An intelligent question and answer system and method for ancient literature teaching

Abstract

The application relates to the technical field of intelligent question answering, and discloses an intelligent question answering system and method for ancient literature teaching, which comprises the following steps: obtaining a creation dynasty timestamp and a genre attribute of a target query text to generate a Boolean type routing flag bit, setting pronunciation attribute data with a time attribute less than or equal to the timestamp as a forced filtering condition, dynamically generating a time-space semantic slicing database, when the flag bit is turned on, converting the target query text into a target matrix, performing bitwise XOR and bitwise logical AND operations on the target matrix, a built-in standard meter matrix and a mask matrix, obtaining a basic meter error total number, extracting a candidate text by using the timestamp, and calculating the comprehensive correlation weight of the target query text and the candidate text in combination with the basic meter error total number. The application excludes interference of pronunciation data in later generations through time dimension isolation, and replaces conventional string comparison with matrix bit operation, so that the operation efficiency of meter error statistics of the system is improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent question-answering technology, specifically to an intelligent question-answering system and method for teaching ancient literature. Background Technology

[0002] The intelligent question-answering system for teaching ancient literature is one of the application forms of natural language processing tools. The system receives ancient literature query texts input by users, performs retrieval and matching operations in the built-in ancient literature corpus, and outputs relevant text parsing or literature association content to assist in the teaching process of ancient literature.

[0003] Existing intelligent question-answering systems primarily utilize word segmentation algorithms and text similarity calculation models to perform retrieval tasks when processing ancient Chinese literature. The system segments the input query text and extracts feature words, then uses these feature words to search for candidate texts in a literature database. Finally, it calculates text similarity weights based on word overlap and outputs matching results in order of weight value. Some question-answering systems also include a separate phonological query function, using a pre-set dictionary database to provide pronunciation annotations and prosody verification results for ancient Chinese characters.

[0004] Existing intelligent question-answering solutions for teaching ancient Chinese literature have several limitations in practical applications. When providing ancient Chinese character pronunciation data, existing systems lack a mandatory filtering mechanism based on historical time, resulting in extracted pronunciation data containing phonetic variations from after the target document's composition date, leading to pronunciation matching deviating from historical facts. Furthermore, when performing phonological and prosodic verification, current techniques rely on conventional character-by-character string comparison to determine tonal matching, which not only consumes significant system computing resources but also has low computational efficiency in prosodic error statistics. Moreover, when calculating the association weight between the query text and candidate texts, current association calculation models only compare the literal content of the texts, failing to incorporate the inherent phonological and prosodic error characteristics of ancient literature into the association calculation dimension. This results in the system being unable to consider both the literal features and underlying prosodic features of the documents, leading to deviations in the final associated document matching results.

[0005] Therefore, the purpose of this invention is to provide an intelligent question-and-answer system and method for teaching ancient literature, in order to address the shortcomings of existing technologies. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides an intelligent question-and-answer system and method for teaching ancient literature. It solves the problems in existing intelligent question-and-answer systems for teaching ancient literature, such as the lack of historical time dimension filtering of pronunciation data leading to pronunciation matching deviating from historical facts, the reliance on conventional string character-by-character comparison for phonological verification resulting in high computational resource consumption and low computational efficiency, and the failure to take into account the underlying metrical features in the calculation of document association weights, leading to deviations in the final matching results.

[0007] To achieve the above objectives, the present invention provides the following technical solution: The first aspect of this invention provides an intelligent question-and-answer system for teaching ancient literature, comprising: The receiving and parsing module is used to obtain the target query text and extract the creation dynasty timestamp and genre attribute, and generate a boolean routing flag based on the genre attribute.

[0008] The data isolation module receives numerical timestamps corresponding to the creation dynasty timestamps, stores pronunciation attribute data with historical time attributes, sets historical time attributes less than or equal to the numerical timestamps as a mandatory filtering condition, and dynamically generates a spatiotemporal semantic slice database. The system constructs mandatory filtering conditions using numerical timestamps to limit the time dimension range of pronunciation attribute data, ensuring that pronunciation data matches the corresponding historical period.

[0009] The phonology verification module, when the Boolean routing flag is enabled, calls the spatiotemporal semantic slice database to convert the target query text into a target matrix. Based on genre attributes, it reads the corresponding standard metrical matrix and mask matrix from the system's built-in rule base. It then performs a bitwise XOR operation between the target matrix and the standard metrical matrix, and performs a bitwise logical AND operation between the result of the XOR operation and the mask matrix to calculate and output the total number of basic metrical errors. The system uses matrix bitwise operations to replace string comparison, calculating and outputting a numerical total number of metrical errors.

[0010] The reverse retrieval module is used to extract candidate text from the spatiotemporal semantic slice database and output it, using the creation dynasty timestamp as a time filtering parameter.

[0011] The association calculation module is used to calculate the comprehensive association weight between the target query text and the candidate text by combining the total number of basic metrical errors.

[0012] Furthermore, the receiving and parsing module includes a text preprocessing component. This component employs a deep learning model architecture with bidirectional long short-term memory network connected to conditional random fields to convert the target query text into a sequence of character vectors, outputting word segmentation and part-of-speech tagging sequences. The receiving and parsing module extracts the author's name or work title from the target query text, compares it against a pre-built literary knowledge graph to extract historical data, maps the historical data to numerical timestamps, and outputs them as the creation dynasty timestamp.

