A reference type analysis method and system based on joint multi-task learning

Abstract

The application provides a reference type analysis method and system based on joint multi-task learning, and relates to the technical field of semantic classification. The method comprises the following steps: configuring input data and label data of a reference intention classification task; defining a reference evaluation classification task, configuring input data and label data of the reference evaluation classification task; constructing a joint multi-task learning framework, embedding the reference intention classification task and the reference evaluation classification task as parallel classification tasks into the joint multi-task learning framework; adopting an alternating iteration strategy of the parallel classification tasks, training the joint multi-task learning framework, balancing gradient propagation between the parallel tasks, obtaining a reference type joint analysis model, and performing reference type analysis processing. Through the application, the technical problem that the deep semantic correlation between the reference intention and the reference evaluation cannot be effectively mined in the prior art, thereby affecting the intelligent analysis capability of scientific research content, can be solved, and the technical effect of improving the interpretation capability of the citation semantics can be achieved.

Description

Technical Field

[0001] This application relates to the field of semantic classification technology, and in particular to a method and system for reference type analysis based on joint multi-task learning. Background Technology

[0002] As science and technology evaluation systems continue to evolve from quantity-oriented to quality-oriented approaches, extracting more interpretive information from citation behavior has become an important research direction in the fields of natural language processing and scientific literature mining. Citations not only carry the connections between research results but also reflect the citer's understanding and evaluation of the cited literature. Therefore, the semantic classification of cited content is gradually becoming a key means to assist in scientific research evaluation, academic influence analysis, and knowledge graph construction.

[0003] Currently, citation semantic analysis mainly focuses on two directions: citation intent classification, which aims to identify the purpose of a citer in using a document, such as whether it is used to introduce background knowledge, borrow research methods, or compare results; and citation evaluation or sentiment classification, which focuses on the citer's attitude in the citation, such as affirmation, skepticism, or neutral description. However, in existing technologies, these two tasks are usually modeled separately, with each building an independent model for training and prediction, lacking a unified semantic coordination mechanism. Especially at the data level, the scarcity of citation evaluation datasets and the high cost of annotation limit model training and result in insufficient generalization ability. Furthermore, existing methods often employ static training strategies, failing to fully consider the data heterogeneity and gradient conflicts between citation tasks, which can easily lead to model bias towards a single task, thereby weakening overall performance and adaptability.

[0004] In summary, existing technologies suffer from several technical problems. Due to the lack of a unified joint modeling mechanism for citation semantics, they fail to effectively uncover the deep semantic relationship between citation intent and citation evaluation, and neglect dynamic balancing and feature sharing strategies among tasks. As a result, the models struggle to accurately capture the complex expressive structures in the cited text, further impacting the depth of understanding of research content, the intelligent parsing ability of citation behavior, and the accuracy and practicality of high-level scientific and technological evaluation. Summary of the Invention

[0005] The purpose of this application is to provide a citation type analysis method and system based on joint multi-task learning, in order to solve the technical problems in the prior art that, due to the lack of a unified joint modeling mechanism for citation semantics, the deep semantic relationship between citation intent and citation evaluation is not effectively explored, and the dynamic balance and feature sharing strategies between tasks are ignored, making it difficult for the model to accurately capture the complex expression structure in the cited text, which further affects the depth of understanding of scientific research content, the intelligent parsing ability of citation behavior, and the accuracy and practicality of high-level scientific and technological evaluation.

[0006] In view of the above problems, this application provides a reference type analysis method and system based on joint multi-task learning.

[0007] Firstly, this application provides a reference type analysis method based on joint multi-task learning, implemented through a reference type analysis system based on joint multi-task learning, comprising: defining a reference intent classification task and configuring the input data and label data for the reference intent classification task; defining a reference evaluation classification task and configuring the input data and label data for the reference evaluation classification task; constructing a joint multi-task learning framework, embedding the reference intent classification task and the reference evaluation classification task as parallel classification tasks into the joint multi-task learning framework, wherein the reference intent classification task and the reference evaluation classification task have independent input spaces and label spaces for inputting reference context texts from different sources; and training the joint multi-task learning framework using an alternating iterative strategy of parallel classification tasks, balancing gradient propagation between parallel tasks, obtaining a joint reference type analysis model, and performing reference type analysis processing.

[0008] Preferably, the reference type analysis method based on joint multi-task learning further includes: constructing reference intent classification and reference evaluation classification training datasets, including input data and label data; inputting the input data of reference intent classification and reference evaluation classification into a shared feature extraction layer, and generating cross-task general semantic features through a pre-trained language model; inputting the shared cross-task general semantic features into task-specific classification heads, and outputting reference intent category and reference evaluation category in parallel; and iteratively training the reference intent classification and reference evaluation classification training datasets alternately, calculating loss values ​​independently and updating model parameters until the convergence target is reached, thereby obtaining the joint reference type analysis model.

