Chinese composition processing method and device

Abstract

The invention relates to the technical field of natural language processing, and discloses a Chinese composition processing method and device, and the method comprises the following steps: S1, obtaining composition data submitted by a student, and generating a digital text from paper and image files through an optical character recognition technology; s2, preprocessing the digital text, including paragraph division, sentence group disassembly, noise character removal, punctuation mark standardization and text sentence segmentation processing; and S3, inputting the sentence group into the semantic coding model to obtain corresponding semantic vectors, and constructing a semantic vector field consisting of a plurality of semantic vectors. By adopting a technical path of constructing a semantic vector field and carrying out scoring modeling based on a deep semantic relationship, the technical effects of dynamically capturing a composition logic structure and truly reflecting semantic coherence are achieved. Compared with a traditional scoring method depending on static word frequency or word bank comparison, the method solves the problems of weak understanding of the context and one-sided scoring.

Description

Technical Field

[0001] The present invention relates to the technical field of natural language processing, and in particular to a method and device for processing Chinese composition. Background Art

[0002] The processing of Chinese language essays still largely relies on manual grading or automated grading systems based on fixed templates. Although some systems have attempted to incorporate natural language processing technology to improve grading efficiency, overall processing solutions still have significant limitations in terms of accuracy, intelligence, and feedback effectiveness. Existing solutions struggle to meet the comprehensive requirements of "understanding, judgment, and suggestions" for grading medium- to high-quality Chinese language essays. Their shortcomings are reflected in the following aspects.

[0003] Traditional essay recognition systems generally rely on optical character recognition (OCR) methods based on image templates. These methods typically combine character segmentation and pattern matching to extract text. While they offer a reasonable recognition rate for printed text, they significantly decrease when dealing with handwritten text, images with traces of erasure, or images taken from distorted angles. This is particularly true in middle school essays, where handwriting styles vary widely and connected characters are common. Traditional algorithms are prone to misjudgments and omissions, resulting in a high level of noise in the recognition results, severely impacting subsequent semantic processing.

[0004] On the other hand, current essay scoring technology primarily relies on static feature extraction and rule-based matching, such as word frequency statistics, keyword matching, and grammatical template comparison. These methods lack contextual modeling capabilities and offer a very limited understanding of the semantic intent behind words, making it difficult to capture the underlying logical structure and thematic development within students' writing. For example, even if an essay is coherent, if it employs flexible sentence structures or innovative structures, it will often receive a low score in existing scoring models due to "deviation from the template," thus inhibiting the rationality of individualized expression.

[0005] Furthermore, existing writing systems often present feedback in the form of overall scores or rough sub-scores, lacking the ability to identify and explain specific problematic sentences. Many systems offer only fixed suggestions, such as "lack of vocabulary" or "incoherent sentences," without specifying the specific problematic sentences or suggesting alternative expressions. This templated, abstract feedback approach lacks specificity and a path for improvement, making it difficult to help students truly understand and improve their expression skills. Summary of the Invention

[0006] In response to the deficiencies of the existing technology, the present invention provides a method and device for processing Chinese compositions, which solves the problems of low recognition accuracy, weak semantic understanding ability, and one-sided scoring mechanism in the existing composition processing methods.

[0007] To achieve the above objectives, the present invention is implemented through the following technical solutions: a method and device for processing Chinese composition, comprising the following steps: S1. Obtain student-submitted essay data and generate digital text from paper and image files using optical character recognition technology; S2. Preprocessing of digital text, including paragraph division, sentence group decomposition, noise character removal, punctuation standardization, and text sentence segmentation; S3. Input the sentence group into the semantic encoding model, obtain the corresponding semantic vector, and construct a semantic vector field composed of multiple semantic vectors; S4. Calculate the semantic relevance between semantic units within the composition based on the semantic vector field; S5. comprehensively scoring the composition text using a scoring function of multiple scoring dimensions; S6. Generate composition optimization suggestions based on the optimization goal of the scoring function.

[0008] Preferably, the optical character recognition technology in step S1 is a convolutional neural network or a long short-term memory network model based on a deep learning algorithm.

[0009] Preferably, the pre-processing in step S2 cleans the composition text of noise, including removing non-text symbols, redundant spaces, and meaningless line breaks in the text, and performing paragraph segmentation to ensure that the sentence group structure within the paragraph is correct.

[0010] Preferably, the semantic encoding model in step S3 adopts a pre-trained Transformer model, which can generate high-dimensional semantic vectors based on the context, and one sentence group corresponds to one semantic vector.

[0011] Preferably, the semantic vector field in step S3 includes multiple semantic vectors, the semantic vectors have similarity in the semantic space, and the dimension of the semantic vector is d, where d is a characteristic dimension of the semantic vector.

[0012] Preferably, the method for calculating the semantic relevance in step S4 is: calculating the similarity between the semantic vectors of each pair of sentence groups, and the similarity is measured by cosine similarity, and the formula is: ; in: Indicates the The semantic vector of a sentence; Indicates the The semantic vector of a sentence; Represents a vector and The dot product of Represents a vector Length of the module; Represents a vector Length of the module; sim express and The cosine similarity of .

