Context-based chinese intelligent correction and text optimization system

Abstract

The application discloses a context-based Chinese intelligent correction and text optimization system and belongs to the technical field of natural language processing.The system comprises a context semantic coding module, a multi-dimensional error detection module, a style and genre perception module, a language fluency evaluation module and an adaptive correction optimization module.The context semantic coding module realizes multi-granularity semantic coding through a hierarchical attention mechanism.The multi-dimensional error detection module detects homonym errors, homograph errors, idiom misuse and syntax errors in parallel.The style and genre perception module identifies the style type of the text and provides targeted optimization.The language fluency evaluation module evaluates language fluency and recommends alternative expressions.The adaptive correction optimization module generates a correction result by comprehensively integrating the outputs of the modules and realizes closed-loop parameter optimization based on user feedback.The correction accuracy of the application reaches 96%, and the quality of Chinese writing is effectively improved.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, specifically to a context-based intelligent Chinese error correction and text optimization system. Background Technology

[0002] With the rapid development of information technology and the widespread adoption of digital office work, the demand for creating and editing Chinese texts has exploded. Whether writing academic papers, drafting business documents, editing news articles, or conducting daily written communication, ensuring the accuracy and standardization of text content is crucial for effectively conveying information. However, as a language system with unique characteristics, Chinese text correction faces numerous technical challenges and practical difficulties.

[0003] From the perspective of the inherent characteristics of the Chinese language, there are a large number of homophones and similar-looking characters, which can easily lead to confusion and misselection during input. Taking Pinyin input methods as an example, when a user enters a Pinyin sequence, the input method usually provides multiple candidate words for the user to choose from. These candidates often include several homophones, and it is easy to make a mistake if one is not careful. The situation with similar-looking characters is equally serious. The construction of Chinese characters is based on radicals and strokes, and many Chinese characters differ only in a few strokes or radicals, which can easily cause visual confusion during rapid writing or reading. In addition, idioms, as an important part of Chinese and a unique cultural carrier, require accurate understanding of their semantic connotations and applicable contexts. Misuse of idioms not only affects the expressive effect of the text but may also cause semantic ambiguity or even make jokes. The flexibility of Chinese grammatical structure also increases the difficulty of detecting grammatical errors. Grammatical problems such as subject-verb agreement, incomplete components, and incorrect word order are common in actual writing.

[0004] From the current state of technological development, Chinese text correction technology mainly falls into two major categories: rule-based methods and statistical learning-based methods. Rule-based methods rely on manually constructed error correction rule bases, which are typically summarized and categorized by linguistics experts based on Chinese grammatical norms and common error patterns. This method boasts high accuracy in specific scenarios, with clear and easily interpretable rules. However, the coverage of these rules has inherent limitations, making it difficult to address the diversity and complexity of language use, and the maintenance and updating of these rules require significant manual input. Statistical learning-based methods identify potential errors by analyzing language patterns and word collocation rules in large-scale corpora. This method can automatically learn language rules from data, offering broader coverage. However, traditional statistical methods often rely on fixed feature templates, ignoring rich contextual information, resulting in poor performance when handling context-related errors.

[0005] In recent years, the rapid development of deep learning technology has brought new technological opportunities and solutions to the field of text correction. The emergence of pre-trained language models enables computers to better understand the semantic information and contextual relationships of text, providing a powerful feature representation foundation for text correction tasks. In existing technologies, such as Chinese Patent Publication No. CN118446221A, a deep learning-based intelligent writing assistance system and method are disclosed. This system collects different types of text data for preprocessing, selects and trains a deep learning model, acquires the current writing content during user writing for semantic analysis and key information extraction, uses the trained deep learning model to generate text, and selects the generated text with the highest similarity as the output. This technical solution mainly focuses on text generation and supplementary auxiliary content, but it has significant shortcomings in terms of the accuracy and context sensitivity of text correction.

[0006] Specifically, the existing technology has the following technical problems: First, its text processing flow focuses on generating supplementary content rather than accurately correcting errors in existing text, which cannot meet users' actual needs for original text correction; Second, the method based on stemming and similarity matching is difficult to accurately identify and correct character-level errors such as homophones and similar-looking characters, because these errors often do not affect the overall similarity of the text but will seriously affect the accuracy of the text; Third, it lacks the ability to perceive different writing styles and cannot provide targeted optimization suggestions and normative constraints according to different text types such as academic papers, business documents, and literary creations; Fourth, it does not consider the assessment of language authenticity, making it difficult to help non-native speakers or users with limited Chinese writing skills improve the naturalness and fluency of their expression.

[0007] In summary, existing text processing solutions still have significant technical shortcomings in areas such as context-sensitive error correction, multi-type error detection, stylistic awareness, and language authenticity assessment, making it difficult to meet users' urgent needs for high-quality Chinese writing assistance. Therefore, there is an urgent need for a Chinese intelligent error correction and text optimization system that can comprehensively utilize contextual semantic information, achieve accurate multi-dimensional error detection, perceive stylistic features, and assess language authenticity. Summary of the Invention

[0008] To address the shortcomings of existing technologies, this invention provides a context-based intelligent Chinese error correction and text optimization system. By constructing a context-sensitive error correction engine based on a pre-trained language model, it achieves accurate error correction and optimization of Chinese text.

[0009] The technical solution of the present invention is as follows:

[0010] The context-based Chinese intelligent error correction and text optimization system includes a context semantic encoding module, a multi-dimensional error detection module, a style perception module, a language authenticity assessment module, and an adaptive error correction and optimization module.

[0011] The context semantic encoding module receives the Chinese text to be processed and performs context-sensitive semantic encoding on each character in the Chinese text based on a pre-trained language model. It captures the context dependencies at the character, word, and sentence levels through a hierarchical attention mechanism, and generates a context semantic vector containing multi-granular semantic information.

[0012] The multi-dimensional error detection module receives the context semantic vector and processes it in parallel through homophone error detection unit, similar-looking character error detection unit, idiom misuse detection unit, and grammatical error detection unit. Based on the error confidence output by each detection unit, the error type and error location are determined, and the error detection result is generated.

[0013] The style perception module receives the context semantic vector, determines the style type of the Chinese text to be processed through the style classifier, extracts the corresponding style norm features from the preset style norm knowledge base based on the determined style type, and fuses the style norm features with the context semantic vector to generate style type labels and style feature vectors.

