A method and system for identifying opinion sentences in scientific literature based on sentiment enhancement

Abstract

The present invention provides a method and system for identifying opinion sentences in scientific and technological documents based on sentiment enhancement, which relates to the field of natural language processing technology and includes: obtaining a target document for which opinion sentences are to be identified; constructing an opinion sentence identification model, wherein the opinion sentence identification model includes a context feature extraction component, a local text feature extraction component, and a weighted sentiment feature fusion enhancement component; inputting the target document into the opinion sentence identification model, and outputting the opinion sentence identification result. The present invention solves the technical problem that the opinion sentence identification models of the prior art mainly focus on opinion sentence mining in the fields of social media, product reviews, etc., and rely on explicit sentiment words for analysis. It is difficult to directly migrate opinion sentences in scientific and technological documents that generally lack explicit sentiment words, resulting in insufficient adaptability to scientific and technological documents.

Description

Technical Field

[0001] The present invention relates to the technical field of natural language processing, and in particular to a method and system for identifying opinion sentences in scientific and technological literature based on emotion enhancement. Background Art

[0002] Identifying opinion sentences in scientific literature is a natural language processing (NLP) task aimed at automatically identifying sentences in text that contain the author's subjective evaluations, reasoning, or judgments. These sentences often reflect the author's subjective views, speculations, or evaluations, rather than simply objective statements of fact. Opinion sentences are often based on the author's personal understanding, research findings, or theoretical hypotheses, expressing the author's attitude, insights, or conclusions on a specific research question.

[0003] In scientific and technical literature, opinion sentences are not only subjective but may also contain complex reasoning, academic commentary, or theoretical discussions. They are typically found in core sections such as the abstract, introduction, background, discussion, and conclusion. They involve evaluating the strengths and weaknesses of research methods, interpreting the significance of experimental results, comparing and analyzing existing research, and predicting future research directions. Accurately identifying opinion sentences is of great value to academic literature, not only revealing the author's academic stance and research attitude, but also providing critical support for information extraction, literature summarization, and knowledge discovery.

[0004] Existing opinion sentence recognition has the following limitations:

[0005] 1. Research Limitations: Most current opinion sentence mining techniques focus on analyzing subjective information in areas such as product reviews and social media. These methods typically focus on sentences that explicitly express sentiment, such as positive or negative evaluations. However, opinions expressed in scientific literature are often more complex, involving implicit reasoning, specialized terminology, and academic argumentation. These characteristics differ significantly from those expressed in popular texts, making existing techniques difficult to directly apply to the task of identifying opinion sentences in scientific literature.

[0006] 2. Lack of dedicated research on opinion sentences in scientific literature: Opinion sentences in scientific literature often contain complex reasoning logic and technical language, implicitly presenting the author's subjective views. This complexity makes it difficult for existing opinion sentence recognition technologies to accurately capture the author's academic attitudes and implicit judgments when processing scientific literature. Furthermore, research on opinion mining methods that closely relate to the research content and academic context of scientific texts is limited. Summary of the Invention

[0007] This application aims to address the shortcomings of existing technologies in this field by providing a method and system for identifying opinion sentences in scientific and technological literature based on sentiment enhancement. Current opinion sentence recognition models are mainly focused on areas such as social media and product reviews, relying on explicit sentiment words for analysis. This makes it difficult to effectively transfer these models to scientific and technological literature involving implicit reasoning, professional terminology, and academic argumentation, resulting in limitations in adaptability. To address this issue, this application provides a more adaptable recognition solution to improve recognition performance in scientific and technological literature.

[0008] The first aspect disclosed in the present application provides a method for identifying opinion sentences in scientific and technological documents based on sentiment enhancement, the method comprising: obtaining a target document for which opinion sentences are to be identified; constructing an opinion sentence identification model, wherein the opinion sentence identification model comprises a context feature extraction component, a local text feature extraction component, and a weighted sentiment feature fusion enhancement component; inputting the target document into the opinion sentence identification model, and outputting an opinion sentence recognition result.

[0009] The second aspect disclosed in the present application provides a system for identifying opinion sentences in scientific and technological literature based on emotion enhancement. The system is used for the above-mentioned method for identifying opinion sentences in scientific and technological literature based on emotion enhancement. The system includes: a target document acquisition module, used to acquire the target document to be subjected to opinion sentence recognition; a recognition model construction module, used to construct an opinion sentence recognition model, wherein the opinion sentence recognition model includes a context feature extraction component, a local text feature extraction component and a weighted emotion feature fusion enhancement component; a recognition result acquisition module, used to input the target document into the opinion sentence recognition model, and output the obtained opinion sentence recognition result.