[0013] Preferably, the scoring function in step S5 is a weighted function covering multiple scoring dimensions, including: Semantic coherence score, which measures the semantic consistency between sentence groups; Logical scoring, assessing the logical relationship between the argument and evidence in the essay; The language diversity score measures the richness of vocabulary and sentence structure in the essay; Grammar standardization scoring, checking grammatical errors and writing standardization in the composition.

[0014] Preferably, the optimization suggestion in step S6 is directed optimization based on various scoring indicators of the scoring function, including the following contents: Regarding semantic coherence, it is recommended to modify the unnatural transitions between sentence groups; Regarding logic, it is recommended to strengthen the argument support and reasoning process in the argument part; In view of language diversity, it is recommended to use diverse vocabulary or sentence structures; Regarding grammatical standardization, specific grammatical errors and suggested modifications are listed.

[0015] Preferably, the generation process of the optimization suggestions in step S6 is customized according to the user-specified goals, including: style optimization, refinement of language expression, and improvement of argument structure.

[0016] The Chinese composition processing device includes: A data acquisition module is used to obtain the composition data submitted by students and generate digital text through optical character recognition technology; The text preprocessing module is connected to the data acquisition module and is used to divide the digital text into sentences, paragraphs and clear noise; The semantic analysis module is connected to the text preprocessing module and is used to input the segmented text into the semantic encoding model to generate semantic vectors; The scoring module is connected to the semantic analysis module and is used to comprehensively score the essay based on the scoring function; The optimization suggestion generation module is connected to the scoring module and is used to generate and output targeted optimization suggestions based on the scoring results; The display module is connected to the optimization suggestion generation module and is used to display the composition data and optimization suggestions.

[0017] The present invention provides a method and device for processing Chinese composition, which has the following beneficial effects: 1. By constructing a semantic vector field and performing scoring modeling based on deep semantic relationships, this paper achieves the technical effect of dynamically capturing the logical structure of essays and truly reflecting semantic coherence. Compared with traditional scoring methods that rely on static word frequency or vocabulary comparison, this method solves the problems of weak understanding of context and one-sided scoring.

[0018] 2. This invention achieves highly targeted, intuitive optimization output by employing a personalized suggestion matching mechanism driven by itemized scoring. Existing technologies often only provide general feedback or templated suggestions, failing to pinpoint specific sentence or segment issues. This invention achieves a breakthrough in providing precise guidance and item-by-item improvement.

[0019] 3. By integrating OCR recognition with an LSTM neural network, this invention achieves the technical effect of efficiently extracting high-precision digital text from essays in various formats. Compared to existing character recognition methods based solely on rule-based templates, this solves the problem of low recognition accuracy in cases of poor image quality or irregular handwriting.

[0020] 4. This invention achieves a clear composition evaluation structure and intuitive user operation by combining graphical structure output with interactive annotation display. Traditional methods often present feedback in text lists, which fragments the information and makes it difficult for users to compare, understand, and quickly make changes. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 is a flow chart of the method of the present invention; Figure 2 This is a system framework diagram of the present invention. DETAILED DESCRIPTION

[0022] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the present specification. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0023] Please see the attached Figure 1 The embodiment of the present invention provides a method and device for processing a Chinese composition, comprising the following steps: S1. Obtain student-submitted essay data and generate digital text from paper and image files using optical character recognition technology; Specifically, in this embodiment, step S1 is to obtain the composition data submitted by the student through optical character recognition (OCR). This process mainly includes the technical solutions of image data preprocessing, text recognition and conversion into digital text.

[0024] First, the system uses an image acquisition module to capture student-submitted essay data, which can be either paper documents or scanned image files. This image data is then processed using optical character recognition (OCR) technology, converting it into editable digital text. This OCR technology is based on a deep learning algorithm, specifically using a long short-term memory (LSTM) model for character recognition.

[0025] The LSTM model is a deep learning model based on a recurrent neural network (RNN). Its structure possesses effective memory capabilities, allowing it to maintain contextual dependencies over long periods of time when processing sequential data. By training on large amounts of annotated data, the LSTM can efficiently recognize characters in images and convert them into corresponding digital text.

[0026] The core of optical character recognition (OCR) is to locate, identify, and convert characters in an image. This process involves the following mathematical models and algorithms: Image preprocessing: First, the acquired image is grayscaled, binarized, and denoised. These preprocessing steps help reduce image noise and improve the accuracy of subsequent character recognition.

[0027] Character region localization: A convolutional neural network (CNN) is used to extract features from the input image to identify character regions. This process uses a sliding window method to decompose the input image into multiple small blocks, gradually performing feature detection to obtain candidate regions for each character.

[0028] Character Recognition: Based on an LSTM network, each character region is identified. LSTM has strong sequence modeling capabilities and can process the relationships between characters in an image. Through training, LSTM can recognize characters presented in an image and convert them into digital text.