[0013] Furthermore, the spatiotemporal semantic slice database is equipped with a weighted priority disambiguation strategy. When loading the spatiotemporal semantic slice database, the data isolation module calculates the authority score of the documents through a built-in document scoring neural network model, and loads the data record with the highest authority score into the active cache area of ​​the spatiotemporal semantic slice database for use by the phonological verification module. The document scoring neural network model adopts a multilayer perceptron architecture, and the input data includes the time difference between the document's publication date and the present, the number of times the document has been cited in subsequent academic papers, and the document's collation version level.

[0014] Furthermore, the phonology verification module extracts the glyph codes of individual characters from the word segmentation and part-of-speech tagging sequences, and combines these codes with the word phrases containing the individual characters and the part-of-speech tags to generate a joint query key. It then retrieves the pronunciation attribute data corresponding to the joint query key from the spatiotemporal semantic slice database. If pronunciation attribute data is not obtained, the phonology verification module triggers a semantic disambiguation neural network model. This model employs a Transformer encoder architecture based on a self-attention mechanism, outputting a contextual dynamic feature vector for each individual character. The pronunciation attribute data is obtained by calculating the cosine similarity between the contextual dynamic feature vector and the baseline feature vector of historical example sentences.

[0015] Furthermore, the reverse retrieval module uses term frequency and inverse document frequency algorithms to extract words in the target query text whose weight distribution values ​​exceed a preset term frequency threshold as specified feature words, and obtains a candidate text set containing the specified feature words. The reverse retrieval module uses a dynamic programming algorithm to construct an edit distance matrix, calculates the minimum number of single-character editing operations required to convert candidate text into target query text, and uses the minimum number of single-character editing operations as the absolute edit distance. The absolute edit distance is used to calculate the basic text similarity, and the difference character pairs between the target query text and each candidate text in the candidate text set are extracted and output to the association calculation module.

[0016] Furthermore, the association calculation module receives the differing character pairs and extracts the historical original words and the current variant words from the differing character pairs. It then expands the preset truncation radius bidirectionally to the left and right to generate a historical context window and a target context window. The association calculation module calls the static word vector lookup table pre-trained on a massive unlabeled corpus of ancient books using a continuous bag-of-words model architecture, mapping the characters in the historical context window and the target context window into historical context feature matrices and target context feature matrices, respectively.

[0017] Furthermore, the association calculation module replaces the corresponding current variant words in the target query text with historical original words to generate a hypothetical reconstruction text, which is then sent to the phonology verification module. The phonology verification module receives the hypothetical reconstruction text, converts it into a hypothesis matrix, and returns it to the association calculation module. The logic operation unit configured within the association calculation module performs Boolean bitwise operations on the hypothesis matrix, performs a bitwise XOR operation between the hypothesis matrix and the standard metrical matrix, and performs a bitwise logical AND operation between the result of the bitwise XOR operation and the mask matrix to obtain the total number of hypothetical metrical errors. The association calculation module calculates the difference between the total number of basic metrical errors and the total number of hypothetical metrical errors to generate a metrical repair gain index, and uses this metrical repair gain index to calculate the phonological confidence weight.

[0018] Furthermore, when the Boolean routing flag is off, the target query text bypasses the phonological verification module and is transmitted to the reverse retrieval module and the association calculation module. The association calculation module includes a semantic density evaluation component. This component calls the entity recognition interface to extract the sum of the number of low-frequency content words and proper nouns in the target query text, and calculates the semantic density index by combining the total character length of the target query text with a smoothing factor. The association calculation module then calls a global semantic encoding neural network model employing a bidirectional long short-term memory network architecture to output the global semantic feature tensor of the target query text, which reflects the overall historical narrative intent of the long essay.

[0019] Furthermore, the system also includes a graph generation module. This module instantiates the target query text and candidate documents found in historical searches as discrete nodes in the graph. It uses the numerical timestamp attribute of the discrete nodes to construct a directed acyclic graph (DAG). The graph generation module integrates a path ranking neural network model using a multilayer perceptron architecture. It traverses the DAG to extract complete unidirectional paths and concatenates the equivalent feature tensors of discrete nodes on the complete unidirectional paths with comprehensive association weights to generate a variable-length path feature matrix. After performing adaptive average pooling on the path feature matrix, it is input into the path ranking neural network model, outputting a confidence probability distribution. The model then sorts the outputs in descending order based on the confidence probability distribution and encapsulates them into a multi-dimensional structured output file.

[0020] A second aspect of this invention provides an intelligent question-and-answer method for teaching ancient literature, applied to an intelligent question-and-answer system for teaching ancient literature, comprising the following steps: Obtain the target query text and extract the creation dynasty timestamp and genre attribute. Generate a boolean routing flag based on the genre attribute and output the creation dynasty timestamp and boolean routing flag.

[0021] It receives the numerical timestamp corresponding to the creation dynasty timestamp, and dynamically generates a spatiotemporal semantic slice database based on the stored pronunciation attribute data with historical time attributes, setting historical time attributes less than or equal to the numerical timestamp as a mandatory filtering condition.