[0014] The language authenticity assessment module receives contextual semantic vectors and calculates the authenticity score of each expression unit in the Chinese text to be processed through a multilayer perceptron-based authenticity assessment network. For expression units with authenticity scores lower than a preset authenticity threshold, alternative expressions with similar semantics and authenticity scores higher than the preset authenticity threshold are retrieved from the authentic expression corpus, and authenticity scores and alternative expression suggestions are generated.

[0015] The adaptive error correction and optimization module receives error detection results, genre labels, style feature vectors, authenticity scores, and alternative expression suggestions. Based on the genre norm constraints corresponding to the genre labels, it filters and sorts the error detection results. The genre norm constraints include word usage constraints, sentence structure constraints, and punctuation usage constraints. Combining the authenticity scores and alternative expression suggestions, it generates the error correction result text and converts the user feedback information during the error correction process into feedback optimization parameters, which are sent to the context semantic encoding module and the multi-dimensional error detection module to update the model parameters.

[0016] The beneficial effects of this invention are:

[0017] First, this invention achieves multi-granularity contextual semantic encoding through a hierarchical attention mechanism, which can simultaneously capture semantic dependencies at the character, word, and sentence levels. Compared with traditional similarity-based matching methods, it can more accurately understand text semantics and contextual relationships, with an error correction accuracy of up to 96%.

[0018] Secondly, this invention achieves parallel detection of homophone errors, similar-looking character errors, misuse of idioms, and grammatical errors by setting up a multi-dimensional error detection module, covering the main error types in Chinese text, and significantly improving the comprehensiveness and accuracy of error detection.

[0019] Third, the present invention identifies the style type of a text through a style perception module and provides optimization suggestions that conform to specific style norms, enabling the system to provide differentiated error correction services based on different scenarios such as academic papers, business documents, or literary creations.

[0020] Fourth, this invention uses a language authenticity assessment module to detect common unnatural expressions among non-native speakers, providing alternative expressions that are more in line with Chinese habits, thus effectively improving the language quality of the text.

[0021] Fifth, this invention achieves closed-loop updates of model parameters based on user feedback through an adaptive error correction optimization module, enabling the system to continuously learn and optimize, and its error correction performance continuously improves with use. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the overall architecture of the system of the present invention.

[0023] Figure 2 This is a structural diagram of the context semantic encoding module.

[0024] Figure 3 This is a structural diagram of a multi-dimensional error detection module.

[0025] Figure 4 This is a structural diagram of the style perception module.

[0026] Figure 5 This is a schematic diagram of the language authenticity assessment module.

[0027] Figure 6 This is a schematic diagram of the adaptive error correction and optimization module. Detailed Implementation

[0028] Please refer to the attached document. Figures 1-6 The present invention will now be described in detail with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0029] This invention provides a context-based intelligent Chinese error correction and text optimization system, such as... Figure 1As shown, the system comprises a context semantic encoding module 1, a multi-dimensional error detection module 2, a stylistic perception module 3, a language authenticity assessment module 4, and an adaptive error correction and optimization module 5. These five core modules form a deeply coupled closed-loop collaborative system architecture. The output data of the previous module serves as the key input of the next module, and the feedback optimization parameters generated by the adaptive error correction and optimization module 5 are passed back to the context semantic encoding module 1 and the multi-dimensional error detection module 2, enabling dynamic optimization and continuous learning of system parameters.

[0030] like Figure 2 As shown, the context semantic encoding module 1 receives the Chinese text to be processed and performs context-sensitive semantic encoding on each character in the Chinese text based on a pre-trained language model, generating a context semantic vector containing multi-granular semantic information. In one embodiment of the present invention, the design of this module fully considers the character-level characteristics and multi-level semantic structure of Chinese text, and adopts a hierarchical processing architecture to realize the step-by-step extraction and fusion of semantic information.

[0031] The context semantic encoding module 1 mainly consists of four components: text preprocessing unit, character embedding unit, position encoding unit, and hierarchical attention encoding unit. These units form a sequential data processing flow and work together to complete the task of converting raw text into semantic vectors.

[0032] The text preprocessing unit is the first processing step in the entire encoding process, responsible for standardizing the input raw Chinese text. The processing operations performed by this unit include unified conversion of full-width and half-width characters, filtering and removing abnormal characters and garbled text, merging and compressing consecutive whitespace characters, and intelligent segmentation of excessively long texts. In this embodiment, full-width and half-width conversion uses a Unicode mapping table for character replacement, converting full-width characters to their corresponding half-width forms to ensure consistency in subsequent processing. Abnormal character filtering employs a whitelist mechanism, retaining only Chinese characters, common punctuation marks, numbers, and English letters, while deleting other non-standard characters. Text segmentation uses a sliding window method based on sentence boundaries, segmenting text into complete sentences to avoid truncating sentences and causing semantic fragmentation.

[0033] The character embedding unit is responsible for converting the preprocessed Chinese character sequence into a dense vector representation. Internally, this unit maintains a character embedding matrix, where each row corresponds to a vector representation of a Chinese character. In this implementation, the embedding matrix is ​​initialized using the character embedding parameters of a pre-trained language model, and can be fine-tuned and optimized through feedback learning. The semantic vector has a dimension of 768 to 1024, preferably 768 or 1024. This embodiment uses 768 dimensions to achieve a better balance between representational power and computational efficiency. For the input character sequence, the character embedding unit maps each character to its corresponding embedding vector through a lookup table, forming a character embedding matrix as the basic input for subsequent processing.

[0034] The positional encoding unit adds positional information to each character, enabling the model to perceive the relative and absolute positional relationships of characters within a text sequence. This invention employs a sine-cosine positional encoding method, which exhibits good extrapolation properties and can handle input sequences of arbitrary length. Specifically, for a character at position pos, the i-th dimension of its positional encoding vector is calculated as follows:

[0035] When i is even: ,

[0036] When i is odd: ,

[0037] in, Let i be the i-th dimension of the position encoding vector. The index of the character's position in the sequence. In this embodiment, the dimension is vector. The value is 768. The position encoding vector and the character embedding vector are added element-wise to obtain the character representation vector containing position information.