[0029] Output digital text: Finally, the corresponding character sequence is output through the LSTM network to obtain the complete digital text representation.

[0030] Formula derivation and variable definition For character recognition in images, the key mathematical formula is the state update equation in the LSTM model. Assume that the input image data is (in is the time step), the state update equation of LSTM is: ; ; ; ; ; in: It is the output of the forget gate, which determines how much of the memory content of the previous moment is retained; The output of the input gate determines how much of the current input content is remembered; is an activation function, and the output value is between [0,1]; is the hidden state at the previous moment is the output gate, which determines the output at the current moment; is the unit state, indicating the memory content at the current moment; is the hidden state, which represents the network output at the current moment; It is a hyperbolic tangent activation function that controls the scale of information; The input at the current moment.

[0031] LSTM training is done by optimizing the parameters in the above gating function 、 、 、 and bias 、 、 、 This enables the network to correctly recognize character sequences in images.

[0032] In this embodiment, the image acquisition module and the OCR processing module are connected via a data transmission interface. The image acquisition module is responsible for acquiring data from paper documents or scanned images and transmitting it to the OCR module for character recognition. The LSTM network within the OCR module processes the input data and outputs the corresponding digital text.

[0033] In this embodiment, step S1 realizes the OCR recognition of the paper or image documents submitted by students by using the LSTM model in deep learning. This method can not only efficiently recognize the characters in the image, but also ensure the accuracy of the recognition result through the advantage of the LSTM network in time series modeling. Through the collaborative work of the image acquisition module and the OCR processing module, the system can automatically and accurately convert the composition image of the student into digital text, providing reliable input for the subsequent processing steps.

[0034] S2. Preprocess the digital text, including paragraph division, sentence group disassembling, removing noise characters, standardizing punctuation marks, and text clause processing; Specifically; in this embodiment, step S2 processes the digital text data generated in step S1, thereby extracting the features related to the composition quality and further transforming them into a scoring feature vector. This process mainly includes the technical solutions of text preprocessing, feature extraction, and feature vectorization. The specific implementation method is as follows: First, the digital text data generated in step S1 is passed to the text preprocessing module, which converts the text into structured information through basic processing such as word segmentation, part-of-speech tagging, and stop word removal. Subsequently, through the feature extraction module, multiple scoring features of the composition are extracted from the text, including grammatical structure, semantic accuracy, lexical richness, sentence structure, etc.

[0035] In the process of feature extraction, the system preferably uses the TF-IDF (term frequency-inverse document frequency) method to calculate the importance of each word in the composition and generates the corresponding lexical features based on this. Then, through the convolutional neural network (CNN), high-level features in the text, such as context information, syntactic structure, etc., are extracted to further enhance the expressiveness of the features; Text preprocessing: First, segment the composition data, disassembling each sentence into individual words. Then perform part-of-speech tagging to identify the grammatical roles of each word, such as nouns, verbs, etc. Then remove stop words, such as meaningless words like "de", "le", etc.

[0036] Feature extraction: Feature extraction is performed on the text based on the TF-IDF model. The calculation formula of the TF-IDF model is as follows: ; ; ; Where: is the frequency of word www in the document.

[0037] is the inverse document frequency, measuring the importance of word www in the document library.

[0038] is the final weight of the word www, indicating the importance of the word in the composition; It is a logarithmic function, usually with the base being the natural logarithm or base 10.

[0039] Convolutional Neural Network (CNN) feature extraction: CNN models are used to further extract high-level features from text. CNNs use convolution operations to find specific syntactic patterns and contextual information in text, thereby obtaining deeper feature representations.

[0040] Feature vectorization: Finally, all extracted scoring features are converted into scoring feature vectors through the feature vectorization module. This vector contains information such as grammar, semantics, syntax, and vocabulary of the essay, which serves as the input for the subsequent scoring model. The feature extraction process involves the calculation formula in the TF-IDF model. The above formula describes how to calculate the weight of each word by using word frequency and inverse document frequency, and then evaluate the importance of the word to the essay score. In addition, the convolution operation of CNN can be described by the following formula: ; in: is the input feature vector.

[0041] is the convolution kernel, which represents the important pattern extracted from the input features; It is a bias term to ensure the accuracy of the output.

[0042] is the output of the convolutional layer, representing the high-level features extracted from the text; is the activation function.

[0043] Through multi-layer convolution and pooling operations, CNN can effectively extract high-level grammatical and contextual features from text.

[0044] In this embodiment, step S2 successfully converts the digital text data generated in step S1 into a scoring feature vector through a multi-layered process involving text preprocessing, feature extraction, and feature vectorization. This feature vector integrates information about the composition's grammatical, semantic, and syntactic structure, providing reliable data support for the subsequent scoring algorithm. The combination of TF-IDF and CNN ensures the efficiency and accuracy of the feature extraction process.