[0010] One or more technical solutions provided in this application have at least the following beneficial effects:

[0011] By extracting target documents to be analyzed from scientific and technological literature, efficient management of input data is achieved, ensuring that the data input to the model is targeted. A model for identifying opinion sentences is constructed, which includes a contextual feature extraction component, a local text feature extraction component, and a weighted sentiment feature fusion enhancement component. It can realize multi-level text feature extraction and processing. This design helps to comprehensively capture the semantic information, contextual associations, and sentiment features in the target documents, providing multi-dimensional feature support for opinion sentence identification. The contextual feature extraction component can accurately capture the implicit information in the sentence, improving the model's ability to capture domain terms and context, and enhancing the accuracy of opinion sentence identification. The local text feature extraction component can accurately capture local semantic features of different scales in the sentence, which is particularly suitable for complex language expressions in opinion sentences. It can effectively extract deep semantic information across multiple words in opinion sentences, improving the ability to understand complex opinion sentences. The weighted sentiment feature fusion enhancement component can accurately identify the implicit sentiment features in opinion sentences, improving the model's ability to recognize opinion sentences. After the target document is input into the model, the contextual feature extraction, local text feature extraction, and weighted sentiment feature fusion enhancement components work together to improve the recognition effect of opinion sentences.

[0012] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 A flow chart of a method for identifying opinion sentences in scientific literature based on sentiment enhancement is provided in an embodiment of the present application.

[0014] Figure 2 An exemplary structural diagram of an opinion sentence recognition model in a method for identifying opinion sentences in scientific literature based on emotion enhancement provided in an embodiment of the present application.

[0015] Figure 3 A schematic diagram of the structure of a scientific literature opinion sentence recognition system based on emotion enhancement provided in an embodiment of the present application.

[0016] Description of reference numerals: target document acquisition module 10, recognition model construction module 20, recognition result acquisition module 30. DETAILED DESCRIPTION

[0017] The embodiments of the present application provide a method and system for identifying opinion sentences in scientific literature based on sentiment enhancement. This addresses the limitation of existing technologies, which are mainly applied in the fields of social media and product reviews and rely on explicit sentiment words for analysis. It solves the problem of insufficient adaptability in scientific literature involving implicit reasoning, professional terminology and academic argumentation, where opinion sentences generally lack explicit sentiment words and are difficult to directly migrate.

[0018] After introducing the basic principles of this application, various non-limiting embodiments of this application will be specifically described below in conjunction with the accompanying drawings. It should be understood that the specific embodiments described here are only used to explain this application and are not used to limit this application.

[0019] Example 1, as Figure 1 As shown, the embodiment of the present application provides a method for identifying opinion sentences in scientific and technological literature based on sentiment enhancement, the method comprising:

[0020] Step S100: Obtaining a target document for opinion sentence recognition;

[0021] Obtain target documents. Target documents refer to scientific and technological documents containing the opinion and factual sentences to be recognized. These documents can include academic papers, research reports, and technical patents. Agricultural scientific and technological documents are preferred. Furthermore, preprocess the target documents, including segmentation, sentence segmentation, and removal of irrelevant information such as headers, footers, and figure captions, to improve subsequent recognition accuracy.

[0022] Step S200: constructing an opinion sentence recognition model, wherein the opinion sentence recognition model includes a context feature extraction component, a local text feature extraction component, and a weighted sentiment feature fusion enhancement component;

[0023] Construct an opinion sentence recognition model for opinion sentence recognition in scientific literature, including context feature extraction components, local text feature extraction components, and weighted sentiment feature fusion enhancement components, such as Figure 2The figure below shows an exemplary structure of the opinion sentence recognition model. The context feature extraction component uses AgriBERT (Agricultural Bidirectional Encoder Representations from Transformers) as an encoder to extract contextual features from the input sentence. AgriBERT can effectively capture domain-specific language features and expressions in agricultural science and technology literature, providing high-quality semantic representations for opinion sentence recognition. The local text feature extraction component uses a Deep Pyramid Convolutional Neural Network (DPCNN) for hierarchical feature extraction, leveraging its hierarchical convolutional structure to capture multi-scale opinion expression features in sentences. The weighted sentiment feature fusion enhancement component uses a weighted sentiment feature fusion enhancement mechanism to organically combine sentiment features with text features to enhance the model's ability to capture implicit emotions and subjective expressions in opinion sentences.

[0024] Step S300: input the target document into the opinion sentence recognition model, and output the opinion sentence recognition result.

[0025] After the preprocessed target document text is divided into sentences, it is input into the opinion sentence recognition model sentence by sentence. The context feature extraction component generates the context representation vector of the sentence, the local text feature extraction component extracts hierarchical multi-scale sentence features, the weighted sentiment feature fusion enhancement component combines sentiment features and text features, and the fully connected layer generates the classification probability of each sentence through the Softmax activation function. Based on the classification probability, the category of each sentence is determined. Finally, the model outputs the classification label of each sentence and the corresponding probability distribution to obtain the opinion sentence recognition result.

[0026] Furthermore, constructing an opinion sentence recognition model includes:

[0027] Step S210: constructing a corpus for identifying opinion sentences;

[0028] Step S220: constructing a context feature extraction component, a local text feature extraction component, and a weighted sentiment feature fusion enhancement component within the opinion sentence recognition model;

[0029] Step S230: Based on the opinion sentence recognition corpus, the opinion sentence recognition model is trained on the training set, the model parameters with the best performance are selected on the validation set, and the performance of the final opinion sentence recognition model is evaluated on the test set.