[0009] Preferably, the citation type analysis method based on joint multi-task learning further includes: configuring citation intent classification label categories, including background, method, and result comparison; configuring citation evaluation classification label categories, including important evaluation and general evaluation; collecting a citation sample set, using citation context text as input data, and labeling each citation context text in the citation sample set according to the citation intent classification label category and the citation evaluation classification label category, thereby constructing a citation intent classification dataset and a citation evaluation classification training dataset.

[0010] Preferably, the citation type analysis method based on joint multi-task learning further includes: using a pre-trained language model as a shared feature extraction layer to encode the input citation context text into a text sequence to obtain an input text sequence; based on the input text sequence, using the pre-trained language model to generate corresponding hidden state representations, and performing average pooling on the hidden states; introducing the LeakyReLU activation function to enhance the features of the average pooled hidden states, generating cross-task general semantic features.

[0011] Preferably, the reference type analysis method based on joint multi-task learning further includes: using the formula: Nonlinear representation feature enhancement is performed to generate the cross-task general semantic features, where W shared and b shared These are the weights and biases of the shared fully connected layer, To mitigate neuron death, the LeakyReLU activation function introduces a negative slope to the hidden state after average pooling.

[0012] Preferably, the reference type analysis method based on joint multi-task learning further includes: the negative slope is 0.01.

[0013] Preferably, the reference type analysis method based on joint multi-task learning further includes: setting an independent classification head for each task; using the classification heads of each task to map the shared cross-task general semantic features to the category labels of the reference intent classification and the reference evaluation classification, respectively, to obtain the reference intent category and the reference evaluation category.

[0014] Preferably, the reference type analysis method based on joint multi-task learning further includes: defining a cross-entropy loss function for each task, calculating the loss value of each classification task respectively, and summing the loss values ​​of parallel multi-tasks as the total loss; updating the model parameters based on the loss values, and iteratively performing alternating training until the training objective is reached, thereby obtaining the joint reference type analysis model.

[0015] Preferably, the reference type analysis method based on joint multi-task learning further includes: during alternating iterative training, when the dataset of one task is exhausted in advance, resetting the iterator for the number of tasks is performed; and continuing sampling of other tasks until the parallel classification tasks are exhausted synchronously.

[0016] Secondly, this application also provides a reference type analysis system based on joint multi-task learning, used to execute a reference type analysis method based on joint multi-task learning as described in the first aspect, comprising: a reference intent classification task definition module, used to define a reference intent classification task and configure the input data and label data of the reference intent classification task; a reference evaluation classification task definition module, used to define a reference evaluation classification task and configure the input data and label data of the reference evaluation classification task; a learning framework embedding module, used to construct a joint multi-task learning framework, embedding the reference intent classification task and the reference evaluation classification task as parallel classification tasks into the joint multi-task learning framework, wherein the reference intent classification task and the reference evaluation classification task have independent input spaces and label spaces, used to input reference context texts from different sources; and a learning framework training module, used to train the joint multi-task learning framework using an alternating iterative strategy of parallel classification tasks, balancing gradient propagation between parallel tasks, obtaining a joint reference type analysis model, and performing reference type analysis processing.

[0017] The technical solution provided in this application has at least the following technical effects or advantages: by achieving the technical goal of constructing a joint analysis model of citation types for multidimensional semantic understanding and improving the collaborative recognition ability of intent and evaluation features in cited text, it achieves the technical effects of enhancing the explanatory power and stability of the model for citation semantics and supporting the intelligent application of scientific research results impact assessment and science and technology management.

[0018] The above description is merely an overview of the technical solution of this application. To better understand the technical means of this application and to facilitate its implementation according to the description, and to make the above and other objects, features, and advantages of this application more apparent, specific embodiments of this application are described below. It should be understood that the content described in this section is not intended to identify key or important features of the embodiments of this application, nor is it intended to limit the scope of this application. Other features of this application will become readily apparent through the following description. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating a reference type analysis method based on joint multi-task learning proposed in this application.

[0021] Figure 2This is a schematic diagram of the structure of a reference type analysis system based on joint multi-task learning according to this application.

[0022] Figure labeling: Reference intent classification task definition module 11, reference evaluation classification task definition module 12, learning framework embedding module 13, learning framework training module 14. Detailed Implementation

[0023] This application provides a citation type analysis method and system based on joint multi-task learning. It addresses the technical problems in existing technologies where the lack of a unified joint modeling mechanism for citation semantics prevents effective mining of the deep semantic connections between citation intent and citation evaluation, and neglects dynamic balancing and feature sharing strategies among tasks. This results in models struggling to accurately capture the complex expressive structures in cited texts, further impacting the depth of understanding of research content, the intelligent parsing ability of citation behavior, and the accuracy and practicality of high-level scientific and technological evaluation. The application aims to construct a joint citation type analysis model for multi-dimensional semantic understanding and enhance the collaborative recognition ability of intent and evaluation features in cited texts. Ultimately, this enhances the model's explanatory power and stability of citation semantics, supporting the assessment of the impact of scientific research achievements and the intelligent application of science and technology management.