[0045] S3. Input the sentence group into the semantic encoding model, obtain the corresponding semantic vector, and construct a semantic vector field composed of multiple semantic vectors; Specifically, after text preprocessing (step S2) in this embodiment is complete, the system now possesses digitized text data with a clear structure and standardized symbols. To conduct in-depth analysis of the essay's semantic structure, thereby supporting subsequent scoring and optimization suggestion generation, semantic modeling and encoding of the preprocessed sentence clusters is required. This process requires not only that the model capture the contextual information of the sentence clusters but also that it be converted into a high-dimensional semantic representation suitable for computer processing. Therefore, in step S3, a semantic encoding model is introduced to extract and vectorize the text's semantic information, constructing a unified semantic vector field that provides the basis for subsequent semantic relevance analysis and scoring.

[0046] After the text is segmented and decomposed into sentence groups, each sentence group is taken as an independent input and input into the semantic encoding model for semantic feature extraction.

[0047] Generally, semantic encoding models employ deep neural networks based on pre-trained language model architectures, preferably using a Transformer architecture with the ability to capture contextual semantics. In specific implementations, general-purpose semantic modeling models with semantic understanding capabilities, such as BERT (Bidirectional Encoder Representations from Transformers) and RoBERTa (Robustly Optimized BERT Pretraining Approach), can be used. These models utilize a self-attention mechanism to synthesize semantic relationships within sentence groups and between contexts, generating vector representations with enhanced semantic expressiveness.

[0048] Specifically, this step aims to extract the contextual semantic features contained in each sentence group and quantify them as high-dimensional vectors, thereby forming a "semantic vector field" within the composition that describes the underlying semantic relationships between semantic units. This structure not only provides basic data for subsequent semantic similarity calculations but also serves as a key intermediary for determining the quality of the composition's semantic structure.

[0049] In this embodiment, the semantic vector is obtained using a pre-trained Transformer encoding model, such as a language modeling architecture such as BERT, RoBERTa or DeBERTa. The model has good context modeling capabilities and can perform semantic aggregation on languages in different contexts.

[0050] In one possible implementation, the system uses each independent sentence cluster (i.e., the continuous sentence group output by the sentence segmentation module) as the model input unit. After multi-layer encoding processing by the Transformer model, the system extracts the corresponding sentence cluster's CLS vector from the hidden layer of the model output, or extracts the representations of all tokens through an average pooling operation and aggregates them to form the final semantic vector representation of the sentence cluster.

[0051] The sentence group semantic vector is recorded as: ; in: Indicates the Semantic vector of a sentence group; Represents the feature dimension of the vector, which is usually the dimension of the model output (such as 768 or 1024) and is a fixed parameter; express dimensional real vector space.

[0052] In some embodiments, a dynamic context window mechanism can be used to jointly encode adjacent sentence groups, that is, the model input is not limited to a single sentence group, but includes the previous and next contexts of the sentence group, thereby enhancing the sensitivity of the semantic vector to context changes.

[0053] The construction of the semantic vector field is completed by combining the semantic vectors of all sentence groups into a set. In mathematical expression, the semantic vector field can be expressed as: ; in: Represents a semantic vector field; Indicates the total number of sentence groups in the composition; Every one For one dimensional semantic vector.

[0054] In some extended implementations, multilingual pre-trained models (such as mBERT and XLM-R) can also be introduced to support semantic encoding of multilingual composition content, thereby expanding the scope of application of this method.

[0055] This step uses a pre-trained Transformer model to encode text sentence groups to form a set of semantic vectors with consistent dimensions and context sensitivity, providing a basic structure for the semantic structure analysis of the composition. S4. Calculate the semantic relevance between semantic units within the composition based on the semantic vector field; Specifically, in this embodiment, after obtaining the semantic vector set output by the semantic encoding model, the semantic vector field construction phase (step S3) is completed. Further work is required to measure and model the internal connections between semantic units within the composition. This process aims to reveal the semantic cohesion and semantic consistency characteristics between different sentence groups within the composition, providing a quantitative basis for subsequent comprehensive scoring based on content comprehension. Therefore, step S4 focuses on calculating the semantic correlations between pairs of sentence groups within the composition based on the semantic vector field.

[0056] Semantic relevance is a direct reflection of a composition's semantic coherence and structural consistency, expressed as the relative similarity between vectors in semantic space. Measuring semantic relevance effectively characterizes the semantic distribution relationships between sentence groups and reveals the overall semantic organization of a composition.

[0057] In this embodiment, to achieve a quantitative measure of semantic relevance, cosine similarity based on the vector space model is used as the core calculation method. Specifically, the system calculates the similarity of the semantic vectors corresponding to each pair of sentence groups to form a semantic similarity matrix between the sentence groups.

[0058] In mathematical expression, let the semantic vectors of any two sentence groups in the semantic vector field be and , then the semantic relevance (i.e. semantic similarity) between them is defined as: ; in: Indicates the The semantic vector of a sentence; Indicates the The semantic vector of a sentence; Represents a vector and The dot product of Represents a vector Length of the module; Represents a vector Length of the module; sim express and The cosine similarity of .