[0030] Obtain a sample document dataset from scientific and technological literature, ensure that the sample documents contain opinion sentences with a variety of writing styles and different ways of expression, such as directly expressed opinions and implicitly inferred opinions, annotate the opinion sentences and factual sentences in each sample document data, obtain a sample opinion sentence recognition result set, divide the sample document dataset and the sample opinion sentence recognition result set into a training set, a validation set, and a test set, and thus construct an opinion sentence recognition corpus.

[0031] Using AgriBERT as the encoder, we built a contextual feature extraction component to extract contextual features from the input sentence, generating domain-specific representation vectors and enhancing the model's understanding of agricultural terminology and context. Leveraging its pre-trained contextual awareness, AgriBERT accurately captures implicit information in sentences, thereby improving the accuracy of opinion sentence recognition.

[0032] The Deep Pyramid Convolutional Neural Network (DPCNN) is used to construct a local text feature extraction component. This hierarchical convolutional structure captures semantic features at different scales within a sentence, making it particularly suitable for the complex language expressions found in opinion sentences. The DPCNN effectively extracts deep semantic information across multiple words in opinion sentences, enhancing the model's ability to understand complex opinion sentences.

[0033] By using a weighted sentiment feature fusion enhancement mechanism, we constructed a weighted sentiment feature fusion enhancement component, organically combining sentiment features with text features to enhance the model's ability to capture implicit emotions and subjective expressions in opinion sentences. This component, through the weighted sentiment feature fusion mechanism, captures implicit sentiment features in opinion sentences and improves the model's ability to distinguish opinion sentences from factual sentences.

[0034] Use the training set in the opinion sentence recognition corpus to train the model, calculate the loss value and update the model parameters based on the prediction results and the true label; use the validation set for mid-term evaluation and adjust the model's hyperparameters, such as the learning rate and hidden layer dimension; evaluate the model performance on the test set and calculate the final indicators, such as the F1 value. When the F1 value meets the preset requirements, the opinion sentence recognition model is completed.

[0035] Furthermore, we construct a corpus for identifying opinion sentences, including:

[0036] Step S211: Obtain a sample document dataset;

[0037] Step S212: manually annotating the sentences in each sample document dataset to distinguish opinion sentences from factual sentences, so as to construct an opinion sentence recognition corpus;

[0038] Step S213: According to experimental requirements, the constructed opinion sentence recognition corpus is divided into a training set, a validation set, and a test set to support the training and evaluation of the opinion sentence recognition model.

[0039] Obtain a sample document dataset from scientific and technological literature, including academic papers, research reports, and technical patent documents. The dataset must have a distribution of opinion sentences and factual sentences. That is, it must include opinion sentences with a variety of writing styles and different ways of expression, such as direct opinions and implicit inferences. Preferably, scientific and technological literature in the agricultural field should be selected, covering different topics such as plant protection, soil improvement, and crop yield.

[0040] We manually annotated the sentences in each sample document dataset to distinguish opinion sentences from factual sentences. Opinion sentences are sentences that express the author's subjective opinion, reasoning, hypothesis, evaluation, or conclusion, such as "The study showed that this method significantly increased yield." Factual sentences are sentences that state objective facts or describe experimental methods, such as "The experiment included three variable groups and one control group."

[0041] Specialized annotation tools are used for manual annotation at the sentence level. For example, doctoral students and researchers with backgrounds in relevant agricultural disciplines are invited to participate in the annotation to ensure professionalism. Each sentence is labeled as "opinion sentence" (such as labeled as 1) or "fact sentence" (such as labeled as 0) to ensure the consistency and reliability of the annotation results. After the annotation is completed, a corpus of opinion sentence recognition is obtained.

[0042] Based on experimental requirements, the opinion sentence recognition corpus was partitioned into training, validation, and test sets in a 6:2:2 ratio. The training set, with the largest amount of data, was used for model training; the validation set was used for model parameter tuning; and the test set was used for final model performance evaluation. A random sampling method was used to ensure an even distribution of opinion and factual sentences across each dataset. The resulting partitioning served as the opinion sentence recognition corpus.

[0043] Furthermore, the context feature extraction component, the local text feature extraction component, and the weighted sentiment feature fusion enhancement component within the opinion sentence recognition model are constructed, including:

[0044] Step S221: Construct the context feature extraction component in the opinion sentence recognition model, use AgriBERT as the Embedding layer to vectorize the input text, and for the input sequence S = {s1,...,s i ,...,s L}, and its output sequence is represented as X={x1,...,x i ,...,x L}, where L is the length of the input sequence, s iis the i-th token in the input sequence, is the corresponding representation vector, and d is the hidden layer dimension;

[0045] Step S222: constructing a local text feature extraction component within the opinion sentence recognition model;

[0046] Step S223: constructing a weighted sentiment feature fusion enhancement component within the opinion sentence recognition model;

[0047] Step S224: Connect the context feature extraction component, the local text feature extraction component and the weighted sentiment feature fusion enhancement component to obtain an opinion sentence recognition model.