[0024] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. It should be understood that this application is not limited to the exemplary embodiments described herein. 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. It should also be noted that, for ease of description, only the parts related to this application are shown in the accompanying drawings, not all of them.

[0025] Example 1, please refer to the appendix. Figure 1 This application provides a reference type analysis method based on joint multi-task learning, applied to a reference type analysis system based on joint multi-task learning, specifically including the following steps:

[0026] Define a referencing intent classification task, and configure the input data and label data for the referencing intent classification task.

[0027] Specifically, the task of classifying citation intent is defined, which involves determining the purpose of citation in the context of cited literature in a text, such as introducing background knowledge, referencing methods, or comparing results, and then identifying the functional use of the cited literature by the citer.

[0028] Define a citation evaluation classification task, and configure the input data and label data for the citation evaluation classification task.

[0029] Specifically, the task of citation evaluation is defined to determine the evaluation attitude towards the cited literature, that is, whether the semantic tendency expressed in the cited text is positive recognition, neutral statement or implicit criticism, and thus to obtain the subjective evaluation position shown by the citer when citing the research results of others, rather than the specific purpose of the citation. Therefore, it is complementary to but independent of the citation intention classification.

[0030] A joint multi-task learning framework is constructed, in which the citation intent classification task and the citation evaluation classification task are embedded as parallel classification tasks. The citation intent classification task and the citation evaluation classification task have independent input spaces and label spaces, which are used to input citation context texts from different sources.

[0031] Specifically, a joint multi-task learning framework is constructed, building a deep learning model architecture capable of handling multiple tasks simultaneously. This framework facilitates inter-task communication by sharing parts of the model structure and performs independent prediction optimization in task-specific components. Multi-task learning can improve the model's generalization ability across various tasks, especially performing better in scenarios where tasks have some semantic correlation but limited training data.

[0032] The citation intent classification task and the citation evaluation classification task are embedded as parallel classification tasks into the joint multi-task learning framework. As two independent learning units that share some feature extraction capabilities, the parallel approach means that the model learns two types of tasks simultaneously within a training cycle, which not only improves training efficiency but also uncovers the potential semantic connections between the two types of tasks.

[0033] The citation intent classification task and the citation evaluation classification task have independent input spaces. The training input text data used by each task is independent and not mixed with each other. The citation intent task focuses on the functional use of the cited text, while the citation evaluation task focuses on the implicit sentiment in the cited statements. Therefore, the input corpora for each task can come from different sources, forming two separate input sets. For example, one task can use citation corpora from papers in the computer science field, while the other task uses citation samples from the biomedical field.

[0034] The citation intent classification task and the citation evaluation classification task have independent label spaces, meaning that each task has a different set of output labels. Labels for citation intent classification might include background, method, and result comparison, while labels for citation evaluation classification might include important evaluations and general evaluations. This separation of label spaces allows each task to design the most suitable classification method based on its own semantic features, without introducing interference from shared labels. The citation intent classification task and the citation evaluation classification task use different data sources to train the model's input representation capabilities, which helps to capture the semantic patterns of citation expressions under different tasks. For example, intent-type citations have a more explicit structure and intuitive semantics, while evaluation-type citations rely more on subtle wording variations; therefore, processing them separately can improve the model's ability to recognize textual details.

[0035] The joint multi-task learning framework is trained by adopting an alternating iterative strategy for parallel classification tasks, balancing gradient propagation between parallel tasks, and obtaining a joint reference type analysis model for reference type analysis.

[0036] Specifically, the training process considers two different but related classification tasks simultaneously: citation intent classification and citation evaluation classification. The training is conducted by alternating between the two tasks, selecting a batch of samples from one task in each round before moving to the other, thus achieving a balanced training pace between tasks. A joint multi-task learning framework is trained, which uses a unified neural network structure with shared semantic feature extraction layers and independent classification modules, optimizing the performance of both tasks simultaneously. Balancing gradient propagation between parallel tasks means calculating the loss of each task separately and controlling the impact of their gradients generated in the feature-sharing layer on the model update with each model parameter update. This prevents the error of one task from dominating the entire network's learning direction, thereby improving training stability and overall accuracy. For example, if the amount of data for the citation intent task is much larger than that for the citation evaluation task, without balancing, gradient propagation might favor the former, causing the latter to perform poorly. The resulting joint citation type analysis model, formed after multiple iterations, is a model capable of simultaneously determining citation intent and citation evaluation, with citation context as input and two types of labels as output. Performing citation type analysis means that the citation type joint analysis model can be practically applied to academic literature processing, such as identifying whether a text is borrowing methods or criticizing results, thereby supporting scenarios such as scientific research impact evaluation or citation motivation tracking.