[0059] In general, sim The value range is , where values closer to 1 indicate closer semantics, values close to 0 indicate no obvious association, and negative values indicate possible semantic opposition.

[0060] In some embodiments, in order to suppress the influence of semantic vector amplitude on similarity results, semantic vectors are uniformly normalized before entering the calculation. Specifically: ; ; The above processing can ensure that the similarity value is determined only by the semantic directionality, thereby improving the measurement accuracy.

[0061] Specifically, if there are many sentence groups, in order to control the computational complexity, a threshold mechanism can be introduced into the vector field. For example, setting the semantic similarity threshold ,when When , it is determined to be semantically related. This threshold can be dynamically adjusted based on training data.

[0062] In step S4 of this embodiment, by systematically calculating the similarity between semantic vectors, the structural relationship between semantic units within the composition is quantitatively expressed, and a data basis and modeling support are provided for subsequent scoring functions.

[0063] S5. comprehensively scoring the composition text using a scoring function of multiple scoring dimensions; Specifically, in this embodiment, after constructing the semantic vector field and calculating semantic relevance, the system further utilizes a scoring mechanism to quantify the overall quality of the composition. This step (step S5) involves comprehensively evaluating the composition's quality based on the semantic relationships and structural features between sentence groups. This evaluation not only considers the logic and coherence of the composition's content, but also includes a comprehensive scoring of multiple dimensions, such as grammatical correctness and vocabulary richness. Using these multi-dimensional scoring metrics, the system can provide more accurate and comprehensive feedback on composition quality.

[0064] Specifically, the core task of step S5 is to integrate the semantic similarity matrix calculated above with other scoring criteria to form the final score of the essay. This scoring mechanism relies on a comprehensive consideration of multiple aspects of feature information, such as semantic coherence, content richness, and language expression.

[0065] In this embodiment, the comprehensive scoring function is modeled using a weighted linear model. Specifically, the quality score of the composition can be expressed as: ; in: Indicates the final score of the composition; represents the semantic coherence score of the composition, which is calculated based on the semantic similarity matrix; The fluency score of the composition is obtained by analyzing the grammatical structure and vocabulary usage of the sentence group; Indicates the grammatical correctness score of the composition, evaluating the number and types of grammatical errors in the composition; It represents the lexical richness score of the composition, which examines the variety and complexity of the vocabulary used in the composition; It is the weight coefficient of each score, which can be adjusted according to the actual application scenario.

[0066] Typically, these weights are automatically learned from training data using machine learning methods, such as regression analysis or gradient boosted tree (GBDT) models. During training, the system optimizes these weights based on labeled essay data to maximize consistency between scoring results and human review.

[0067] The overall score can also take into account the relevance of the essay's topic. For example, in a multi-topic essay, a text-based topic model (such as LDA) can be used to assess whether the essay's content is relevant to the topic or requirements, thereby weighting and adjusting the overall score.

[0068] Specifically, each sentence group in the composition can be scored based on its relevance to the topic. The topic relevance score of each sentence group is , the overall topic relevance score of the essay is: ; in: Indicates the total number of sentence groups in the composition; Indicates the The topic relevance score of a sentence group.

[0069] The score can be obtained by comparing the output of the topic model with the topic distribution of the sentence group.

[0070] Step S5 comprehensively considers the multi-dimensional characteristics of the composition and conducts a comprehensive evaluation of the composition from various aspects such as semantic coherence, grammatical structure, vocabulary richness and emotional expression.

[0071] S6. Generate composition optimization suggestions based on the optimization goal of the scoring function; Specifically, after completing the comprehensive processing of essay scoring (step S5), the system needs to display and provide feedback on the final evaluation results. Step S6 primarily involves visualizing the scoring results and providing appropriate feedback to the user. This process includes not only the digital display of the scoring results but also a quality analysis of the essay content, suggestions for improvement, and detailed guidance for specific issues. The core goal of this step is to enhance the user experience, enabling users to accurately understand the strengths and weaknesses of their essays through system feedback and further optimize their writing skills based on this feedback.

[0072] In this embodiment, the system displays the scoring results in a variety of ways, including graphical display, detailed scoring rules, and specific improvement suggestions. These display methods can help users fully understand the various dimensions of the composition and make effective improvements based on the scoring results. Specifically, the system will output the overall score of the composition, and will also display the performance of the composition in various aspects such as semantic coherence, grammatical correctness, and vocabulary richness. To further enhance the user experience, the system will also provide specific feedback and suggestions to guide users to make improvements in the corresponding aspects.

[0073] This embodiment uses a graphical interface to display the scoring results, presenting the essay scores and scoring items in the form of bar charts, radar charts, or other visual graphics to facilitate intuitive understanding by the user. For example, the overall score of an essay can be displayed as a numerical value and compared with the scores of various scoring items such as semantic coherence, grammatical correctness, and fluency.