[0048] The embedding layer is used to encode the input text sequence into a representation vector with contextual information. In specific domain tasks, domain-specific pre-trained language models can significantly improve the performance of tasks such as text classification and information extraction by pre-training on large-scale domain corpora. AgriBERT is a pre-trained language model fine-tuned on a diverse agricultural corpus that can capture scientific knowledge and practical application information in the agricultural field and generate a unique embedding representation with domain context relevance. The input sequence of the constructed context feature extraction component is S = {s1,...,s i ,...,s L}, under the encoding of AgriBERT, the generated output sequence is represented as X = {x1,...,x i ,...,x L}, where L is the length of the input sequence, s i is the i-th token in the input sequence, x i ∈R d is the corresponding representation vector, and d is the hidden layer dimension.

[0049] The local text feature extraction component is built based on a deep pyramid convolutional neural network, a highly efficient hierarchical architecture designed specifically for modeling long-distance dependencies in text classification tasks. Its core structure consists of multiple stacked convolutional blocks with shortcut connections, forming a pyramid-shaped feature extraction structure that combines high performance and computational efficiency.

[0050] The weighted sentiment feature fusion enhancement component adopts a weighted sentiment feature fusion mechanism, aiming to enhance the recognition ability of opinion sentences by integrating the sentiment features, subjective features and contextual features in the text. As subjective statements, opinion sentences usually contain the author's emotional tendencies, subjective judgments and contextual context. Therefore, combining text features and sentiment features can more comprehensively capture the characteristics of opinion sentences.

[0051] A weighted fusion mechanism is employed to perform a weighted fusion of the outputs of the local text feature extraction component and the weighted sentiment feature fusion enhancement component, combining them into a unified representation. This fused representation is then mapped to a two-dimensional space via a fully connected layer, and opinion sentences are classified via a softmax layer. This fusion mechanism organically combines sentiment features with textual contextual features, enabling the model to more accurately capture the implicit emotions and subjective expressions in opinion sentences. This comprehensive representation of multi-dimensional features enhances the model's robustness and classification capabilities, particularly with respect to the complex sentiment implicit in academic opinion sentences.

[0052] The context feature extraction component, the local text feature extraction component, and the weighted sentiment feature fusion enhancement component are connected to obtain a complete opinion sentence recognition model.

[0053] Furthermore, constructing a local text feature extraction component within the opinion sentence recognition model includes:

[0054] Step S2221: using a deep pyramid convolutional neural network, wherein the deep pyramid convolutional neural network is composed of a text region embedding layer and a plurality of convolutional blocks with shortcut connections stacked together, presenting a pyramid-shaped feature extraction structure;

[0055] Step S2222: Construct a text region embedding layer to extract local context features across multiple words from the hidden layer output of AgriBERT. The text region embedding layer adopts F num Filters and convolution kernels of size H×d, where H is the height of the convolution kernel and d is the embedding dimension of the hidden layer;

[0056] Step S2223: Construct a convolution block with shortcut connection to extract deep text features layer by layer. Each convolution block consists of two convolution layers, and each layer uses F num Filter and convolution kernel of size H×H;

[0057] Step S2224: A fixed feature map downsampling mechanism is used to control the computational complexity of the model and maintain key features. After each convolution block, the model performs downsampling through a pooling operation with a stride of 2, reducing the size of the internal representation by half while maintaining the same number of feature maps. The number of repetitions T of convolution and downsampling can be calculated as follows:

[0058]

[0059] Where L is the sequence length.

[0060] In the original deep pyramid convolutional neural network, this layer captures larger text block information by converting the input text into a regional embedding representation, rather than just embedding a single word. In this study, this layer is transformed to extract local context features from the hidden layer output of AgriBERT. Specifically, F num The convolution operation is performed with a filter and a convolution kernel of size H×d, where H is the size of the convolution kernel and d is the embedding dimension of the hidden layer. This design can effectively capture local features across multiple words and provide high-quality input for subsequent convolution operations.

[0061] Each convolution block consists of two convolutional layers, each using F num The shortcut connection adopts a pre-activation mechanism, that is, applying the activation function before the convolution operation, which helps to accelerate convergence and improve the stability of the training process.

[0062] After each convolution block, a pooling layer with a stride of 2 is configured to perform downsampling through the maximum pooling operation, halving the size of the internal representation while maintaining the same number of feature maps. This pyramid-like downsampling reduces the computational cost while retaining important contextual information.

[0063] The downsampling and convolution block operations are repeated until the size of the internal representation reaches the minimum feasible value. The number of repetitions T can be calculated by the following formula:

[0064]

[0065] Where L is the sequence length.

[0066] The final output of the deep pyramid convolutional neural network is recorded as in Represents the feature corresponding to the i-th filter, F num Represents the total number of filters. Through this structure, the deep pyramid convolutional neural network can effectively capture the hierarchical features of text at multiple scales, providing strong feature support for the deep semantic modeling of opinion sentences.