[0037] Furthermore, this application also includes: constructing separate training datasets for citation intent classification and citation evaluation classification, including input data and label data; inputting the input data for citation intent classification and citation evaluation classification into a shared feature extraction layer, and generating cross-task general semantic features through a pre-trained language model; inputting the shared cross-task general semantic features into task-specific classification heads, and outputting citation intent category and citation evaluation category in parallel; and iteratively training the citation intent classification and citation evaluation classification training datasets alternately, calculating loss values ​​independently and updating model parameters, until the convergence target is reached, thereby obtaining the joint analysis model for obtaining citation types.

[0038] Specifically, separate training sample sets are prepared for two different text classification tasks: citation intent and citation evaluation. The input data consists of contextual text containing citation markers, while the label data comprises manually or automatically labeled category information. Labels for citation intent classification are generally categorized into types such as background, method, and result comparison, while labels for citation evaluation classification are categorized into categories such as important evaluation and general evaluation. By constructing two separate datasets, the model can understand cited text from different semantic perspectives.

[0039] The pre-trained language model can be BERT or SciBERT, capable of extracting context-dependent semantic information from the input text. The input data for citation intent classification and citation evaluation classification are fed separately into a shared feature extraction layer. The pre-trained language model generates cross-task general semantic features, allowing both tasks to benefit from a shared understanding of language structure, thereby improving learning efficiency and feature representation capabilities under limited data conditions.

[0040] Shared, cross-task general semantic features are input into task-specific classification heads, which output citation intent category and citation evaluation category in parallel. In other words, the extracted semantic features are simultaneously fed into two independent classification modules. Each module makes predictions for either the citation intent or citation evaluation task, mapping abstract features to specific category labels. This allows the two tasks to share knowledge while maintaining independence in output, ensuring that each task accurately achieves its respective objective. For example, the same sentence is predicted as "method citation + importance evaluation," rather than confusing the judgments of the two dimensions.

[0041] By alternately sampling training datasets for citation intent classification and citation evaluation classification, independently calculating loss values ​​and updating model parameters, and iteratively training until convergence is achieved, a joint citation type analysis model is obtained. That is, during training, the model does not process all task samples simultaneously, but rather extracts data from the intent and evaluation tasks alternately in a certain order, calculates the classification error for each, and updates the model parameters based on backpropagation of the error. The training process continues iteratively until the model converges on the validation set, meaning its classification performance stabilizes and no longer shows significant improvement. The final model is the joint analysis model, capable of simultaneously recognizing citation intent and citation evaluation.

[0042] Furthermore, this application also includes: configuring citation intent classification label categories, including background, method, and result comparison; configuring citation evaluation classification label categories, including important evaluation and general evaluation; collecting a citation sample set, using citation context text as input data, and labeling each citation context text in the citation sample set according to the citation intent classification label category and the citation evaluation classification label category, thereby constructing a citation intent classification dataset and a citation evaluation classification training dataset.

[0043] Specifically, in the citation intent classification task, the range of categories that the task needs to identify is pre-defined, which refers to the types of labels that the joint multi-task learning framework will output during training and prediction. The citation intent classification label categories include background, method, and result comparison, which respectively indicate that the cited text is used to introduce the research background, draw on existing methods, or compare with the results of others. This helps the joint multi-task learning framework understand the functional purpose of the cited text, thereby enabling more accurate semantic classification.

[0044] In the citation evaluation classification task, the evaluation type labels to be predicted by the joint multi-task learning framework are defined. Citation evaluation classification label categories include important evaluation and general evaluation. Important evaluation indicates that the citer gives high recognition or explicit praise to the cited literature, while general evaluation indicates that the citer only makes a neutral citation or has no obvious emotional bias. Configuring citation evaluation classification label categories can guide the joint multi-task learning framework to capture the subjective evaluation characteristics in citation language, which is helpful for fine-grained analysis of the academic value in scientific literature.

[0045] Text fragments containing citations are extracted from historical texts as samples. The citation sample set is located before and after the citation annotations, for example, containing sentences such as "the method proposed by Smith et al. (2015)...". The citation context text is used as input data, indicating that the citation-related statements are used as training input for a joint multi-task learning framework to train the understanding of semantic features in the cited text.

[0046] Each reference context text in the reference sample set is labeled according to the reference intent classification label and the reference evaluation classification label. Then, each sample is labeled to ensure that each input text has both a reference intent label and a reference evaluation label. The labeled reference samples are divided into two datasets according to the task requirements: one is used to train the model to recognize the semantic function of references, i.e., the intent classification task, and the other is used to train the model to recognize the evaluation stance of references, i.e., the sentiment classification task. The reference intent classification dataset and the reference evaluation classification training dataset are constructed to obtain a multi-task training format, thereby realizing the joint learning of reference semantics by the joint multi-task learning framework.

[0047] Furthermore, this application also includes: using a pre-trained language model as a shared feature extraction layer to encode the input reference context text into a text sequence to obtain an input text sequence; based on the input text sequence, using the pre-trained language model to generate a corresponding hidden state representation, and performing average pooling on the hidden state; introducing the LeakyReLU activation function to enhance the features of the average pooled hidden state, generating cross-task general semantic features.