[0074] Specifically, the final score of the essay The following graphical displays can be used: ; in: Indicates the Rating items; is the weight coefficient of the corresponding scoring item; Indicates the total number of rated items.

[0075] The score for each rating category, combined with its corresponding weight, creates a radar chart or bar graph that clearly illustrates the strengths and weaknesses of each aspect of the essay. Specifically, each axis in the radar chart represents a rating category, with higher scores corresponding to points further from the center, helping users intuitively view their essay's performance across different dimensions.

[0076] As an option, the system can also provide detailed error analysis and suggestions. By automatically analyzing the grammatical errors, structural problems, or inappropriate vocabulary in the composition, the system can provide users with specific improvement suggestions. For example, if the composition contains many grammatical errors, the system can point out the specific grammatical rules and provide corresponding correction suggestions. If the vocabulary is not rich enough, the system will also prompt the user to add more synonyms or more advanced vocabulary.

[0077] In some embodiments, the system not only displays the essay score, but also provides targeted feedback, such as: If the "Grammar" score is low, the system can suggest that users improve it through online grammar checking tools or language learning platforms; When the "lexical richness" score is low, the system can suggest that users use a wider range of vocabulary in their writing, or provide some replacement suggestions for advanced vocabulary.

[0078] Specifically, for different scoring items, feedback suggestions can be generated using the following formula: ; in: Indicates that for Feedback suggestions generated by each rating item; is the score of the rating item; The improvement threshold for this scoring item. If the score is lower than this threshold, an improvement suggestion is triggered. A function that generates recommendations, specifically generating different recommendations based on the relationship between the score and the threshold.

[0079] In this embodiment, step S6 not only clearly displays the essay scoring results, but also provides users with personalized, targeted feedback and suggestions. Through intuitive score visualization, error analysis, and improvement suggestions, the system can help users better understand the quality of their essays and continuously improve their writing skills through targeted training.

[0080] The Chinese composition processing device includes: A data acquisition module is used to obtain the composition data submitted by students and generate digital text through optical character recognition technology; Specifically, the data acquisition module in this embodiment of the present invention primarily functions to capture student-submitted essay data and convert it into editable digital text using optical character recognition (OCR). This module includes an input interface and OCR technology, and connects to the text preprocessing module to provide raw text data for subsequent processing steps.

[0081] During implementation, students can submit their essays by uploading scanned copies, photos, or electronic documents. The data acquisition module first receives the student's essay image or scanned document and uses optical character recognition (OCR) technology to recognize the text within the image and convert it into digital text. OCR technology parses the characters in the image, identifying the individual characters and words and outputting them in a processable digital format. The recognized digital text provides the raw input data for subsequent steps such as text preprocessing and semantic analysis.

[0082] The module's OCR technology adapts to a variety of fonts and formats with high accuracy, ensuring clear text extraction from essays from diverse sources. Furthermore, OCR technology can filter out noise in images to improve text recognition accuracy. Through optical character recognition, the data acquisition module can automatically process large numbers of handwritten or printed essays submitted by students, significantly improving essay processing efficiency.

[0083] The text preprocessing module is connected to the data acquisition module and is used to divide the digital text into sentences, paragraphs and clear noise; Specifically, the text preprocessing module is a key component of the Chinese composition processing device, primarily responsible for further processing the digital text acquired by the data acquisition module. Its core functions include sentence and paragraph segmentation, as well as noise removal. The text preprocessing module is integrated with the data acquisition module to ensure that the digital text is optimized through the preprocessing process for subsequent semantic analysis, scoring, and generation of optimization suggestions.

[0084] First, the text preprocessing module receives the digital text from the data acquisition module. After being converted into a computer-processable character format using OCR technology, the digital text enters the text preprocessing module for structural processing. This module uses a sentence segmentation algorithm to segment the input text, breaking long text into several shorter sentence units. This process helps the system more accurately understand the grammatical structure and semantic relationships of each sentence.

[0085] During sentence segmentation, the system first uses punctuation marks (such as periods, question marks, and exclamation points) as segmentation points to divide the text into multiple sentences. For example, in the sentence "He ran to school," the system will divide it into complete sentence units based on the period. In some complex cases, the system will also combine grammatical rules to further determine sentence boundaries to ensure accurate sentence segmentation.

[0086] Next, the text preprocessing module performs paragraph segmentation. By detecting blank lines or paragraph markers within the text, the system accurately identifies and divides the text into paragraphs. This step ensures a clear overall structure for the essay, facilitating subsequent semantic analysis and scoring. The paragraph segmentation process provides useful information to the scoring module, for example, by checking whether the essay has appropriate paragraph structure, thereby improving scoring accuracy.