[0038] The hierarchical attention encoding unit is the core component of the context semantic encoding module 1. It captures multi-granular semantic information through the hierarchical attention context encoding algorithm proposed in this invention. The core idea of ​​this algorithm is to calculate attention weights at three levels: character, word, and sentence, and then obtain a comprehensive context semantic vector through weighted fusion. Compared with a single-level encoding method, this multi-level encoding method can capture the semantic information and structural features in the text more comprehensively.

[0039] The specific implementation process of the hierarchical attention context encoding algorithm proposed in this invention is as follows. For the input character representation sequence, the attention representation is first calculated at the character level. The character-level attention calculation adopts a multi-head self-attention mechanism, and its calculation process can be expressed as follows:

[0040] ,

[0041] in, This is a character-level attention representation matrix, with dimensions equal to the sequence length multiplied by the hidden dimension. For character-level query matrices, It is a key matrix at the character level. A matrix of character-level values. This represents the multi-head attention computation function, which receives a query matrix. Key matrix Sum matrix The input vectors are weighted and represented by an attention-weighted representation. Multi-head attention computation first projects the input vectors into multiple subspaces through different linear transformations, independently computes the attention in each subspace, and finally concatenates the outputs of each head and performs a linear transformation to obtain the final result. Preferably, the number of attention heads is set to 12, and each attention head has a 64-dimensional dimension.

[0042] After obtaining the character-level representations, the character sequences need to be aggregated into word-level representations. This invention employs a word boundary detection mechanism to identify the start and end positions of words. This mechanism determines word boundaries based on the output of a pre-trained word segmentation model. For each identified word, its contained character-level representations are subjected to average pooling to obtain the initial representation vector of the word. Then, attention representations are calculated at the word level:

[0043] ,

[0044] in, The word-level attention representation matrix is ​​as follows: , , These are the query matrix, key matrix, and value matrix at the word level, respectively. Word-level attention can capture the dependencies and collocation patterns between words, playing an important role in detecting errors such as improper word collocation and misuse of idioms.

[0045] Similarly, sentence boundary detection divides the text into sentence units. Sentence boundary identification is based on a comprehensive judgment of punctuation and grammatical rules, using terminating punctuation marks such as periods, question marks, and exclamation marks as sentence delimiters. For each sentence, its word-level representations are pooled to obtain the initial sentence representation, and then the sentence-level attention representation is calculated.

[0046] ,

[0047] in, The sentence-level attention representation matrix, , , These are the query matrix, key matrix, and value matrix at the sentence level, respectively. Sentence-level attention can capture the logical relationships and discourse structure between sentences, which is of great value for understanding the overall semantics of text and detecting cross-sentence errors.

[0048] After obtaining the three levels of attention representations, they need to be fused into the final context semantic vector. This invention employs a hierarchical fusion mechanism, which uses learnable weight coefficients to perform a weighted sum of the three levels of representations:

[0049] ,

[0050] in, This is the final context semantic vector matrix. , , The weighting coefficients for hierarchical fusion satisfy the constraints. These weights are optimized through gradient descent during model training, or they can be manually set according to specific task requirements. Preferably, The value range is from 0.4 to 0.5. The value range is from 0.3 to 0.4. The value range is from 0.1 to 0.3. In this embodiment, Set to 0.45, Set to 0.35. Setting the parameter to 0.20 has been verified through extensive experiments, demonstrating optimal performance in overall error correction. The highest weight is given to character-level errors because they constitute the largest proportion of errors in Chinese error correction tasks, requiring a focus on character-level information.

[0051] Regarding the selection of the context window length, this invention fully considers the paragraph structure characteristics of Chinese text and practical application requirements. The context window length determines the range of context that the model can refer to when encoding a certain character. A window that is too short will result in insufficient context information, while a window that is too long will increase computational overhead and may introduce noise. Preferably, the context window length ranges from 64 to 512 characters, and this embodiment uses 256 characters as the default context window length. This value can cover most sentences and common paragraph structures in Chinese, while avoiding the problem of decreased computational efficiency caused by an excessively long window. The hierarchical attention mechanism has 8 to 16 attention heads, preferably 12. Each attention head is responsible for capturing the semantic relationships of different subspaces, and the parallel computation of the multi-head mechanism improves the expressive power of the model.

[0052] The context semantic vector output by the context semantic encoding module 1 is simultaneously passed to the multi-dimensional error detection module 2, the style perception module 3, and the language authenticity assessment module 4, serving as the basic input data for analysis and processing by these three modules. This one-to-many data transmission architecture enables the system to execute multiple analysis tasks in parallel, significantly improving overall processing efficiency and response speed.

[0053] like Figure 3 As shown, the multi-dimensional error detection module 2 receives the context semantic vector output by the context semantic encoding module 1, performs parallel detection processing on it for multiple types of errors, and generates an error detection result containing error type and error location information. In one embodiment of the present invention, this module designs specialized detection units for the four most common types of errors in Chinese text, including a homophone error detection unit, a similar-looking character error detection unit, an idiom misuse detection unit, and a grammatical error detection unit. The four detection units work together to achieve comprehensive coverage and accurate identification of errors in Chinese text.

[0054] The multi-dimensional error detection module 2 adopts a parallel processing architecture. Four detection units simultaneously receive context semantic vectors and perform independent analysis. Each unit outputs its own error confidence score, and finally, the error synthesis and judgment unit aggregates the detection results from all units to form a unified output. This parallel architecture significantly improves detection efficiency compared to serial processing, while the independence of each unit facilitates subsequent module-level upgrades and maintenance.

[0055] The homophone error detection unit is responsible for detecting homophone substitution errors caused by improper use of the Pinyin input method. These errors are extremely common in daily Chinese input. When using the Pinyin input method, users may inadvertently select the wrong character due to the presence of multiple homophones in the candidate list. This unit first constructs a homophone confusion matrix, which records the confusion probabilities between common homophone pairs. These probabilities are obtained based on large-scale user input log statistics and can reflect homophone confusion patterns in actual use.