[0067] Furthermore, constructing a weighted sentiment feature fusion enhancement component within the opinion sentence recognition model includes:

[0068] Step S2231: Extract multi-dimensional emotional features: including positive emotional score S pos , negative sentiment score S neg , sentiment polarity score S pol and subjective score S sub, to characterize the emotional tendency and subjectivity of the text, wherein the positive sentiment score and negative sentiment score are calculated by the VADER Sentiment tool, and the emotional polarity score and subjectivity score are calculated by the TextBlob tool;

[0069] Step S2232: weighted fusion of text features and sentiment features: by fusion of weighted sentiment features to enhance the fusion mechanism, the text features V extracted by the local text feature extraction component are combined into a text With the above emotional characteristics S pos , S neg , S pol and S sub Perform joint representation to generate a unified feature vector As follows:

[0070] V fusion =Concat(λ text V text ,λ senti ⊙S senti );

[0071] Among them, is the learnable weight of text features; S senti =(S pos ,S neg ,S pol ,S sub ) is the sentiment feature vector, λ senti =(λ pos ,λ neg ,λ pol ,λ sub ) is the learnable weight of the sentiment feature vector; ⊙ is the element-by-element multiplication operation, and Concat(·) is the vector concatenation operation;

[0072] Step S2233: Opinion sentence recognition based on fusion representation: The weighted fusion vector V fusion Opinion sentence recognition is achieved by mapping it to a two-dimensional space through a fully connected layer and normalizing it through a Softmax function.

[0073] Construct a positive sentiment score S pos , negative sentiment score S neg , sentiment polarity score S pol and subjective score S sub The calculation mechanism is to capture the emotional and subjective information of different dimensions in the text, where the positive sentiment score S pos , the value range is [0,1], indicating the intensity of positive sentiment; negative sentiment score S neg , the value range is [0,1], indicating the intensity of negative sentiment; the sentiment polarity score S pol, the value range is [-1,1], indicating the overall polarity of the emotion, from negative to positive; the subjective score S sub , with a value range of [0,1], indicating the degree to which a text is objective or subjective. Positive and negative sentiment scores are calculated using the VADER Sentiment tool, while sentiment polarity and subjectivity scores are calculated using the TextBlob tool. These sentiment features can characterize the sentiment and subjectivity of a sentence from multiple dimensions.

[0074] Through the weighted sentiment feature fusion enhancement fusion mechanism, the text features extracted by the local text feature extraction component are combined with the output of the weighted sentiment feature fusion enhancement component to generate a unified feature vector. The fusion process is expressed by the following formula:

[0075] V fusion =Concat(λ text V text ,λ senti ⊙S senti );

[0076] The element-by-element multiplication operation ⊙ is used to ensure that each sentiment feature is properly adjusted according to its weight. Finally, the weighted text features and sentiment features are combined into a unified representation V through the Concat function. fusion The fusion representation is mapped to a two-dimensional space through a fully connected layer. This fusion mechanism organically combines sentiment features with text context features, enabling the model to more accurately capture the implicit emotions and subjective expressions in opinion sentences.

[0077] The weighted fusion vector V fusion Opinion sentence recognition is achieved by mapping it to a two-dimensional space through a fully connected layer and normalizing it through a Softmax function. This comprehensive representation of multi-dimensional features enhances the robustness and classification ability of the model, especially for the complex emotional expressions implied in academic opinion sentences.

[0078] In this study's comparative experiments, we selected the model with the highest F1 score on the validation set for test set evaluation. The experimental results showed that the AgriBERT-SentiDPCNN model performed best in the opinion sentence recognition task, achieving an F1 score of 90.19%. In comparison, the baseline model, AgriBERT-DPCNN, achieved an F1 score of 89.50%, further validating the effectiveness of improving model performance through the weighted sentiment feature fusion mechanism.

[0079] Compared with traditional neural network models, AgriBERT-SentiDPCNN has significant advantages. For example, the F1 values ​​of FastText and TextCNN are 64.31% and 67.58% respectively, which shows that models that rely solely on standard embeddings and lack large-scale pre-training support are significantly limited in their performance on complex tasks. Among pre-trained language models, AgriBERT performed best with an F1 value of 88.72%, outperforming BERT (87.97%) and DeBERTa (87.90%). This verifies the high efficiency of domain-specific language models (such as AgriBERT) in domain-specific tasks, which can provide feature representations that are more suitable for domain tasks.

[0080] The hybrid neural network model further improves performance over traditional models. The baseline model, AgriBERT-DPCNN, achieves an F1 score of 89.50%, the best among the hybrid models. This demonstrates the robustness of DPCNN's hierarchical feature extraction capabilities for opinion recognition tasks. Furthermore, AgriBERT-CNN and AgriBERT-RCNN achieve F1 scores of 89.23% and 88.99%, respectively, also demonstrating strong competitiveness.