[0087] The text preprocessing module also handles noise cleaning. This refers to removing irrelevant information, redundant characters, or erroneous characters from the text. For example, OCR technology may produce misidentified characters during the recognition process. The text preprocessing module automatically detects and corrects these erroneous characters to ensure the accuracy and usability of the input text. Common noise includes extra spaces, incorrect characters, and repeated words, all of which can affect the quality of subsequent analysis and scoring. Therefore, the text preprocessing module uses regular expressions or other text cleaning algorithms to remove meaningless characters and redundant spaces.

[0088] Through standardized processing procedures, the text preprocessing module can efficiently and accurately provide optimized text for the composition scoring system, ensuring that the entire system can run stably and generate reasonable scores and feedback under different composition data inputs.

[0089] The semantic analysis module is connected to the text preprocessing module and is used to input the segmented text into the semantic encoding model to generate semantic vectors; Specifically, the semantic analysis module is responsible for converting preprocessed digital text into semantic vectors. This module primarily uses deep learning techniques, particularly Transformer-based models (such as BERT and GPT), to perform semantic analysis on text, capturing the deeper meaning of words and the logical relationships between sentences.

[0090] First, the semantic analysis module receives structured text data from the text preprocessing module. The system converts the words in each sentence into word vectors. These word vectors capture the semantic information of the words in context using a trained semantic encoding model. The model then processes the word vectors for each sentence to generate a semantic vector for that sentence. These semantic vectors not only reflect the meaning of individual words but also express the relationships between them.

[0091] By analyzing the text sentence by sentence, the system can further aggregate the semantic vectors of sentences into semantic representations of paragraphs or the entire text. These semantic representations provide deep semantic information about each part of the composition, helping to assess the overall coherence and logic of the composition.

[0092] Furthermore, the semantic analysis module evaluates the structure and fluency of essays by calculating the semantic similarity between sentences. By comparing the semantic vectors of different sentences, the system can identify logical connections or incoherence in the essay, providing a basis for scoring and feedback.

[0093] The scoring module is connected to the semantic analysis module and is used to comprehensively score the essay based on the scoring function; Specifically, the scoring module is responsible for comprehensively scoring the essays based on the semantic vectors and similarity information provided by the semantic analysis module. The scoring module uses multi-dimensional scoring criteria, including grammatical correctness, semantic coherence, sentence fluency, and vocabulary richness, to comprehensively evaluate the essays.

[0094] First, the scoring module receives the output data from the semantic analysis module, including the semantic vector for each sentence and the semantic similarity between sentences. Based on this data, the scoring module evaluates the composition's structure and logical coherence. A high level of semantic similarity between sentences indicates a coherent and coherent composition, resulting in a higher score; a low score indicates a low score.

[0095] Next, the scoring module assigns a comprehensive score to the essay based on the established scoring criteria. Each scoring dimension (such as grammar, semantic coherence, fluency, and vocabulary) is weighted, and the scoring module uses this weighted calculation to determine the essay's final score. Grammar is scored based on the results of a grammatical analysis, semantic coherence is assessed by the similarity between sentences, fluency considers sentence structure and paragraph cohesion, and lexical richness is measured by analyzing lexical diversity.

[0096] The optimization suggestion generation module is connected to the scoring module and is used to generate and output targeted optimization suggestions based on the scoring results; Specifically, the module connects the scoring module and the user interaction component to generate targeted optimization suggestions based on the scoring results and semantic analysis data.

[0097] First, a content quality model is established to identify the weak links in grammar, coherence, vocabulary, etc. of the composition based on the scores of each dimension and semantic vectors provided by the scoring module.

[0098] Then, a suggestion matching algorithm is used to extract low-scoring dimensions and locate the semantic vectors of the corresponding sentences or paragraphs. The algorithm then combines this with the suggestion database to select the rewriting suggestions that most closely match the semantic features.

[0099] The module includes a semantic comparison unit, a matching selection unit and a suggestion output unit. The semantic comparison unit is responsible for extracting problem areas from the scoring results, the matching selection unit calls the suggestion database and performs similarity matching, and the suggestion output unit feeds back the matching results to the user end.

[0100] Furthermore, each optimization suggestion includes problem location information, rewritten text, and improvement instructions to facilitate user understanding and adoption.

[0101] The output of recommendations is automatically triggered based on the preset threshold. If the score is lower than the set standard, the system will start the corresponding recommendation generation process.

[0102] The suggestion database is generated by expert corpus and automatic induction, and the content can be updated dynamically to ensure that common writing problems are covered.

[0103] In the overall process, the optimization suggestion generation module is started after the scoring is completed, automatically matching suggestions based on semantic content, and outputting the optimization information to the interactive interface.

[0104] A display module, connected to the optimization suggestion generation module, for displaying composition data and optimization suggestions; Specifically, the display module includes a score display unit, a dimension analysis unit, a content annotation unit, and a suggestion prompt unit. Each component communicates with the processing module through a logic control bus to achieve synchronous information output.

[0105] The display module receives the total score and dimension score data generated by the scoring module, and displays them in the form of graphics, tables or text according to the set template to intuitively reflect the performance of the composition.