[0022] When the received Boolean routing flag is in the enabled state, the spatiotemporal semantic slice database is called to convert the target query text into a target matrix. Based on the genre attribute, the corresponding standard metrical matrix and mask matrix are read from the system's built-in rule base. The target matrix and the standard metrical matrix are XORed, and the result of the XOR operation is ANDed with the mask matrix to obtain the total number of basic metrical errors and output it.

[0023] Using the era timestamp of the creation as a time filtering parameter, candidate texts are extracted from the spatiotemporal semantic slice database and output.

[0024] Receive the total number of basic metrical errors and candidate texts, and calculate the comprehensive association weight between the target query text and the candidate texts by combining the total number of basic metrical errors.

[0025] This invention provides an intelligent question-and-answer system and method for teaching ancient literature. It has the following beneficial effects: 1. This invention sets up a data isolation module, sets historical time attributes less than or equal to numerical timestamps as a mandatory filtering condition, and dynamically generates a spatiotemporal semantic slice database. It performs time-dimensional filtering on pronunciation attribute data, eliminates interference from later data after the dynasty corresponding to the target query text, and makes the pronunciation data called by the system match the corresponding historical period.

[0026] 2. This invention converts the target query text into a target matrix by setting up a phonology verification module. The target matrix is ​​then XORed with the standard metrical matrix and ANDed with the mask matrix to obtain the total number of basic metrical errors. The matrix bitwise operation mechanism replaces the conventional string comparison process for phonology verification, which reduces the system's computational resource consumption and improves the computational efficiency of metrical error statistics.

[0027] 3. By setting up an association calculation module, this invention uses the total number of basic prosody errors as a combined parameter when calculating the comprehensive association weight between the target query text and the candidate text. Based on conventional text content comparison, it introduces the phonological and prosodic error dimension of ancient literature, so that the matching between documents takes into account both the literal features of the text and the underlying prosodic features, thereby improving the effectiveness of text association matching in the question-answering system. Attached Figure Description

[0028] Figure 1 This is a system architecture diagram of the present invention; Figure 2 This is a flowchart illustrating the overall process of the method of the present invention. Figure 3 This is a comparison chart of the traceability accuracy and cross-generational false alarm rate of different systems according to the present invention; Figure 4 This is a comparison chart showing the trend of system response time as a function of the underlying data volume in this invention. Detailed Implementation

[0029] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] See attached document Figure 1 This invention provides an intelligent question-and-answer system for teaching ancient literature, comprising: a receiving and parsing module, a data isolation module, a phonetic verification module, a reverse retrieval module, an association calculation module, and a graph generation module.

[0031] The receiving and parsing module is used to obtain the input target query text and extract metadata attributes. Metadata attributes include the creation dynasty timestamp and genre attribute.

[0032] The data isolation module stores historical vocabulary definitions and pronunciation attribute data with a time dimension identifier. The data isolation module is configured with a view isolation component to control the data reading range based on the time dimension.

[0033] The phonology verification module is equipped with a matrix conversion unit and a logic operation unit. The phonology verification module receives text characters, converts the text characters into mathematical matrix form, and performs bitwise operations.

[0034] The reverse retrieval module is used to locate candidate texts on the timeline. It is equipped with a character comparison component to extract difference character pairs between texts.

[0035] The association calculation module receives basic text similarity data and matrix verification error data, and calculates the final association weight between the target query text and candidate texts based on the set weight allocation logic.

[0036] The graph generation module is used to construct topological paths based on the final association weights, generate graph structure data with nodes and connections, and output the system's parsing results.

[0037] See attached document Figure 2 This invention provides a structured question-answering method for ancient literature based on spatiotemporal semantic isolation and phonological adaptive association, comprising the following steps: The receiving and parsing module obtains the target query text passed from the front end, parses the target query text, and extracts the creation dynasty timestamp and genre attribute corresponding to the target query text. The data isolation module receives the creation dynasty timestamp, locks the spatiotemporal semantic slice database of the corresponding time dimension according to the set logic, and restricts the subsequent word meaning query operation and phonological attribute query operation to the scope of the spatiotemporal semantic slice database. The current processing logic branch is determined based on the genre attribute. When the genre attribute meets the set metrical requirements, the phonetics verification module calls the pronunciation attribute data in the spatiotemporal semantic slice database, converts the target query text into a target matrix, reads the corresponding standard metrical matrix and mask matrix, performs XOR operation and logical AND operation on the target matrix, and calculates the total number of basic metrical errors of the target query text. The reverse retrieval module uses the creation dynasty timestamp as the time filter parameter and performs a reverse timeline retrieval in the data isolation module to obtain a set of candidate texts containing specified feature words. The reverse retrieval module calculates the basic text similarity between the target query text and each candidate text in the candidate text set, and extracts the difference character pairs between the target query text and the candidate text. The difference character pairs include the historical original words in the candidate text and the current variant words in the target query text. The association calculation module triggers the association calculation logic based on the genre attribute. When the genre attribute has metrical requirements, the association calculation module replaces the current variant words in the target query text with the original historical words to generate hypothetical text. The phonology verification module receives the hypothetical text, performs matrix transformation and Boolean operation on the hypothetical text, obtains the total number of hypothetical errors, and the association calculation module calculates the difference between the total number of hypothetical errors and the total number of basic metrical errors. The difference is combined with the basic text similarity to calculate the final association weight of the candidate text. The graph generation module obtains the final association weight of each candidate text output by the association calculation module, and establishes the candidate text whose final association weight meets the threshold setting condition as the source node of the allusion. The graph generation module constructs a topological directed edge from the target query text to each source node of the allusion within the storage space, generates a directed acyclic graph for tracing the origins of the allusions, and integrates the text interpretation data with the matrix verification report to form structured parsed data for external output.