[0081] The AgriBERT-SentiDPCNN model also outperforms large language models (LLMs). For example, Llama3-70b and GPT-4 achieve F1 scores of 75.21% and 75.30%, respectively. Although LLMs demonstrate high versatility across multiple tasks, their performance on the opinion sentence recognition task is lower than that of specialized domain-optimized models because they are based solely on prompts without fine-tuning.

[0082] In summary, the experimental results show that AgriBERT-SentiDPCNN performs best in the opinion sentence recognition task, significantly outperforming not only traditional neural network models and pre-trained language models, but also untuned large language models, demonstrating the powerful performance advantages of domain-specific models in complex tasks.

[0083] In summary, the method for identifying opinion sentences in scientific literature based on sentiment enhancement provided by the embodiments of the present application has the following technical effects:

[0084] By extracting the target documents to be analyzed from scientific and technological literature, efficient management of input data is achieved, ensuring that the data of the input model is targeted; constructing an opinion sentence recognition model, which includes a contextual feature extraction component, a local text feature extraction component, and a weighted sentiment feature fusion enhancement component, which can realize multi-level text information extraction and processing. This design helps to fully capture the semantic information, contextual associations and sentiment features in the target documents, and provide multi-dimensional feature support for opinion sentence recognition; the contextual feature extraction component can accurately capture the implicit information in the sentence, improving the model's ability to capture domain terms and context. , improving the accuracy of opinion sentence recognition; the local text feature extraction component can accurately capture the semantic features of different scales in the sentence, and is especially suitable for the complex language expressions in the opinion sentences. It can effectively extract the deep semantic information across multiple words in the opinion sentences, and improve the understanding of complex opinion sentences; the weighted sentiment feature fusion enhancement component can accurately identify the implicit sentiment features in the opinion sentences, and improve the model's ability to recognize opinion sentences; after the target document is input into the model, the context feature extraction, local text feature extraction component and weighted sentiment feature fusion enhancement component work together to achieve efficient classification of each sentence and improve the recognition effect of opinion sentences.

[0085] Example 2, based on the same inventive concept as the method for identifying opinion sentences in scientific literature based on emotion enhancement in the above embodiment, Figure 3 As shown, the embodiment of the present application provides a system for identifying opinion sentences in scientific and technological literature based on emotion enhancement, the system comprising:

[0086] A target document acquisition module 10 is used to acquire a target document for opinion sentence recognition;

[0087] The recognition model construction module 20 is used to construct an opinion sentence recognition model, wherein the opinion sentence recognition model includes a context feature extraction component, a local text feature extraction component, and a weighted sentiment feature fusion enhancement component;

[0088] The recognition result acquisition module 30 is used to input the target document into the opinion sentence recognition model and output the opinion sentence recognition result.

[0089] Furthermore, the recognition model building module 20 includes:

[0090] Corpus construction unit, used to construct the opinion sentence identification corpus;

[0091] A module construction unit, used to construct a context feature extraction component, a local text feature extraction component, and a weighted sentiment feature fusion enhancement component within the opinion sentence recognition model;

[0092] A model training unit is used to train the opinion sentence recognition model on a training set based on the opinion sentence recognition corpus, screen the model parameters with the best performance on a validation set, and perform performance evaluation on the final opinion sentence recognition model on a test set.

[0093] Furthermore, the corpus construction unit includes:

[0094] Sample document data acquisition channel, used to obtain sample document data sets;

[0095] The annotation channel is used to manually annotate the sentences in each sample document dataset, distinguishing opinion sentences from factual sentences, and constructing an opinion sentence recognition corpus;

[0096] The partitioning channel is used to divide the constructed opinion sentence recognition corpus into training sets, validation sets, and test sets according to experimental requirements to support the training and evaluation of the opinion sentence recognition model.

[0097] Furthermore, the module building unit includes:

[0098] The context feature extraction component construction channel is used to construct the context feature extraction component in the opinion sentence recognition model. AgriBERT is used as the Embedding layer to vectorize the input text. For the input sequence S = {s1,...,s i ,...,s L}, and its output sequence is represented as X={x1,...,x i ,...,x L}, where L is the length of the input sequence, s i is the i-th token in the input sequence, is the corresponding representation vector, and d is the hidden layer dimension;

[0099] A local text feature extraction component construction channel is used to construct the local text feature extraction component in the opinion sentence recognition model;

[0100] A weighted sentiment feature fusion enhancement component construction channel is used to construct the weighted sentiment feature fusion enhancement component in the opinion sentence recognition model;

[0101] The opinion sentence recognition model acquisition channel is used to connect the context feature extraction component, the local text feature extraction component and the weighted sentiment feature fusion enhancement component to obtain the opinion sentence recognition model.