[0048] Specifically, a pre-trained language model is used as a shared feature extraction layer. This means that within the joint multi-task learning framework, a natural language understanding model that has been deeply trained on a large-scale corpus, such as BERT or SciBERT, is selected to extract semantic features from the input text. This shared feature extraction layer is universal across multiple tasks, effectively extracting the underlying semantic information contained in the input text, avoiding independent modeling for each task, thus reducing the number of parameters and improving learning efficiency. The input quoting context text is then encoded into a text sequence. This involves converting a quoted text, typically a complete sentence containing quoting markers, into a numerical vector sequence that the joint pre-trained language model can process through steps such as word vectorization, positional encoding, and masking. This ultimately generates a continuous semantic representation of the original sentence, resulting in the input text sequence.

[0049] The encoded text sequence is input into a pre-trained language model. After processing by the multi-layer neural network of the pre-trained language model, the hidden semantic representation of each word or position is output. The hidden states at all positions are averaged over the vector dimension to obtain the global semantic representation of the entire text. The entire citation context is compressed into a unified vector in the semantic space, allowing subsequent tasks to process the input holistically without relying on word-by-word information. This improves the stability and abstract representation ability of the joint multi-task learning framework.

[0050] After average pooling of the hidden states, an activation function is introduced to enhance the nonlinear representation capability of the joint multi-task learning framework. LeakyReLU retains a small output value even when the input is negative, which can effectively alleviate the problem that some neurons in deep networks always output zero during training, thereby improving the robustness of model training.

[0051] Vectors processed by the LeakyReLU activation function possess stronger non-linear expressive power, making it easier for joint multi-task learning frameworks to capture complex semantic changes in input text. The processed representation vectors can be used as shared input for multiple tasks, containing both structural information at the intent level and subjective information at the evaluation level. Therefore, they can support classification heads for different tasks in discriminating specific targets, serving as general semantic features across tasks.

[0052] Furthermore, this application also includes: via the formula: Nonlinear representation feature enhancement is performed to generate the cross-task general semantic features, where W shared and b shared These are the weights and biases of the shared fully connected layer, To mitigate neuron death, the LeakyReLU activation function introduces a negative slope to the hidden state after average pooling.

[0053] Specifically, the text hidden state vector after average pooling is input into a calculation formula with a non-linear activation function to further transform the text, making it more semantically expressive. The weight W... shared and bias b shared These are the parameters of the shared fully connected layer, used for linear mapping of the input vector. Weights W shared Represents a matrix that determines how each dimension of the input vector is reconstructed or scaled; bias b shared It is a vector that has been comprehensively translated based on the weight mapping result. A shared fully connected layer means that the parameter settings are consistent across all tasks, ensuring that the generated semantic representation has cross-task interpretability and consistency. Hidden state after average pooling. Average pooling refers to averaging the hidden states generated by a pre-trained language model after encoding text, resulting in a global vector representation. This operation eliminates the influence of text length differences, ensuring a uniform semantic representation regardless of input text length, thus facilitating subsequent fully connected layers and classifier processing. The LeakyReLU activation function introduces a negative slope to retain more gradient information. The introduction of a negative slope effectively alleviates the problem of some neurons becoming completely inactive and unable to continue updating during neural network training—a phenomenon known as neuron death. By mitigating neuron death, the joint multi-task learning framework can more stably complete parameter learning, improving the ability to model complex semantic relationships.

[0054] Furthermore, this application also includes: the negative slope is 0.01.

[0055] Specifically, in the LeakyReLU activation function, for the less-than-zero numerical part of the input, the corresponding output is set to the product of the input value and a fixed ratio, which is the negative slope. LeakyReLU does not directly set the output to zero when dealing with negative inputs; instead, it retains a small portion of the gradient, preventing neurons in the neural network from completely failing during training, thereby improving the overall learning ability. Setting the negative slope to 0.01 means that when the input value is negative, the output value is the input value multiplied by 0.01, thus preserving gradient propagation capability while controlling the influence of negative values.

[0056] Furthermore, this application also includes: setting an independent classification head for each task; using the classification heads of each task to map the shared cross-task general semantic features to the category labels of the reference intent classification and the reference evaluation classification, respectively, to obtain the reference intent category and the reference evaluation category.

[0057] Specifically, a separate classification head is set up for each task, with dedicated output modules designed for the citation intent classification task and the citation evaluation classification task, respectively. The classification head consists of a set of linear layers or a feedforward neural network, whose structure maps the input semantic feature vector to a specific classification probability distribution. Since the two tasks have different objectives—citation intent focuses on functional purpose judgment, while citation evaluation focuses on subjective stance expression—using separate classification heads avoids confusing the semantic boundaries of different tasks, allowing each task to focus on its own feature decoding.