[0038] During the front-end data access phase, parameter extraction and logical branch setting are mainly performed through the receiving and parsing module.

[0039] The receiving and parsing module receives the target query text from the front end. To transform natural language text into structured feature vectors that can be processed by computers, the receiving and parsing module is equipped with a text preprocessing component. This component performs basic word segmentation and part-of-speech tagging on the target query text. The text preprocessing component adopts a deep learning model architecture with a bidirectional long short-term memory network connected to a conditional random field. Specifically, the input layer of the model transforms the target query text into a sequence of character vectors. The hidden layer captures the global contextual features of the character vector sequence through the bidirectional long short-term memory network and outputs an emission probability matrix. The state transition layer learns the transition probability constraints between classical Chinese part-of-speech tags based on the conditional random field, and finally outputs the globally optimal word segmentation and part-of-speech tagging sequence.

[0040] The receiving and parsing module triggers a metadata extraction mechanism to obtain the creation dynasty timestamp. It then uses a built-in entity recognition interface to extract the author's name or work title from the target query text and compares it against the system's pre-built literary knowledge graph using this name or title as the primary search key. If no matching author's name or work title is found in the literary knowledge graph, the receiving and parsing module automatically triggers a fallback strategy, extracting dynasty-specific keywords or imperial reign titles from the target query text for estimation and matching, thus preventing the system from stalling due to missing metadata. The receiving and parsing module also retrieves historical period data returned from the comparison and maps this data into continuous numerical timestamps. These numerical timestamps represent the upper limit of the creation era of the target query text. Converting discrete dynasty names into continuous numerical forms provides filtering parameters for subsequent data isolation modules to support mathematical comparison operations.

[0041] While extracting the numerical timestamps, the receiving and parsing module simultaneously extracts the genre attributes of the target query text. Based on the classification labels returned by the literary knowledge graph, or, if the graph's classification labels are missing, by statistically analyzing the word count, sentence symmetry, and rhyme distribution characteristics of the target query text, the receiving and parsing module determines the genre attributes of the target query text. Genre attributes include data categories such as regulated verse, ci poetry, classical poetry, fu (a type of classical Chinese prose), and prose.

[0042] The receiving and parsing module sets the routing flag based on the genre attributes to complete the configuration of the underlying processing branch of the system. The system has a pre-built genre classification mapping table. The genre classification mapping table sets genres with strict tonal constraints as metrical texts and genres without fixed tonal constraints as non-metrical texts. The receiving and parsing module inputs the extracted genre attributes into the genre classification mapping table for Boolean logic judgment and generates Boolean routing flags.

[0043] When the genre attribute is determined to be metrical text, the receiving and parsing module sets the Boolean routing flag to 1. The flag being on means that the system will activate the phonetic verification module in the subsequent workflow, and then perform matrix transformation and verification operations. When the genre attribute is determined to be non-metrical text, the receiving and parsing module sets the Boolean routing flag to 0. The flag being off means that the system will bypass the phonetic verification module in the subsequent workflow. This bypass design is intended to prevent the system from attempting to construct a standard metrical matrix for texts such as prose that lack standard tonal patterns, thereby causing abnormalities in the underlying logic operations.

[0044] By setting a Boolean routing flag, the system can be compatible with various ancient documents with significant structural differences within the same processing architecture, improving the stability of the underlying control logic. After completing parameter extraction and routing configuration, the receiving and parsing module sends a numerical timestamp to the data isolation module to trigger the subsequent data isolation reading process.

[0045] After receiving the numerical timestamp from the parsing module, the data isolation module triggers a dynamic loading workflow for spatiotemporal semantic slices.

[0046] The data isolation module is configured with a relational database or graph database at its core. The relational database or graph database stores historical vocabulary definition data and pronunciation attribute data. To support time-based retrieval operations, each record in the historical vocabulary definition data and pronunciation attribute data is configured with a dynasty time tag corresponding to the historical period. During the data entry stage, the dynasty time tag is mapped to a continuous historical time attribute. Both the historical time attribute and the numerical timestamp are represented by integer years in the Gregorian calendar, with a value range of integers including both negative and positive numbers. This enables unified calculation across the timeline before and after the Common Era.

[0047] To reduce the possibility of the target query text being affected by semantic changes or pronunciation changes in subsequent processing stages, the data isolation module is configured with a view isolation component. The view isolation component uses numerical timestamps to build a data access barrier. Specifically, the view isolation component achieves data isolation by dynamically generating structured query statement views in the underlying query engine. The view generation logic sets historical time attributes less than or equal to numerical timestamps as a mandatory filtering condition. The data isolation module uses the mandatory filtering condition to generate a temporary data access view for the current query session.