[0056] The specific implementation process of the homophone error detection algorithm proposed in this invention is as follows. For each character in the text... First, through the Pinyin conversion function Obtain its Pinyin encoding ,in The function takes a Chinese character as input and returns the corresponding Pinyin string. Then, it retrieves a set of homophones based on the Pinyin encoding. This set contains all those related to Other Chinese characters with the same pinyin. For each candidate character in the set of homophones. The context fit is calculated, that is, whether replacing the current character with the candidate character would be semantically more reasonable. The context fit is calculated using the following formula:

[0057] ,

[0058] in, Candidate characters In the current context The fit probability in the value ranges from 0 to 1. Candidate characters The embedding vector is obtained by looking up a table in the character embedding matrix. The context semantic vector for the current position is output by context semantic encoding module 1. For characters A collection of homophones, superscript This represents the vector transpose operation. This represents an exponential function with the natural constant e as its base. The formula uses a softmax normalized exponential function to normalize the context matching score of each candidate character. The softmax function is calculated by taking the natural exponent of each element. Then divide by the sum of the natural indices of all elements to obtain the fitness representation in the form of a probability distribution.

[0059] When there is a candidate homophone The context adaptation probability is significantly higher than that of the original character. When determining the adaptation probability, it is determined that there may be a homophone error at the current position. The specific rules for error determination are as follows:

[0060] ,

[0061] in, For characters The function flags homophone errors. A result of 1 indicates that a homophone error was detected, while a result of 0 indicates that no error was detected. This is an indicator function, defined as outputting 1 when the condition within the parentheses is true, and 0 when the condition is false. The threshold for identifying homophone errors is a preset hyperparameter. Preferably, The threshold value ranges from 0.75 to 0.90, and this embodiment uses 0.85 as the default threshold. This higher threshold is set to control the false positive rate and prevent correct characters from being incorrectly marked as homophones. The specific value of the threshold can be adjusted according to the precision and recall requirements of the application scenario. If recall is prioritized, the threshold can be appropriately lowered; if precision is prioritized, the threshold can be appropriately increased.

[0062] The similar-looking character error detection unit is responsible for detecting visual confusion errors caused by the similarity of character shapes. These errors typically occur in handwriting input or image text recognition scenarios, where some Chinese characters are extremely similar in appearance, differing only slightly in individual strokes or radicals, easily leading to confusion during recognition or writing. This unit first constructs a feature map of similar-looking characters, using the stroke structure information of the Chinese characters to perform structured encoding representation of each character.

[0063] The proposed algorithm for detecting similar-looking characters in this invention is based on a joint judgment mechanism of character shape features and contextual semantics. For each character in the text... First, extract its stroke structure feature vector. This feature vector encodes multiple structural attributes of a character, including radical category, total number of strokes, stroke order encoding, structural type (left-right structure, top-bottom structure, enclosing structure, etc.), and the relative positional relationships of each component. The feature vector is set to 128 dimensions, effectively representing the structural features of Chinese characters.

[0064] Based on stroke structure feature vectors, the similarity between characters is calculated, and a set of similar-looking characters is retrieved. ,in The cosine similarity function is calculated as cos(a,b)=a·b / (||a||·||b||), which is the inner product of two vectors divided by the product of their respective magnitudes. The character similarity threshold is set to 0.70. The set of similar characters includes all other characters whose character similarity to the current character exceeds the threshold.

[0065] For each candidate character in the set of similar-looking characters The probability of confusion between similar-looking characters is calculated by combining the contextual semantic vector and the character shape feature vector:

[0066] ,

[0067] in, Candidate characters The confusion probability is based on context and glyph features, and its value ranges from 0 to 1. is the sigmoid activation function, used to map any real value to a probability space between 0 and 1. It is a learnable weight matrix with a dimension of 1 multiplied by (768+128+128). This represents the concatenation operation between the context vector and two glyph feature vectors. This is a bias term. When the confusion probability exceeds a preset threshold, a similar-looking character error is determined to exist. Preferably, the confidence threshold for detecting similar-looking characters ranges from 0.70 to 0.85; in this embodiment, 0.78 is used as the default value.

[0068] The idiom misuse detection unit is responsible for detecting integrity errors and semantic mismatch errors in the use of idioms. Idioms are an important part of the Chinese language, with fixed structural forms and specific semantic connotations. Misuse of idioms not only affects the expressive effect of the text but may also cause semantic ambiguity or misunderstanding. This unit consists of two parts: an idiom integrity detection subunit and an idiom semantic misuse detection subunit, which are responsible for detecting different types of idiom misuse problems, respectively.

[0069] The idiom integrity detection subunit scans and matches text using a pre-built idiom dictionary to detect integrity issues such as missing, replaced, or reversed characters in idioms. This subunit uses a prefix tree data structure to store the idiom dictionary, supporting efficient string matching operations. During the scanning process, for character sequences suspected of being idioms, it checks whether they completely match a specific idiom entry in the dictionary. If character differences are found, they are marked as potential idiom integrity errors. The idiom dictionary contains approximately 50,000 commonly used idioms, and each idiom entry records information such as the standard spelling, pinyin, definition, and usage instructions.

[0070] The idiom semantic fit detection subunit evaluates the appropriateness of idiom usage in the current context by calculating semantic fit metrics. Even if the idiom itself is written perfectly, its use in an inappropriate context is still considered misuse. The formula for calculating idiom semantic fit proposed in this invention is as follows:

[0071] ,

[0072] in, For idioms In context semantic fit in The cosine similarity function is used. This is the semantic vector of the idiom, obtained by average pooling the context semantic vectors of each character in the idiom. The average pooling result of the context semantic vector represents the overall context semantics. For the mutual information between idioms and their context This reflects the degree of co-occurrence of the idiom with the current context. When the semantic fit is lower than a preset threshold, it is determined that there is a possibility of idiom misuse. Preferably, the confidence threshold for idiom misuse detection ranges from 0.80 to 0.95. In this embodiment, 0.88 is used as the default value. This threshold is set relatively high to control the false alarm rate.

[0073] The grammatical error detection unit is responsible for detecting errors at the sentence structure and grammatical rule levels, including but not limited to subject-verb agreement, verb-object agreement, incomplete or redundant components, incorrect word order, and mixed sentence structures. This unit uses a deep learning model based on sequence labeling to analyze the text and identify the specific location and type of grammatical errors.

[0074] The sequence labeling model employs a bidirectional long short-term memory network as a feature extractor, further extracting sequence-level features based on the context semantic vector. Then, a conditional random field layer is used to jointly decode the labels at each position. The labeling system adopts the BIOE labeling scheme, where B represents the start of an error, I represents an internal position of an error, O represents a correct position, and E represents the end of an error. For each position, the model outputs the probability distribution of its corresponding label, and determines whether a syntax error exists based on this probability distribution.