[0102] Furthermore, the local text feature extraction component constructs a channel, including:

[0103] A deep convolutional neural network construction node is used to adopt a deep pyramid convolutional neural network. The deep pyramid convolutional neural network is composed of a text region embedding layer and multiple convolution blocks with shortcut connections, presenting a pyramid-shaped feature extraction structure;

[0104] The text region embedding layer construction node is used to construct a text region embedding layer for extracting local context features across multiple words from the hidden layer output of AgriBERT. The text region embedding layer adopts F num Filters and convolution kernels of size H×d, where H is the height of the convolution kernel and d is the embedding dimension of the hidden layer;

[0105] Convolutional block construction node is used to build a convolutional block with shortcut connection to extract deep text features layer by layer. Each convolutional block consists of two convolutional layers, each layer uses F num Filters and convolution kernels of size H×H are connected through shortcuts to achieve feature transfer;

[0106] The downsampling mechanism uses a node to adopt a fixed feature map downsampling mechanism to control the computational complexity of the model and maintain key features. After each convolution block, the model performs downsampling through a pooling operation with a stride of 2, reducing the size of the internal representation by half while maintaining the same number of feature maps. The number of repetitions T of convolution and downsampling can be calculated by the following formula:

[0107]

[0108] Where L is the sequence length.

[0109] Furthermore, the weighted emotional feature fusion enhancement component constructs a channel, including:

[0110] Multi-dimensional sentiment feature extraction node, used to extract multi-dimensional sentiment features: including positive sentiment score S pos , negative sentiment score S neg , sentiment polarity score S pol and subjective score S sub , to characterize the emotional tendency and subjectivity of the text, wherein the positive sentiment score and negative sentiment score are calculated by the VADER Sentiment tool, and the emotional polarity score and subjectivity score are calculated by the TextBlob tool;

[0111] Weighted fusion node, used for weighted fusion of text features and sentiment features: by enhancing the fusion mechanism through weighted sentiment feature fusion, the text features V extracted by the local text feature extraction component are text With the above emotional characteristics S pos, S neg , S pol and S sub Perform joint representation to generate a unified feature vector As follows:

[0112] V fusion =Concat(λ text V text ,λ senti ⊙S senti );

[0113] Among them, is the learnable weight of text features; S senti =(S pos ,S neg ,S pol ,S sub ) is the sentiment feature vector, λ senti =(λ pos ,λ neg ,λ pol ,λ sub ) is the learnable weight of the sentiment feature vector; ⊙ is the element-by-element multiplication operation, and Concat(·) is the vector concatenation operation;

[0114] Opinion sentence recognition node, used for opinion sentence recognition based on fusion representation: the weighted fusion vector V fusion Opinion sentence recognition is achieved by mapping it to a two-dimensional space through a fully connected layer and normalizing it through a Softmax function.

[0115] Through the above detailed description of a method for identifying opinion sentences in scientific and technological literature based on emotion enhancement, those skilled in the art can clearly understand a system for identifying opinion sentences in scientific and technological literature based on emotion enhancement in this embodiment. Since it corresponds to the method disclosed in the embodiment, the description is relatively simple. For relevant matters, please refer to the method section.

[0116] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present application. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application is not limited to the embodiments shown herein, but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for identifying opinion sentences in scientific literature based on sentiment enhancement, characterized in that: The method comprises: Obtain the target document for opinion sentence recognition; Constructing an opinion sentence recognition model, wherein the opinion sentence recognition model includes a context feature extraction component, a local text feature extraction component, and a weighted sentiment feature fusion enhancement component; Inputting the target document into the opinion sentence recognition model and outputting an opinion sentence recognition result; Construct an opinion sentence recognition model, including: Construct a corpus for identifying opinion sentences; Constructing a context feature extraction component, a local text feature extraction component, and a weighted sentiment feature fusion enhancement component within the opinion sentence recognition model; Based on the opinion sentence recognition corpus, the opinion sentence recognition model is trained on a training set, model parameters with optimal performance are selected on a validation set, and the performance of the final opinion sentence recognition model is evaluated on a test set; Constructing the context feature extraction component, local text feature extraction component, and weighted sentiment feature fusion enhancement component within the opinion sentence recognition model includes: Construct the context feature extraction component in the opinion sentence recognition model, use AgriBERT as the Embedding layer to vectorize the input text, and for the input sequence S = {s1,...,s i ,...,s L }, and its output sequence is represented as X={x1,...,x i ,...,x L }, where L is the length of the input sequence, s i is the i-th token in the input sequence, x i ∈R d is the corresponding representation vector, d is the hidden layer dimension; Constructing a local text feature extraction component within the opinion sentence recognition model; Constructing a weighted sentiment feature fusion enhancement component within the opinion sentence recognition model; The context feature extraction component, the local text feature extraction component and the weighted sentiment feature fusion enhancement component are connected to obtain an opinion sentence recognition model.

2. The method for identifying opinion sentences in scientific and technological literature based on emotion enhancement according to claim 1 is characterized in that: Construct an opinion sentence recognition corpus, including: Obtain sample document dataset; Manually annotate the sentences in each sample document dataset to distinguish opinion sentences from factual sentences, and build an opinion sentence recognition corpus. According to experimental requirements, the constructed opinion sentence recognition corpus is divided into training set, validation set and test set to support the training and evaluation of the opinion sentence recognition model.