[0048] The temporary data access view is the spatiotemporal semantic slice database. The logical representation of the data set inside the spatiotemporal semantic slice database is as follows: ; in, Represents a spatiotemporal semantic slice database; A single data record representing the historical vocabulary definition data and pronunciation attribute data; This represents the historical time attribute corresponding to a single data record. A numeric timestamp representing the target query text.

[0049] By constructing a spatiotemporal semantic slice database, the system limits subsequent dictionary-dependent query operations to a fixed historical time window. When the phonology verification module reads pronunciation attribute data, or when the reverse retrieval module compares text features, it only accesses data records within the spatiotemporal semantic slice database. The mechanism of establishing view isolation based on the time dimension can limit the scope of data reading, which helps to reduce the adverse effects of cross-era semantic drift on the accuracy of tracing the origins of ancient documents at the system architecture level.

[0050] To address the issue of multiple pronunciations and annotations for the same word in multiple versions of ancient books within the same dynasty, the spatiotemporal semantic slice database is configured with a weighted priority disambiguation strategy. When the data isolation module loads the spatiotemporal semantic slice database, if it detects multiple data records with the same glyph and historical time attributes, the view isolation component sorts the multiple data records in descending order based on the document authority score, and only loads the data record with the highest document authority score into the active cache area of ​​the spatiotemporal semantic slice database.

[0051] To obtain accurate document authority scores, the view isolation component integrates a document scoring neural network model. This model employs a multilayer perceptron architecture. The input data for the model consists of multidimensional feature vectors. The specific dimensions of these vectors include the time difference between the document's publication date and the present, the number of times the document has been cited in subsequent academic papers, and the document's collation and version rating. The model comprises an input layer, two fully connected hidden layers, and an output layer. The two fully connected hidden layers use a linear rectified function as the activation function to extract the nonlinear correlation features of the multidimensional feature vectors. The output layer uses a normalized exponential function, and the output result represents the document authority score. The document authority score ranges from 0 to 1, representing a continuous floating-point number. Its business meaning signifies the reliability of a single data record within the academic community.

[0052] Regarding the construction and training process of the document scoring neural network model, the system collects the verification records of historical ancient dictionaries as training samples. The labels of the training samples are manually annotated by domain experts based on the recognized authority of historical classics. The label values ​​are also within the range of [0, 1]. During the training phase, mean squared error is used as the loss function. The gradient is calculated and the network weight parameters are updated using the backpropagation algorithm. By introducing the calculation and filtering logic of document authority scoring, the system helps to improve the uniqueness and certainty of the data provided by the spatiotemporal semantic slice database to the upper-level modules, and avoids branch explosion in downstream verification operations caused by polysemous pronunciation data.

[0053] The receiving and parsing module outputs a Boolean routing flag based on the genre attribute. When the Boolean routing flag is equal to 1, it means that the system determines that the target query text belongs to the metrical text category. At this time, the phonetic verification module is triggered to execute the dynamic context polyphonic character disambiguation workflow.

[0054] The phonological verification module receives the target query text and calls the structured sequence generated by the receiving and parsing module in the early preprocessing stage. The structured sequence contains the word segmentation results and part-of-speech tags of the target query text. In classical Chinese, there are a large number of Chinese characters whose tone changes due to changes in part of speech or meaning. For example, characters with the part of speech of nouns are often pronounced with a level tone, but when converted to the part of speech of verbs, they are pronounced with an oblique tone. Relying on a single character form for static table lookup can easily lead to ambiguity in pronunciation mapping, which in turn can cause a chain of errors in subsequent matrix verification.

[0055] To reduce the risk of ambiguity in pronunciation mapping caused by looking up a single character in a table, the phonological verification module traverses every single character in the structured sequence. The phonological verification module extracts the character's glyph encoding and combines it with the phrase text containing the character and the corresponding part-of-speech tag to generate a joint query key. The structure of the joint query key includes three dimensions: character shape, phrase, and part-of-speech tag.

[0056] The phonology verification module submits the joint query primary key to the data isolation module. The data isolation module restricts the joint query to the loaded spatiotemporal semantic slice database. The pronunciation attribute data within the spatiotemporal semantic slice database is configured with corresponding applicable part-of-speech and applicable phrase fields. By matching the joint query primary key, the phonology verification module obtains unique pronunciation attribute data corresponding to the current context. Specifically, the pronunciation attribute data is represented as binary classification labels for subsequent matrix operations. The phonology verification module maps level tone attributes to the value 0 and oblique tone attributes to the value 1. Converting the pronunciation attributes of natural language into binary values ​​aims to provide a standard input format for the Boolean XOR operations of the underlying computer.

[0057] If the phonology verification module fails to match the corresponding pronunciation attribute data through the joint query primary key in the spatiotemporal semantic slice database, it indicates that there are rare word combinations in the target query text that are not included in the spatiotemporal semantic slice database. In order to improve the completeness of the algorithm logic and avoid the interruption of the verification process, the phonology verification module triggers a downgrade disambiguation strategy based on vector similarity.