[0075] Preferably, the error confidence threshold for syntax error detection ranges from 0.65 to 0.80, and this embodiment uses 0.72 as the default value. This threshold is set relatively low because the boundaries for judging syntax errors are somewhat ambiguous, requiring a high recall rate to avoid missing genuine syntax errors. At the same time, a lower threshold may introduce some false positives, but these false positives can be filtered and selected by the subsequent adaptive error correction optimization module.

[0076] The error comprehensive judgment unit receives the output results from the four detection units mentioned above and performs summary and integration processing. The output of this unit is a structured error detection result, which includes three components: error type label (homophone error, similar-looking character error, idiom misuse, grammatical error), error location information (start character index and end character index), and error confidence (a floating-point number ranging from 0 to 1). The error detection result is passed to the adaptive error correction optimization module 5 for subsequent error correction candidate generation and filtering sorting processing.

[0077] like Figure 4 As shown, the style perception module 3 receives the context semantic vector output by the context semantic encoding module 1, identifies the style type of the text, and generates the corresponding style feature vector. In one embodiment of the present invention, this module classifies the input text using a style classifier and extracts corresponding normative feature information from a preset style norm knowledge base, providing style constraint basis for subsequent error correction and optimization.

[0078] The style perception module 3 mainly consists of three parts: style classification unit, style norm knowledge base, and style feature fusion unit. These parts work together to achieve a complete processing flow from style recognition to feature fusion.

[0079] The text classification unit uses a multilayer perceptron (MLP) structure to process the context semantic vector and output the text type label. The MLP consists of an input layer, two hidden layers, and an output layer. The number of neurons in the hidden layers is set to 512 and 256, respectively, and the ReLU activation function is used. The number of neurons in the output layer is the same as the number of text types, and the probability distribution of each text type is output using a softmax function.

[0080] This invention defines five writing styles: academic papers, business documents, literary works, news reports, and everyday communication, covering common Chinese writing scenarios and application needs. Academic papers are characterized by rigorous and standardized vocabulary, complete sentence structures, sound logical argumentation, and professional terminology. Business documents are characterized by concise and clear expression, formal and polite tone, standardized format, and efficient information delivery. Literary works are characterized by beautiful and vivid language, rich rhetorical devices, delicate emotional expression, and a distinctive style. News reports are characterized by objectivity, accuracy, an inverted pyramid structure, concise and refined language, and complete information elements. Everyday communication is characterized by natural and casual expression, a high degree of colloquialism, clear emotional expression, and relaxed format requirements.

[0081] The specific implementation process of the style feature weighted fusion algorithm proposed in this invention is as follows. First, the style classification unit calculates the probability distribution of the text belonging to each style type:

[0082] ,

[0083] in, The text belongs to a genre type. The probability of is between 0 and 1, and the sum of the probabilities of all types is 1. This is a classification weight matrix with dimensions of 5 x 768. This is the pooling operation function, which aggregates variable-length sequence representations into fixed-length vectors. In this embodiment, average pooling is used, which means taking the arithmetic mean of the vectors at all positions in the sequence. For context semantic vector matrix, This is the bias vector. The genre label is used to select the category with the highest probability through the argmax operation, where the argmax operation returns the value of the independent variable that maximizes the objective function. .

[0084] The stylistic norms knowledge base stores linguistic normative feature information corresponding to each stylistic type. This information includes vocabulary norms (a list of recommended and avoided words for this stylistic type), sentence structure characteristics (typical sentence patterns and length distribution for this stylistic type), punctuation usage rules (frequency and usage scenarios of each punctuation mark), and other stylistic-specific normative requirements. The knowledge base uses a structured storage format, with each stylistic type corresponding to a normative feature vector. The vector dimension is 256.

[0085] The style feature fusion unit fuses the contextual semantic vector with the stylistic normative feature vector to generate a comprehensive style feature vector:

[0086] ,

[0087] in, This is a style feature vector with 256 dimensions. This is a layer normalization function. It calculates the mean and variance of the input vector along the feature dimensions and then standardizes it. This standardization is used to stabilize the training process and accelerate convergence. To fuse the weight matrix, This represents the concatenation operation between the context semantic vector and the stylistic feature vector. This is the bias vector.

[0088] Style matching score measures the degree to which the input text fits the determined style type, reflecting the text's stylistic conformity. The calculation method is as follows:

[0089] ,

[0090] in, This represents the stylistic match score, ranging from -1 to 1. A higher value indicates a higher degree of match. Let be the cosine similarity function. When... When the text matching score falls below the preset threshold, the system will alert the user that the text style deviates from the corresponding genre, and will strengthen genre standard constraints during subsequent error correction. Preferably, the genre matching score threshold ranges from 0.70 to 0.90; this embodiment uses 0.80 as the default value.

[0091] The style perception module 3 outputs style type labels and style feature vectors, which are then passed to the adaptive error correction and optimization module 5 to constrain style norms during the error correction candidate selection process, ensuring that the error correction results meet the language norm requirements of the corresponding style.

[0092] like Figure 5As shown, the language authenticity assessment module 4 receives the context semantic vector output by the context semantic encoding module 1, assesses the language authenticity of the text expression, and provides alternative suggestions for expressions that are not authentic enough. In one embodiment of the present invention, this module specifically targets and provides optimization suggestions for common unnatural expressions encountered by non-native speakers or users with limited Chinese writing skills, helping users improve the language quality and fluency of their texts.

[0093] The Language Authenticity Assessment Module 4 mainly consists of four parts: expression unit segmentation unit, authenticity feature extraction unit, authenticity score calculation unit, and alternative expression retrieval unit, forming a complete processing flow from segmentation to assessment to suggestions.

[0094] The expression unit segmentation unit is responsible for segmenting the input Chinese text into phrase-level expression units, each of which is the basic unit for evaluating authenticity. The segmentation strategy is based on a comprehensive consideration of grammatical structure and semantic integrity, striving to ensure that each expression unit forms a relatively complete semantic fragment. Preferably, each expression unit contains 3 to 8 characters, a length setting that can cover common Chinese phrases and fixed collocations. The segmentation process employs a sequence labeling model based on conditional random fields, labeling each character in the text as the start, interior, or end position of an expression unit, and then performing segmentation based on the labeling results.

[0095] The authenticity feature extraction unit extracts three types of features from each expression unit: lexical richness, collocation rationality, and conformity to expression habits. These three types of features reflect the authenticity of the expression from different perspectives, and comprehensive consideration can yield a more complete and accurate authenticity assessment result.

