recognition method for drug combination patents based on a large language model as described in claim 2, characterized in that, Using the enhanced training set, the initial recognition model is iteratively trained using a multi-task learning framework combined with a contrastive learning mechanism, including: A multi-task learning architecture is constructed to simultaneously optimize multiple associated tasks, including: a binary classification task to determine whether a patent is a drug combination patent, a multi-classification task of combination type, a key sentence fragment extraction task, a drug name entity recognition task, and a patent similarity prediction task. The overall loss function of the multi-task learning framework is a weighted linear combination of the loss functions of each associated task.

7. The intelligent recognition method for drug combination patents based on a large language model as described in claim 1, characterized in that, The method also includes: An attention weight matrix is ​​extracted from the multi-head attention mechanism of the drug combination patent identification model, and a visual heatmap is generated based on the attention weight matrix.

8. The intelligent recognition method for drug combination patents based on a large language model as described in claim 1, characterized in that, After inputting the patent data to be identified into the drug combination patent identification model and obtaining the drug combination information identification result, the process also includes: If the confidence level of the result is greater than or equal to the preset first confidence level threshold, the drug combination information identification result will automatically pass the review. If the confidence level of the result is less than the preset first confidence level threshold but greater than the preset second confidence level threshold, the drug combination information identification result will be marked as pending review. If the confidence level of the result is less than or equal to the second confidence level threshold, the drug combination information identification result is marked as a difficult case and a detailed analysis prompt is generated; The first confidence threshold is greater than the second confidence threshold.

9. A patent intelligent recognition device for drug combination based on a large language model, characterized in that, The intelligent recognition device for drug combination patents based on a large language model includes: The generation module is used to obtain a standard dataset based on drug combination labeling information from patent data, and to expand the standard dataset to generate an initial training set. The fine-tuning module is used to train the base large language model using the initial training set by employing parameter-efficient fine-tuning techniques to generate an initial recognition model. The filtering module is used to filter out difficult-to-identify samples from unlabeled patent data using the initial identification model and add them to the initial training set to form an enhanced training set. The training module is used to iteratively train the initial recognition model using the enhanced training set, employing a multi-task learning framework combined with a contrastive learning mechanism, to obtain a drug combination patent recognition model. The identification module is used to input the patent data to be identified into the drug combination patent identification model to obtain the drug combination information identification result. The drug combination labeling information includes: drug combination patent binary classification label, combination type label, key sentence fragments, drug name entities, and labeling confidence level; The training module is further used for: Training sample pairs for contrastive learning are constructed based on the augmented training set, wherein samples of the same type are paired to form positive sample pairs, the original sample and multiple augmented samples generated by data augmentation are paired to form positive sample pairs, and samples of different types are paired to form negative sample pairs. The supervised contrast loss is calculated based on the constructed positive and negative sample pairs, and additional weights are applied to difficult negative sample pairs whose feature similarity to the positive sample pairs is higher than a preset threshold.

Images