[0058] The phonology verification module is internally configured with a semantic disambiguation neural network model, which adopts a Transformer encoder architecture based on a self-attention mechanism. Specifically, the semantic disambiguation neural network model contains a word embedding layer, a positional encoding layer, and multiple stacked encoder modules. Each encoder module is connected to a multi-head self-attention mechanism sublayer and a feedforward neural network sublayer. The input data of the semantic disambiguation neural network model is a complete sentence sequence containing the target single character. The preprocessing logic transforms the complete sentence sequence into a basic vector matrix through the word embedding layer and superimposes the positional information output by the positional encoding layer to preserve the word order features of Chinese characters. The basic vector matrix is ​​then processed by the multi-head self-attention mechanism sublayer to calculate the dependency weights between different Chinese characters within the sentence, and finally outputs the context dynamic feature vector of the target single character, denoted as . The dimension of the context dynamic feature vector is determined by the number of hidden layer neurons preset by the system, and in business terms, it represents the semantic tendency and pronunciation probability distribution of the target single character in the current syntactic environment.

[0059] For the aforementioned downgrading and disambiguation process, the phonological verification module extracts historical example sentences of the target single character under different pronunciations from the spatiotemporal semantic slice database. The phonological verification module also inputs the historical example sentences into the semantic disambiguation neural network model to extract the corresponding baseline feature vectors.

[0060] The phonology verification module calculates the cosine similarity between the context dynamic feature vector and the baseline feature vector. The cosine similarity is denoted as and the formula for calculating the cosine similarity is: ; in, Represents cosine similarity; The contextual dynamic feature vector representing the target single character; The baseline feature vector representing historical example sentences; Dot product operation representing vectors; and These represent the L2 norm of the corresponding vectors, and the cosine similarity is a continuous floating-point number between [-1, 1]. The larger the value, the closer the current context of the target character is to the context of the historical example sentence.

[0061] The phonology verification module selects the historical example sentences corresponding to the maximum cosine similarity and assigns the pronunciation attribute data associated with the historical example sentences to the target single character. Through the combination of dynamic joint query and downgraded similarity calculation, the phonology verification module assigns relatively certain pronunciation attribute data to each single character, thereby providing a relatively reliable data foundation for subsequent matrix transformation. For the multi-head attention mechanism and feedforward neural network calculation process in the Transformer encoder architecture, those skilled in the art can refer to standard natural language processing models for implementation. The related matrix multiplication and residual connection methods are well-known technologies in this field and will not be elaborated here.

[0062] Regarding the construction and training process of the semantic disambiguation neural network model, the system collects a corpus of ancient books with manually annotated pronunciations as training samples. The training samples are obtained from annotated ancient poetry and prose collections proofread by domain experts. The training samples contain the original sentence sequences of ancient documents and the actual pronunciation labels of the corresponding target single characters. During the training phase, the multi-class cross-entropy loss function is used to calculate the error loss between the probability distribution of the model's predicted output and the actual pronunciation labels. The model uses the backpropagation algorithm to calculate the gradient of the loss function with respect to the network nodes and uses an adaptive moment estimation optimizer to update the weight parameters of the Transformer encoder. Through iterative training until the cross-entropy loss function converges, a precise mapping relationship between the semantic disambiguation neural network model and the contextual features of ancient Chinese is established.

[0063] After completing the dynamic context polyphonic character disambiguation and obtaining the pronunciation binary classification labels of each single character in the target query text, the phonology verification module triggers the orthogonalized phonology matrix verification workflow.

[0064] The phonology verification module extracts the binary pronunciation attribute sequence of the target query text. Based on the physical layout structure of the target query text, the phonology verification module reorganizes the binary pronunciation attribute sequence into a two-dimensional mathematical matrix to generate the target matrix. The number of rows in the target matrix corresponds to the number of sentences contained in the target query text, and the number of columns in the target matrix corresponds to the number of characters contained in a single sentence. The matrix elements inside the target matrix take values ​​of 0 or 1, where a value of 0 represents a level tone and a value of 1 represents an oblique tone. Converting natural language into a two-dimensional matrix aims to provide a standardized data structure for subsequent discrete mathematical operations.

[0065] During the generation of the target matrix, if the actual number of characters in the target query text is inconsistent with the number of characters required by the standard metrical rules, such as missing or extra characters, the phonology verification module will trigger the matrix alignment and completion mechanism. The phonology verification module will perform end-bit truncation or placeholder completion based on null value identifiers on the target matrix according to the column dimension of the standard metrical matrix to ensure the dimensional consistency of subsequent matrix Boolean operations.

[0066] After the target matrix is ​​constructed, the phonological verification module reads the corresponding standard metrical matrix and mask matrix from the system's built-in rule base based on the genre attributes extracted by the receiving and parsing module in the early stage. The system's built-in rule base is a digital repository pre-built based on traditional classic metrical literature such as "Shi Yun He Bi", "Ci Lin Zheng Yun", and "Qin Ding Ci Pu". To ensure the reading efficiency during system operation, the rule base uses a key-value pair mapping data structure for storage. Specifically, the sub-category of genre attributes (such as the rhyme of the first line of a five-character quatrain or the specific name of the ci poem "Shui Diao Ge Tou") is used as the retrieval key. Each retrieval key uniquely maps to a set of standard metrical matrices and mask matrices corresponding to the dimension.