[0096] Lexical richness reflects the diversity and precision of vocabulary use in expression. Authentic Chinese expressions typically use appropriate and precise vocabulary, avoiding simplistic or inappropriate vocabulary. The formula for calculating lexical richness is as follows:

[0097] ,

[0098] in, For expression unit The vocabulary richness feature value, This refers to the number of unique words in an expression unit, counted by deduplication. The total number of words in the unit of expression. This formula represents the frequency of low-frequency words (words that appear below a certain threshold in a large-scale corpus) within a given unit. The first part of the formula reflects the diversity of the vocabulary, while the second part reflects the precision and specialization of the vocabulary.

[0099] Collocational appropriateness features reflect the degree of reasonableness in word collocation within an expression. Chinese contains numerous fixed collocations and idiomatic expressions; authentic expressions should adhere to these collocational rules. Collocational appropriateness features are calculated based on statistical information from word collocations in a large-scale corpus.

[0100] ,

[0101] in, For expression unit The feature value of the reasonableness of the combination, This is the set of word collocations within an expression unit, containing all adjacent word pairs within that unit. For words and The point-to-point mutual information reflects the collocation strength of two words. The calculation of point-to-point mutual information is based on the co-occurrence frequency and marginal frequency of words in the corpus. A higher collocation rationality feature value indicates that the word collocation in the expression conforms more closely to the habitual usage of Chinese.

[0102] The expression habit conformity feature reflects the degree of similarity between the expression and the expression patterns commonly used by native speakers. This feature is obtained through comparison with a corpus of authentic expressions:

[0103] ,

[0104] in, For expression unit The feature value of the degree of conformity of expression habits This is a corpus of authentic expressions, containing a large number of high-quality expression fragments extracted from native speaker writings. and Respectively, they are expression units and representation units in the corpus The semantic vector is calculated by the context semantic encoding module 1. is the cosine similarity function. This feature measures the distance between the current expression and the most similar idiomatic expression in the corpus; a higher value indicates a closer resemblance to an idiomatic expression pattern.

[0105] The authenticity score calculation unit calculates the final authenticity score by combining the above three types of characteristics:

[0106] ,

[0107] in, For expression unit The authenticity score ranges from 0 to 1, with higher values ​​indicating more authentic expression. , , The weight coefficients for the three types of features satisfy the constraints. These weighting coefficients determine the relative importance of various features in the authenticity assessment and can be adjusted according to the specific application scenario. Preferably, The value is 0.25. The value is 0.35. The value is 0.40. This weighting configuration emphasizes the importance of conformity to expression habits in the authenticity assessment, because similarity to native speaker expression patterns is the most direct indicator of authenticity.

[0108] Alternative expression retrieval units are used for expressions with an authenticity score below a preset authenticity threshold. The system retrieves semantically similar alternative expressions with higher authenticity scores from a corpus of idiomatic expressions as optimization suggestions. The retrieval process considers both semantic similarity and authenticity scores.

[0109] ,

[0110] in, The recommended alternative expression is... This is a weighting coefficient for the authenticity score, used to balance the weights of semantic similarity and authenticity in retrieval. Preferably, an authenticity threshold is preset. The value range is from 0.60 to 0.80, and this embodiment uses 0.70 as the default value; weighting coefficient Setting it to 0.5 gives semantic similarity and authenticity equal weight in search ranking.

[0111] The authenticity score and alternative expression suggestions output by the language authenticity assessment module 4 are passed to the adaptive error correction and optimization module 5 as an important reference for generating comprehensive optimization suggestions.

[0112] like Figure 6 As shown, the adaptive error correction and optimization module 5 is used to comprehensively process the output results of the aforementioned modules, generate the final error correction result text, and realize the closed-loop update and optimization of model parameters based on user feedback. In one embodiment of the present invention, this module is a key component for realizing closed-loop collaboration of the system, undertaking the core responsibilities of error correction decision-making and parameter optimization.

[0113] The adaptive error correction optimization module 5 mainly consists of four parts: an error correction candidate generation unit, a stylistic constraint screening unit, an error correction confidence ranking unit, and a feedback learning unit, forming a complete processing flow from candidate generation to screening and ranking and then to feedback learning.

[0114] The error correction candidate generation unit receives the error detection results output by the multi-dimensional error detection module 2 and generates multiple candidate correction schemes for each detected error location. The generation strategy for candidate schemes varies depending on the error type: for homophone errors, candidate schemes are selected from the set of homophones with the highest contextual fit; for similar-looking character errors, candidate schemes are selected from the set of similar-looking characters; for misuse of idioms, candidate schemes include the correct spelling of the idiom and other idioms with similar meanings; for grammatical errors, candidate schemes are generated through a grammatical correction model, including correction operations such as component addition, component deletion, and word order adjustment. The number of candidate schemes generated for each error location is set to 3 to 5 to achieve a balance between candidate diversity and computational efficiency.

[0115] The style constraint filtering unit utilizes the style type labels and style feature vectors output by the style perception module 3 to filter candidate correction schemes based on the language norm constraints of the corresponding style. These language norm constraints include three dimensions: word usage norms, sentence structure constraints, and punctuation usage constraints. The filtering rules are based on information such as word usage norms and sentence structure requirements for each style type stored in the style norm knowledge base, excluding candidates that do not conform to the style norms. For example, for academic papers, candidates with a high degree of colloquialism will be excluded during the filtering process; for everyday communication texts, more colloquial and personalized candidate schemes will be retained.

[0116] The error correction confidence ranking unit calculates and ranks the confidence scores of candidate correction schemes after being filtered by stylistic constraints, and selects the scheme with the highest confidence score as the final correction suggestion. The dynamic error correction confidence calculation algorithm proposed in this invention comprehensively considers four factors: error detection confidence, context fit, stylistic matching, and naturalness score.

[0117] ,

[0118] in, Incorrect location The One candidate amendment The error correction confidence level ranges from 0 to 1. For position The error detection confidence score is output by the multi-dimensional error detection module 2. Candidate solutions Context fit measures the degree to which a candidate solution fits the semantic context. Candidate solutions The stylistic fit measures the degree to which candidate solutions conform to stylistic norms. Candidate solutions The authenticity score measures how natural the expression of the candidate solution is. , , , The weight coefficients of the four factors satisfy the following conditions: .