[0058] The shared cross-task general semantic features are mapped to category labels for both the citation intent and sentiment evaluation dimensions using the classification heads of each task. In other words, the global semantic representation of the text obtained through the shared feature extraction layer is fed into the independent classification heads of the two tasks to predict the citation text's intent category and sentiment evaluation category, respectively. The intent category label includes background, method, and result comparison, explaining the purpose of the citation behavior; the evaluation category label includes important and general evaluations, reflecting the positive or negative attitude towards the citation. The mapping process essentially transforms the continuous vector representation of the input into discrete category probabilities, thus supporting specific classification judgments. Through the mapping and classification process, the softmax function is used to output the probability of each category label. The category with the highest probability is selected as the final prediction result, i.e., the classification label of the current citation context in the intent dimension and the classification label in the sentiment evaluation dimension, serving as the citation intent category and citation evaluation category, respectively.

[0059] Furthermore, this application also includes: defining a cross-entropy loss function for each task, calculating the loss value for each classification task respectively, and summing the loss values ​​of parallel multi-tasks as the total loss; updating the model parameters based on the loss values, and iteratively performing alternating training until the training objective is reached, thereby obtaining the reference type joint analysis model.

[0060] Specifically, independent loss functions are set for the citation intent classification task and the citation evaluation classification task to measure the deviation between the model's predicted results and the true labels. Cross-entropy loss is the objective function used for classification problems, guiding model learning by calculating the difference between the model's predicted probability distribution and the actual label distribution. The closer the predicted result is to the true label, the smaller the cross-entropy value, and vice versa, thus serving as a direct driving indicator for model optimization. The loss values ​​for the citation intent classification task and the citation evaluation classification task are calculated separately. That is, the predicted results for each task are output separately, and the current error values ​​for each task are calculated based on their respective cross-entropy loss functions. This ensures that each task can independently evaluate its training effect and does not lose independent supervision signals due to sharing feature extraction layers with other tasks.

[0061] The total loss is obtained by summing the loss values ​​from multiple parallel tasks, while the loss values ​​calculated separately for each task are directly added together to form a unified optimization objective. This summation method enables joint optimization among multiple tasks, allowing the model to consider the performance of both types of tasks simultaneously during training. This guides the network structure in the shared parts to learn general features more comprehensively, avoiding severe resource competition and imbalance between tasks, and helping to maintain the learning balance of different tasks.

[0062] Model parameters are updated based on the loss value, using optimization algorithms such as gradient descent. The learnable parameters in the model are backpropagated and updated according to the current total loss value. The direction and magnitude of each parameter update depend on the gradient of the loss function with respect to each parameter, thus gradually improving the model's performance on the training data. Parameter updating is the core step in the deep learning model training process, determining whether the model can eventually converge to a state with good generalization ability. In multiple training cycles, training samples are alternately selected between two tasks according to a pre-defined alternating sampling strategy, thus avoiding a situation where one task dominates the training process for a long time. Alternating training not only prevents the gradient direction of a single task from dominating the learning direction of the entire shared network, but also balances the optimization priority issues caused by differences in task data distribution. For example, when citation evaluation task samples are scarce while intent task samples are abundant, direct mixed training may lead to overfitting of intent features; alternating training can balance the influence of both. After multiple rounds of alternating iterative training, the model's various indicators reach the preset standards. At this point, training can be considered complete, and this model can be used as the final result for actual automatic citation intent and citation evaluation identification tasks to obtain a joint analysis model of citation types. The joint analysis model for citation types has the ability to simultaneously identify the functional purpose and emotional attitude of citations, providing multi-dimensional semantic tags for academic text analysis and serving application scenarios such as scientific research evaluation, trend analysis, and knowledge graph construction.

[0063] Furthermore, this application also includes: during alternating iterative training, when the dataset for one task is exhausted prematurely, the iterator for the number of tasks is reset; sampling of other tasks continues until the parallel classification tasks are exhausted synchronously.

[0064] Specifically, during multi-task model training, when all data samples from either the intent classification task or the evaluation classification task are used up first—meaning all samples in the current training round have been traversed—the data reader for that task needs to be reset to return to the beginning of the dataset so it can continue participating in training in subsequent alternating iterations. An iterator is a control mechanism for sequentially accessing data, used to read training samples from the dataset one by one. The reset operation ensures that even if a task has less data in each training round, it will not be ignored by the model, thus improving the balance and generalization effect of the training.

[0065] Once the dataset for one task has been traversed and reset, sampling and training continue on the dataset for another task that has not yet been fully traversed, until both datasets have been fully trained in one round. Synchronous exhaustion means that both datasets reach a state where there are no remaining samples for training in a certain round of training. One round of alternating training is completed, thus ensuring that even if the number of data samples for the two tasks is different, the model training will not favor the task with the larger sample size, thereby improving the fairness and robustness of the overall learning process.