[0067] The standard metrical matrix maintains the same dimension as the target matrix. The standard metrical matrix records the standard tonal arrangement rules of the genre under ideal conditions. The standard metrical matrix is ​​a two-dimensional static array formed by domain experts who transformed the standard tonal patterns specified in traditional classics. The matrix elements are also composed of values ​​0 and 1, where the value 0 represents that the standard requires the position to be pronounced with a level tone, and the value 1 represents that the standard requires the position to be pronounced with an oblique tone.

[0068] The mask matrix is ​​used to mark the positions of characters in the metrical rules that allow for both level and oblique tones. Specifically, there are objective phenomena in classical Chinese poetry where some positions do not have strict requirements for level and oblique tones. In order to transform this ambiguous literary rule into precise calculation logic, the phonology verification module configures the values ​​in the mask matrix. The mask matrix sets the value to 1 at coordinate positions that require strict matching of level and oblique tones, and sets the value to 0 at coordinate positions that allow for both level and oblique tones. By pre-fixing the above rule base, the system avoids performing complex real-time calculations of classical metrical rules during the verification stage, and provides a complete reference tensor for the underlying hardware-level bitwise operations.

[0069] To extract the pronunciation nodes that truly violate metrical constraints in the target query text, the logic operation unit configured within the phonology verification module performs Boolean bitwise operations on the aforementioned matrix. The phonology verification module performs an XOR operation on the target matrix and the standard metrical matrix, and then performs a logical AND operation on the result of the XOR operation with the mask matrix to generate the metrical error matrix. The formula for the relevant matrix operations is expressed as follows: ; in, Represents the fundamental error matrix of the out-of-law rule; The target matrix is ​​generated by transforming the target query text. Represents the standard metrical matrix corresponding to the genre attributes; represents the mask matrix that marks the loose rules. Represents the bitwise XOR operator for matrices; This represents the bitwise logical AND operator for matrices.

[0070] In the XOR operation, the XOR output is 0 when an element in the target matrix is ​​the same as an element in the standard metrical matrix; otherwise, it is 1. Through the XOR operation, the phonology verification module can extract all coordinate nodes in the target query text that are inconsistent with the standard metrical rules. In the logical AND operation, the phonology verification module uses the positions with values ​​of 0 in the mask matrix to force the error results under the relaxed rule coordinates to 0. Through the superposition of the above two matrix operations, the phonology verification module can filter out the out-of-meter nodes in the relaxed positions and retain the error coordinates that truly violate the metrical rules.

[0071] Subsequently, the phonological verification module sums up all the values ​​in the basic metrical error matrix to calculate the total number of basic metrical errors in the target query text. The total number of basic metrical errors is denoted as , and the relevant calculation formula is expressed as: ; in, Represents the total number of basic metrical errors; Represents the summation operator for matrix elements; total number of basic metrical errors. The data type is a non-negative integer, and the business meaning represents the number of absolute metrical errors and the total number of basic metrical errors contained in the target query text in the current state. The calculation results will be cached by the phonology verification module and transmitted to the associated calculation module as a benchmark comparison parameter for the subsequent reverse hypothesis deduction process.

[0072] For the underlying hardware instruction implementation of bitwise XOR and bitwise AND operations of matrices, those skilled in the art can use standard arithmetic logic units for instruction compilation and register calls. Boolean algebra-based underlying bitwise operation scheduling is a well-known technology in this field and will not be elaborated here.

[0073] In order to accurately trace the origins of allusions in ancient literature, the reverse retrieval module triggers a forced timeline filtering mechanism after receiving the target query text and the corresponding numerical timestamp.

[0074] The reverse retrieval module uses numerical timestamps as the upper limit parameter for time filtering and sends historical document retrieval requests to the data isolation module. In order to improve the reliability of the retrieved documents as the source of the target query text, the reverse retrieval module sets strict time constraints in the retrieval request. Specifically, the reverse retrieval module requires that the creation time attribute of the candidate documents must be strictly less than the numerical timestamp. Through the forced time axis filtering mechanism, the reverse retrieval module removes document data from the spatiotemporal semantic slice database that is in the same era as or later than the target query text. This filtering mechanism helps to reduce the probability of chronological inversion tracing errors at the physical and logical level.

[0075] To retrieve potentially related text fragments from massive amounts of historical documents, the reverse retrieval module includes a feature word extraction component. This component employs term frequency (TF) and inverse document frequency (IVF) algorithms to calculate the weight distribution of each word in the target query text. The formulas for the TF and IVF algorithms are as follows: ; in, The numerical values ​​representing the weight distribution of words; This represents the frequency of single-character or multi-character words appearing in the target query text; The inverse document frequency (IVF) is calculated by dividing the total number of documents in the underlying search engine corpus by the number of documents containing the target word and then taking the logarithm. The reverse retrieval module extracts words whose weight distribution values ​​exceed a preset word frequency threshold as specified feature words. The preset word frequency threshold is a continuous floating-point number between [0.1, 0.5], and its business meaning represents the degree of importance of the word in the target query text. The reverse retrieval module inputs the specified feature words into the underlying search engine to obtain a set of candidate texts containing the specified feature words.