[0119] The weighting coefficients reflect the relative importance of each factor in the error correction decision. Preferably, Setting it to 0.30 means the weight of the error detection confidence score is 30%. Setting it to 0.35 indicates that the weight of context fit is 35%. Setting it to 0.20 means the weight of the style matching score is 20%. Setting it to 0.15 indicates that the authenticity score has a weight of 15%. This weight configuration prioritizes contextual fit as the most important factor, because the error correction result must first be semantically consistent with the context.

[0120] The error correction confidence ranking unit sorts the candidate correction schemes for each error location in descending order of error correction confidence and selects the scheme with the highest confidence as the final correction suggestion. If the error correction confidence is lower than a preset decision threshold, the system will leave the original text unchanged and prompt the user with a suspected error, allowing the user to decide whether to modify it, thus avoiding incorrect corrections under uncertain circumstances. Preferably, the error correction decision threshold ranges from 0.70 to 0.90; this embodiment uses 0.80 as the default value.

[0121] The feedback learning unit is the core component for achieving closed-loop optimization of the system. It is responsible for collecting and processing user feedback on the error correction results. After the system presents error correction suggestions to the user, the user can choose to accept or reject the suggestion. The feedback learning unit transforms the user's feedback (acceptance or rejection) into update signals for model parameters, enabling continuous learning and optimization of the system.

[0122] When a user accepts a correction suggestion, it indicates that the suggestion is correct, and the system should positively reinforce the corresponding detection and correction paths. When a user rejects a correction suggestion, it indicates that the suggestion may be problematic, and the system should negatively adjust the corresponding parameters to reduce the probability of similar erroneous judgments. The calculation method for feedback optimization parameters is as follows:

[0123] ,

[0124] in, To provide feedback for optimizing parameters, i.e., the incremental updates to model parameters. The learning rate controls the step size for parameter updates. For parameters The gradient operator is used to calculate the partial derivatives of the loss function with respect to the model parameters θ. Let be the binary cross-entropy loss function, calculated as ℒ(y,ŷ)=-[y·log(ŷ)+(1-y)·log(1-ŷ)], which measures the difference between the system's prediction and the user's feedback. User feedback is tagged with 1 for acceptance and 0 for rejection. This represents the error correction confidence level predicted by the system. The loss function uses binary cross-entropy loss, which effectively transforms user feedback into a supervisory signal for model optimization.

[0125] Preferably, the feedback learning rate The value range is from 0.001 to 0.01, and 0.005 is used as the default value in this embodiment. A smaller learning rate can ensure the smoothness of parameter updates and avoid drastic fluctuations in model performance due to individual error feedback. The feedback optimization parameters are sent to the context semantic encoding module 1 and the multi-dimensional error detection module 2 to update the neural network parameters of these two modules, thereby realizing continuous optimization and personalized adaptation of the system.

[0126] This closed-loop feedback mechanism enables the system to continuously improve its error correction performance based on user habits and personal preferences. For error types frequently encountered by specific users, the system gradually increases the detection sensitivity of those errors; for error correction suggestions that users repeatedly reject, the system reduces the output probability of that type of suggestion. Through continuous feedback learning, the system can provide increasingly accurate and personalized error correction services for each user.

[0127] The overall workflow of the system of this invention can be summarized into the following stages:

[0128] During the input phase, users input the Chinese text to be processed into the system. The text can be a single sentence, a complete document, or an article. The system supports multiple text input methods, including direct input, file upload, and clipboard pasting.

[0129] During the encoding phase, the context semantic encoding module 1 performs semantic encoding processing on the input Chinese text. The text is first normalized by the preprocessing unit, then converted into a vector representation by the character embedding unit and the positional encoding unit, and finally generated into a context semantic vector containing multi-granular semantic information by the hierarchical attention encoding unit. The output of this phase is a context semantic vector matrix, which serves as the shared input for subsequent modules.

[0130] During the analysis phase, the multi-dimensional error detection module 2, the style perception module 3, and the language authenticity assessment module 4 receive the context semantic vectors in parallel and perform their respective analysis and processing. The multi-dimensional error detection module 2 identifies potential homophone errors, similar-looking character errors, misuse of idioms, and grammatical errors in the text through four detection units, and outputs the error detection results. The style perception module 3 determines the text's style type using a style classifier and extracts corresponding normative features from a style norm knowledge base, outputting style type labels and style feature vectors. The language authenticity assessment module 4 calculates the authenticity score of each expression unit through a comprehensive evaluation of the three types of features, retrieves alternative suggestions for expressions with insufficient authenticity, and outputs authenticity scores and alternative expression suggestions. The parallel processing of these three modules significantly improves the overall response speed of the system.

[0131] During the optimization phase, the adaptive error correction optimization module 5 comprehensively receives the output results from the above three modules, performs error correction candidate generation, stylistic constraint screening, and confidence ranking processing, and finally generates the error correction result text. The error correction result is presented to the user in a highlighted form, clearly indicating the location and content of the suggested modifications. The user can review each suggestion and choose whether to accept the system's error correction suggestions.

[0132] During the feedback phase, users review the system's error correction suggestions, choosing to accept or reject each suggestion. The feedback learning unit collects user feedback, calculates parameter update gradients, and generates feedback optimization parameters. These optimization parameters are sent to the context semantic encoding module 1 and the multi-dimensional error detection module 2 to update the corresponding model parameters, completing one round of closed-loop optimization. The system's error correction performance continuously improves with ongoing user use and feedback.

[0133] In the experimental verification phase of this invention, a comprehensive test dataset containing 50,000 Chinese text samples was used to fully evaluate the system. The test dataset covers five text types: academic papers, business documents, literary works, news reports, and daily communication, with 10,000 samples in each type. The sample sources include publicly available Chinese corpora, real users' written texts, and manually constructed error samples, covering various common text error types and language styles.

[0134] The test environment configuration for the experiment is as follows: The server is equipped with an NVIDIA A100 graphics processing unit with 40GB of video memory; the central processing unit is an Intel Xeon Gold 6248R with a clock speed of 3.0GHz and 24 cores; the memory capacity is 256GB DDR4; the operating system is Ubuntu 20.04 LTS; and the deep learning framework is PyTorch 1.12.