[0066] In summary, the citation type analysis method based on joint multi-task learning provided in this application has the following technical effects: by achieving the technical goal of constructing a joint citation type analysis model for multi-dimensional semantic understanding and improving the collaborative recognition ability of intent and evaluation features in cited text, it achieves the technical effects of enhancing the explanatory power and stability of the model for citation semantics and supporting the intelligent application of scientific research results impact assessment and science and technology management.

[0067] Example 2: Based on the same inventive concept as the reference type analysis method based on joint multi-task learning in the foregoing examples, this application also provides a reference type analysis system based on joint multi-task learning. Please refer to the appendix. Figure 2 The system includes: a citation intent classification task definition module 11, used to define a citation intent classification task and configure the input data and label data of the citation intent classification task; a citation evaluation classification task definition module 12, used to define a citation evaluation classification task and configure the input data and label data of the citation evaluation classification task; a learning framework embedding module 13, used to construct a joint multi-task learning framework, embedding the citation intent classification task and the citation evaluation classification task as parallel classification tasks into the joint multi-task learning framework, wherein the citation intent classification task and the citation evaluation classification task have independent input spaces and label spaces, used to input citation context texts from different sources; and a learning framework training module 14, used to train the joint multi-task learning framework using an alternating iterative strategy of parallel classification tasks, balancing gradient propagation between parallel tasks, obtaining a joint citation type analysis model, and performing citation type analysis processing.

[0068] Furthermore, the reference type analysis system based on joint multi-task learning is also used for: constructing reference intent classification and reference evaluation classification training datasets, including input data and label data; inputting the input data of reference intent classification and reference evaluation classification into a shared feature extraction layer, generating cross-task general semantic features through a pre-trained language model; inputting the shared cross-task general semantic features into task-specific classification heads, and outputting reference intent category and reference evaluation category in parallel; and iteratively training by alternately sampling the reference intent classification and reference evaluation classification training datasets, independently calculating loss values ​​and updating model parameters until convergence, thereby obtaining the joint reference type analysis model.

[0069] Furthermore, the reference type analysis system based on joint multi-task learning is also used for: configuring reference intent classification label categories, including background, method, and result comparison; configuring reference evaluation classification label categories, including important evaluation and general evaluation; collecting a reference sample set, taking the reference context text as input data, and labeling each reference context text in the reference sample set according to the reference intent classification label category and the reference evaluation classification label category, thereby constructing a reference intent classification dataset and a reference evaluation classification training dataset.

[0070] Furthermore, the aforementioned citation type analysis system based on joint multi-task learning is also used to: utilize a pre-trained language model as a shared feature extraction layer to encode the input citation context text into a text sequence to obtain an input text sequence; based on the input text sequence, utilize the pre-trained language model to generate corresponding hidden state representations and perform average pooling on the hidden states; introduce the LeakyReLU activation function to enhance the features of the average pooled hidden states and generate cross-task general semantic features.

[0071] Furthermore, the aforementioned reference type analysis system based on joint multi-task learning is also used to: [using the formula:] Nonlinear representation feature enhancement is performed to generate the cross-task general semantic features, where W shared and b shared These are the weights and biases of the shared fully connected layer, To mitigate neuron death, the LeakyReLU activation function introduces a negative slope to the hidden state after average pooling.

[0072] Furthermore, the reference type analysis system based on joint multi-task learning is also used to: have a negative slope of 0.01.

[0073] Furthermore, the reference type analysis system based on joint multi-task learning is also used to: set an independent classification head for each task; and use the classification heads of each task to map the shared cross-task general semantic features to the category labels of the reference intent classification and the reference evaluation classification, respectively, to obtain the reference intent category and the reference evaluation category.

[0074] Furthermore, the reference type analysis system based on joint multi-task learning is also used to: define a cross-entropy loss function for each task, calculate the loss value of each classification task respectively, and sum the loss values ​​of parallel multi-tasks as the total loss; update the model parameters based on the loss value, and iteratively perform alternating training until the training objective is reached, thereby obtaining the reference type joint analysis model.

[0075] Furthermore, the reference type analysis system based on joint multi-task learning is also used in the following way: during alternating iterative training, when the dataset of one task is exhausted in advance, the task quantity iterator is reset; sampling of other tasks continues until the parallel classification tasks are exhausted synchronously.

[0076] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The reference type analysis method and specific example based on joint multi-task learning in the foregoing embodiment one are also applicable to the reference type analysis system based on joint multi-task learning in this embodiment. Through the foregoing detailed description of the reference type analysis method based on joint multi-task learning, those skilled in the art can clearly understand the reference type analysis system based on joint multi-task learning in this embodiment. Therefore, for the sake of brevity, it will not be described in detail here.

[0077] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0078] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of this application and its equivalents, this application also intends to include such modifications and variations.