3. The method for identifying opinion sentences in scientific literature based on emotion enhancement according to claim 2 is characterized in that: Constructing a local text feature extraction component within the opinion sentence recognition model includes: A deep pyramid convolutional neural network is used. The deep pyramid convolutional neural network is composed of a text region embedding layer and multiple convolutional blocks with shortcut connections, presenting a pyramid-shaped feature extraction structure. Construct a text region embedding layer to extract local context features across multiple words from the hidden layer output of AgriBERT. The text region embedding layer adopts F num Filters and convolution kernels of size H×d, where H is the height of the convolution kernel and d is the embedding dimension of the hidden layer; Construct a convolution block with shortcut connection to extract deep text features layer by layer. Each convolution block consists of two convolution layers, and each layer uses F num Filters and convolution kernels of size H×H are connected through shortcuts to achieve feature transfer; A fixed feature map downsampling mechanism is used to control the computational complexity of the model and maintain key features. After each convolution block, the model performs downsampling through a pooling operation with a stride of 2, halving the size of the internal representation while maintaining the same number of feature maps. The number of repetitions T of convolution and downsampling can be calculated as follows: Where L is the sequence length.

4. The method for identifying opinion sentences in scientific literature based on emotion enhancement according to claim 3 is characterized in that: Constructing a weighted sentiment feature fusion enhancement component within the opinion sentence recognition model includes: Extract multi-dimensional emotional features: including positive emotional score S pos , negative sentiment score S neg , sentiment polarity score S pol and subjective score S sub , to characterize the emotional tendency and subjectivity of the text; wherein the positive emotional score and negative emotional score are calculated by the VADER Sentiment tool, and the emotional polarity score and subjectivity score are calculated by the TextBlob tool; Weighted fusion of text features and sentiment features: By fusion of weighted sentiment features to enhance the fusion mechanism, the text features V extracted by the local text feature extraction component are combined into a text With the above emotional characteristics S pos , S neg , S pol and S sub Perform joint representation to generate a unified feature vector As follows: V fusion =Concat(λ text V text ,λ senti ⊙S senti ); Among them, λ text is the learnable weight of text features; S senti =(S pos ,S neg ,S pol ,S sub ) is the sentiment feature vector, λ senti =(λ pos ,λ neg ,λ pol ,λ sub ) is the learnable weight of the sentiment feature vector; ⊙ is the element-by-element multiplication operation, and Concat(·) is the vector concatenation operation; Opinion sentence recognition based on fusion representation: The weighted fusion vector V fusion Opinion sentence recognition is achieved by mapping it to a two-dimensional space through a fully connected layer and normalizing it through a Softmax function.

5. A system for identifying opinion sentences in scientific literature based on sentiment enhancement, characterized in that: A system for implementing a method for identifying opinion sentences in scientific and technological literature based on sentiment enhancement according to any one of claims 1 to 4, comprising: A target document acquisition module is used to acquire the target document for opinion sentence recognition; A recognition model construction module is used to construct an opinion sentence recognition model, wherein the opinion sentence recognition model includes a context feature extraction component, a local text feature extraction component, and a weighted sentiment feature fusion enhancement component; A recognition result acquisition module is used to input the target document into the opinion sentence recognition model and output the opinion sentence recognition result; The recognition model building module includes: Corpus construction unit, used to construct the opinion sentence identification corpus; A component construction unit, used to construct a context feature extraction component, a local text feature extraction component, and a weighted sentiment feature fusion enhancement component within the opinion sentence recognition model; A model training unit, configured to train the opinion sentence recognition model on a training set based on the opinion sentence recognition corpus, select model parameters with optimal performance on a validation set, and perform performance evaluation on the final opinion sentence recognition model on a test set; The component building unit includes: The context feature extraction component construction channel is used to construct the context feature extraction component in the opinion sentence recognition model. AgriBERT is used as the Embedding layer to vectorize the input text. For the input sequence S = {s1,...,s i ,...,s L }, and its output sequence is represented as X={x1,...,x i ,...,x L }, where L is the length of the input sequence, s i is the i-th token in the input sequence, is the corresponding representation vector, and d is the hidden layer dimension; A local text feature extraction component construction channel is used to construct the local text feature extraction component in the opinion sentence recognition model; A weighted sentiment feature fusion enhancement component construction channel is used to construct the weighted sentiment feature fusion enhancement component in the opinion sentence recognition model; The opinion sentence recognition model acquisition channel is used to connect the context feature extraction component, the local text feature extraction component and the weighted sentiment feature fusion enhancement component to obtain the opinion sentence recognition model.

6. The system for identifying opinion sentences in scientific and technological literature based on emotion enhancement according to claim 5 is characterized in that: The corpus construction unit includes: Sample document data acquisition channel, used to obtain sample document data sets; The annotation channel is used to manually annotate the sentences in each sample document dataset, distinguishing opinion sentences from factual sentences, and constructing an opinion sentence recognition corpus; The partitioning channel is used to divide the constructed opinion sentence recognition corpus into training sets, validation sets, and test sets according to experimental requirements to support the training and evaluation of the opinion sentence recognition model.

Images