[0135] Experimental results show that the system of this invention achieves excellent performance indicators in various error detection tasks. In the homophone error detection task, the system achieves an accuracy of 97.2%, a recall of 95.8%, and an F1 score of 96.5%; in the similar-looking character error detection task, the system achieves an accuracy of 96.5%, a recall of 94.3%, and an F1 score of 95.4%; in the idiom misuse detection task, the system achieves an accuracy of 98.1%, a recall of 93.6%, and an F1 score of 95.8%; and in the grammatical error detection task, the system achieves an accuracy of 94.8%, a recall of 91.2%, and an F1 score of 92.9%. The overall error correction accuracy reaches 96.2%, which is approximately 15 percentage points higher than existing similarity-based matching solutions and approximately 20 percentage points higher than traditional rule-based methods.

[0136] In the task of genre classification, the system achieved a five-category accuracy of 94.5%, with each genre type achieving an accuracy rate of over 92%. The results of the language authenticity assessment showed a high correlation with the human assessment by professional linguists, with a Pearson correlation coefficient of 0.87, indicating that the system's authenticity assessment results have good reliability.

[0137] In terms of system performance, the average processing time for a single text message is 120 milliseconds, with a maximum processing time of no more than 500 milliseconds, meeting the performance requirements of real-time interaction. The system supports batch processing mode, with a throughput of up to 200 text messages per second during batch processing.

[0138] The embodiments of the present invention are not limited to the specific embodiments described above. Those skilled in the art can make various equivalent changes or substitutions based on the technical solutions of the present invention, and all such changes or substitutions should be included within the protection scope of the present invention.

Claims

1. A context-based intelligent Chinese error correction and text optimization system, characterized in that, include: The context semantic encoding module receives the Chinese text to be processed, performs context-sensitive semantic encoding on each character in the Chinese text based on a pre-trained language model, captures the context dependencies at the character, word, and sentence levels through a hierarchical attention mechanism, and generates a context semantic vector containing multi-granular semantic information. A multi-dimensional error detection module is used to receive the context semantic vector, and process the context semantic vector in parallel through a homophone error detection unit, a similar-looking character error detection unit, an idiom misuse detection unit, and a grammatical error detection unit, respectively. Based on the error confidence output by each detection unit, the error type and error location are determined, and an error detection result is generated. The style perception module is used to receive the context semantic vector, determine the style type of the Chinese text to be processed through a style classifier, extract the corresponding style norm features from a preset style norm knowledge base based on the determined style type, and fuse the style norm features with the context semantic vector to generate style type labels and style feature vectors. The language authenticity assessment module is used to receive the context semantic vector, calculate the authenticity score of each expression unit in the Chinese text to be processed through a multilayer perceptron-based authenticity assessment network, and for expression units whose authenticity scores are lower than a preset authenticity threshold, retrieve semantically similar alternative expressions with authenticity scores higher than the preset authenticity threshold from the authentic expression corpus, and generate authenticity scores and alternative expression suggestions. An adaptive error correction and optimization module is used to receive the error detection results, the text type labels, the style feature vectors, the authenticity scores, and the alternative expression suggestions. Based on the text type label's corresponding text style norm constraints, the module filters and sorts the error detection results. The text style norm constraints include word usage norm constraints, sentence structure constraints, and punctuation usage constraints. The module combines the authenticity scores and the alternative expression suggestions to generate error correction result text. The module also converts user feedback information during the error correction process into feedback optimization parameters and sends them to the context semantic encoding module and the multi-dimensional error detection module to update the model parameters.

2. The system according to claim 1, characterized in that, In the context semantic encoding module, the length of the context window ranges from 64 to 512 characters, the number of attention heads in the hierarchical attention mechanism ranges from 8 to 16, and the dimension of the semantic vector ranges from 768 to 1024.

3. The system according to claim 1, characterized in that, In the multi-dimensional error detection module, the confidence threshold for homophone error detection is 0.75 to 0.90, the confidence threshold for similar-looking character error detection is 0.70 to 0.85, the confidence threshold for idiom misuse detection is 0.80 to 0.95, and the confidence threshold for grammatical error detection is 0.65 to 0.

80.

4. The system according to claim 1, characterized in that, The style perception module includes academic papers, business documents, literary works, news reports, and daily communication, with a style matching threshold of 0.70 to 0.

90.

5. The system according to claim 1, characterized in that, The homophone error detection unit constructs a homophone confusion matrix, calculates the homophone substitution probability based on the character's pinyin encoding and the context semantic vector, and determines that a homophone error exists when the homophone substitution probability exceeds a preset homophone error detection confidence threshold.

6. The system according to claim 1, characterized in that, The similar-character error detection unit constructs a similar-character feature map, calculates the similar-character confusion probability based on the character's stroke structure encoding and the context semantic vector, and determines that a similar-character error exists when the similar-character confusion probability exceeds a preset similar-character error detection confidence threshold.

7. The system according to claim 1, characterized in that, The idiom misuse detection unit includes an idiom integrity detection subunit and an idiom semantic adaptation detection subunit. The idiom integrity detection subunit is used to detect missing or incorrectly replaced characters in the idiom. The idiom semantic adaptation detection subunit is used to calculate the semantic adaptation degree between the idiom and the context based on cosine similarity and point mutual information. The semantic adaptation degree ranges from 0 to 1.

8. The system according to claim 1, characterized in that, The language authenticity assessment module includes: The expression unit segmentation unit is used to segment the Chinese text to be processed into phrase-level expression units; The authenticity feature extraction unit is used to extract the lexical richness feature, collocation rationality feature, and expression habit conformity feature of each expression unit; The authenticity score calculation unit is used to calculate the authenticity score of each expression unit based on the vocabulary richness feature, the collocation rationality feature, and the expression habit conformity feature.

9. The system according to claim 1, characterized in that, The adaptive error correction optimization module includes: An error correction candidate generation unit is used to generate an error correction candidate set based on the error detection results; The style constraint filtering unit is used to filter the error correction candidate set based on the style specification constraints corresponding to the style type label; The error correction confidence ranking unit is used to rank the selected error correction candidates based on the error correction confidence and output the optimal error correction result.

10. The system according to claim 1, characterized in that, The adaptive error correction optimization module further includes a feedback learning unit, which is used to collect user feedback on whether they accept or reject the error correction result, update the gradient based on the feedback calculation parameters, and generate the feedback optimization parameters.

Images