[0106] The dimension analysis section displays scores for grammar, coherence, vocabulary, etc. in categories and supports expansion of sub-items to facilitate users' understanding of the performance of each dimension.

[0107] The content annotation department locates key sentences based on the results of semantic analysis, and highlights or marks them in the composition text, highlighting the problem sentences and the scoring dimensions to which they belong.

[0108] The suggestion prompt unit calls the output content of the optimization suggestion generation module, compares the suggested text with the original sentence and presents the result in the composition interface, with optimization instructions attached for users to refer to and modify.

[0109] The display module supports interactive operations. When users click on a rating item or annotated sentence, it can trigger the display of corresponding suggestions and explanations, thereby achieving traceability of the analysis content.

[0110] The display module is compatible with multi-terminal applications, supports local client or web-based deployment, and the interface layout adapts to different screen sizes to ensure display integrity and response speed.

[0111] Through structured display and semantic alignment annotation, the display module can clearly express the results of composition quality evaluation, improving users' understanding efficiency and modification experience.

[0112] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. The method for processing Chinese composition is characterized by: The following steps are involved: S1. Obtain student-submitted essay data and generate digital text from paper and image files using optical character recognition technology; S2. Preprocessing of digital text, including paragraph division, sentence group decomposition, noise character removal, punctuation standardization, and text sentence segmentation; S3. Input the sentence group into the semantic encoding model, obtain the corresponding semantic vector, and construct a semantic vector field composed of multiple semantic vectors; S4. Calculate the semantic relevance between semantic units within the composition based on the semantic vector field; S5. comprehensively scoring the composition text using a scoring function of multiple scoring dimensions; S6. Generate composition optimization suggestions based on the optimization goal of the scoring function.

2. The method for processing Chinese composition according to claim 1, characterized in that: The optical character recognition technology in step S1 is a convolutional neural network or a long short-term memory network model based on a deep learning algorithm.

3. The method for processing Chinese composition according to claim 1, characterized in that: The pre-processing in step S2 is to clean up the noise of the composition text, including removing non-text symbols, redundant spaces, and meaningless line breaks in the text, and performing paragraph segmentation to ensure that the sentence group structure within the paragraph is correct.

4. The method for processing Chinese composition according to claim 1, characterized in that: The semantic encoding model in step S3 adopts a pre-trained Transformer model, which can generate high-dimensional semantic vectors based on the context, and one sentence group corresponds to one semantic vector.

5. The method for processing Chinese composition according to claim 1, characterized in that: The semantic vector field in step S3 includes multiple semantic vectors, which have similarity in the semantic space and a dimension of the semantic vector is d, where d is a characteristic dimension of the semantic vector.

6. The method for processing Chinese composition according to claim 1, characterized in that: The method for calculating the semantic relevance in step S4 is: calculating the similarity between the semantic vectors of each pair of sentence groups, and the similarity is measured by cosine similarity, and the formula is: ； in: Indicates the The semantic vector of a sentence; Indicates the The semantic vector of a sentence; Represents a vector and The dot product of Represents a vector Length of the module; Represents a vector Length of the module; sim express and The cosine similarity of .

7. The method for processing Chinese composition according to claim 1, characterized in that: The scoring function in step S5 is a weighted function that covers multiple scoring dimensions, including: Semantic coherence score, which measures the semantic consistency between sentence groups; Logical scoring, assessing the logical relationship between the argument and evidence in the essay; The language diversity score measures the richness of vocabulary and sentence structure in the essay; Grammar standardization scoring, checking grammatical errors and writing standardization in the composition.

8. The method for processing Chinese composition according to claim 1, characterized in that: The optimization suggestion in step S6 is directed optimization based on the various scoring indicators of the scoring function, including the following: Regarding semantic coherence, it is recommended to modify the unnatural transitions between sentence groups; Regarding logic, it is recommended to strengthen the argument support and reasoning process in the argument part; In view of language diversity, it is recommended to use diverse vocabulary or sentence structures; Regarding grammatical standardization, specific grammatical errors and suggested modifications are listed.

9. The method for processing Chinese composition according to claim 1, characterized in that: The generation process of the optimization suggestions in step S6 is customized according to the user-specified goals, including: style optimization, refinement of language expression, and improvement of argument structure.

10. A Chinese composition processing device, according to the Chinese composition processing method according to any one of claims 1 to 9, characterized in that: include: A data acquisition module is used to obtain the composition data submitted by students and generate digital text through optical character recognition technology; The text preprocessing module is connected to the data acquisition module and is used to divide the digital text into sentences, paragraphs and clear noise; The semantic analysis module is connected to the text preprocessing module and is used to input the segmented text into the semantic encoding model to generate semantic vectors; The scoring module is connected to the semantic analysis module and is used to comprehensively score the essay based on the scoring function; The optimization suggestion generation module is connected to the scoring module and is used to generate and output targeted optimization suggestions based on the scoring results; The display module is connected to the optimization suggestion generation module and is used to display the composition data and optimization suggestions.

Images