Claims

1. A reference type analysis method based on joint multi-task learning, characterized in that, include: Define a referencing intent classification task, and configure the input data and label data for the referencing intent classification task; Define a citation evaluation classification task, and configure the input data and label data for the citation evaluation classification task; A joint multi-task learning framework is constructed, in which the citation intent classification task and the citation evaluation classification task are embedded as parallel classification tasks into the joint multi-task learning framework. The citation intent classification task and the citation evaluation classification task have independent input spaces and label spaces, which are used to input citation context texts from different sources. The joint multi-task learning framework is trained by adopting an alternating iterative strategy for parallel classification tasks, balancing gradient propagation between parallel tasks, obtaining a joint reference type analysis model, and performing reference type analysis processing. The method employs an alternating iterative strategy for parallel classification tasks to train the joint multi-task learning framework, balancing gradient propagation between parallel tasks to obtain a joint reference type analysis model, including: Construct training datasets for citation intent classification and citation evaluation classification, including input data and label data; The input data for the reference intent classification and reference evaluation classification are respectively input into the shared feature extraction layer, and cross-task general semantic features are generated through a pre-trained language model; The shared cross-task general semantic features are input into the task-specific classification head, and the reference intent category and reference evaluation category are output in parallel. By alternately sampling the reference intent classification and reference evaluation classification training datasets, independently calculating the loss value and updating the model parameters, and iteratively training alternately until the convergence target is reached, the joint analysis model for obtaining reference types is obtained. Specifically, training datasets for citation intent classification and citation evaluation classification are constructed, including: Configure the reference intent category tags, including background, method, and result comparison; Configure the category tags for the evaluation, including important evaluations and general evaluations; Collect a citation sample set, take the citation context text as input data, and label each citation context text in the citation sample set according to the citation intent classification label category and the citation evaluation classification label category, and construct a citation intent classification dataset and a citation evaluation classification training dataset; The input data for the reference intent classification and reference evaluation classification are respectively input into the shared feature extraction layer, and cross-task general semantic features are generated through a pre-trained language model, including: By using a pre-trained language model as a shared feature extraction layer, the input reference context text is encoded into a text sequence to obtain the input text sequence. Based on the input text sequence, the corresponding hidden state representation is generated using the pre-trained language model, and the hidden state is then averaged and pooled. We introduce the LeakyReLU activation function to enhance the features of the average pooling hidden state, generating general semantic features across tasks.

2. The reference type analysis method based on joint multi-task learning according to claim 1, characterized in that, Through the formula: Nonlinear representation feature enhancement is performed to generate the cross-task general semantic features, wherein... and These are the weights and biases of the shared fully connected layer, To mitigate neuron death, the LeakyReLU activation function introduces a negative slope to the hidden state after average pooling.

3. The reference type analysis method based on joint multi-task learning according to claim 2, characterized in that, The negative slope is 0.

01.

4. The reference type analysis method based on joint multi-task learning according to claim 1, characterized in that, The shared cross-task general semantic features are input into the task-specific classification head, and the reference intent category and reference evaluation category are output in parallel, including: Set an independent category header for each task; By using the classification heads of each task, the shared cross-task general semantic features are mapped to the category labels of the reference intent category and the reference evaluation category, respectively, to obtain the reference intent category and the reference evaluation category.

5. The reference type analysis method based on joint multi-task learning according to claim 1, characterized in that, By alternately sampling the reference intent classification and reference evaluation classification training datasets, independently calculating the loss value and updating the model parameters, and iteratively training alternately until the convergence target is reached, the joint analysis model for obtaining reference types is obtained, including: Define a cross-entropy loss function for each task, calculate the loss value for each classification task separately, and sum the loss values ​​of multiple parallel tasks as the total loss; The model parameters are updated based on the loss value, and iterative training is performed until the training objective is achieved, thereby obtaining the joint analysis model of the reference type.

6. The reference type analysis method based on joint multi-task learning according to claim 5, characterized in that, Iterative training is performed alternately until the training objective is achieved, including: During alternating iterative training, when the dataset for one task is consumed prematurely, the task quantity iterator is reset. Continue sampling for other tasks until the parallel classification tasks are exhausted.

7. A reference type analysis system based on joint multi-task learning, characterized in that, The steps for implementing the reference type analysis method based on joint multi-task learning as described in any one of claims 1 to 6 include: The reference intent classification task definition module is used to define reference intent classification tasks and configure the input data and label data of the reference intent classification tasks. The citation evaluation classification task definition module is used to define the citation evaluation classification task and configure the input data and label data of the citation evaluation classification task. The learning framework embedding module is used to construct a joint multi-task learning framework. The citation intent classification task and the citation evaluation classification task are embedded as parallel classification tasks into the joint multi-task learning framework. The citation intent classification task and the citation evaluation classification task have independent input spaces and label spaces, which are used to input citation context texts from different sources. The learning framework training module is used to train the joint multi-task learning framework by adopting an alternating iterative strategy of parallel classification tasks, balancing gradient propagation between parallel tasks, obtaining a joint reference type analysis model, and performing reference type analysis processing.

Images