[0076] After obtaining the candidate text set, the reverse retrieval module calculates the basic text similarity between the target query text and each candidate text in the candidate text set. In order to capture the variation and adaptation of texts in ancient literature, the reverse retrieval module is equipped with a character comparison component. The character comparison component uses a dynamic programming algorithm to construct an edit distance matrix. In the process of constructing the edit distance matrix, the character comparison component initializes a two-dimensional state matrix. The number of rows and columns of the two-dimensional state matrix correspond to the character length of the target query text plus one and the character length of the candidate text plus one, respectively. The character comparison component iteratively calculates the value of each element in the two-dimensional state matrix according to the state transition equation. The logic of the state transition equation is as follows: when the two characters at the aligned position are the same, the current state value inherits the state value of the adjacent position at the top left corner; when the two characters at the aligned position are different, the current state value takes the minimum value of the three adjacent positions at the left, top, and top left corners and adds one.

[0077] By calculating the minimum number of single-character editing operations required to transform candidate text into target query text, the character comparison component extracts the value of the lower right element of the two-dimensional state matrix as the absolute edit distance. The number of single-character editing operations includes the total number of insertion, deletion, and replacement operations.

[0078] The reverse retrieval module uses absolute edit distance to calculate normalized basic text similarity. The formula for calculating basic text similarity is: ; in, Represents the basic text similarity; Represents the absolute edit distance between the target query text and the candidate texts; A sequence of characters representing the target query text. A sequence of characters representing the candidate text; Represents the length of the target query text in characters; The length of the candidate text in characters; This represents the function that takes the maximum value. The value range is a continuous floating-point number between [0, 1]. The closer the value is to 1, the more similar the target query text and the candidate text are in terms of literal structure.

[0079] To provide foundational data for subsequent adaptive association calculations, the reverse retrieval module extracts the difference character pairs between the target query text and candidate texts after calculating the basic text similarity. Specifically, the reverse retrieval module locates the character coordinates where the replacement operation occurred by tracing the optimal path of the two-dimensional state matrix. The module extracts the replaced characters from the candidate text as historical original words and extracts the corresponding characters from the target query text as current variant words. These historical original words and current variant words together constitute difference character pairs. The reverse retrieval module then sends the basic text similarity and difference character pairs to the association calculation module to support downstream hypothesis deduction and weight calculation.

[0080] After receiving the basic text similarity and difference character pairs from the reverse retrieval module, the association calculation module triggers the mutated character context extraction logic.

[0081] To provide complete contextual data for subsequent semantic feature comparison, the association calculation module is equipped with a context extraction component. The context extraction component extracts the historical original words and the current variant words from the difference character pairs. The association calculation module uses the absolute position coordinates of the historical original words in the candidate text and the absolute position coordinates of the current variant words in the target query text as the center reference points, and expands the preset extraction radius to the left and right in both directions to generate the historical context window and the target context window.

[0082] The preset truncation radius is set to an integer, and its value range is a positive integer between [1, 3]. The preset truncation radius represents the number of adjacent environment characters referenced by the business meaning algorithm when evaluating the mutation of a single character. The data structure inside the history context window and the target context window is a one-dimensional character array. The truncation logic formula is expressed as follows: ; ; in, Represents the target context window; Represents the historical context window; This represents a one-dimensional ordered array composed of discrete characters; Represents the absolute coordinates of the current mutated word within the target query text; Represents the absolute position coordinates of the original historical words in the candidate text; This represents the preset cutoff radius; The coordinates in the target query text are The current variant words; and These represent the surrounding context characters in the target query text that extend to the left and right to the truncation boundary, respectively. The coordinates in the candidate text are Historical original words; and These represent the surrounding context characters in the candidate text that extend to the left and right to the truncation boundary, respectively.

[0083] When performing a bidirectional extended truncation operation, if the index value after superimposing the absolute position coordinates with the preset truncation radius exceeds the physical beginning and end boundaries of the target query text or candidate text, the context truncation component triggers a boundary placeholder completion mechanism. Specifically, the context truncation component fills the out-of-bounds position with a system-preset empty character identifier, which is mapped to a padding vector with a value of 0 at the computer's underlying layer. Through the boundary placeholder completion mechanism, the associated calculation module helps maintain a fixed character length dimension between the historical context window and the target context window, with the fixed character length dimension valued at 2k+1.

[0084] After generating the historical context window and the target context window, the association calculation module calls the feature encoding component to perform vectorized mapping on the discrete characters in the historical context window and the target context window. In order to convert natural language into a tensor format that can be used for similarity inference, the feature encoding component integrates a static word vector lookup table. The static word vector lookup table is pre-trained on a massive unlabeled corpus of ancient books using a continuous bag-of-words model architecture. The association calculation module matches each character in the historical context window and the target context window in the static word vector lookup table and outputs the corresponding historical context feature matrix and target context feature matrix.

[0085] When matching the static word vector lookup table, if rare or uncommon characters not included in the lookup table exist in either the historical or target context window, the feature encoding component will trigger an unknown word degradation strategy. Specifically, the feature encoding component extracts a randomly initialized vector corresponding to a globally unknown identifier pre-defined by the system to replace the rare or uncommon characters, thus avoiding interruptions and errors during the vectorization mapping process. Through the static word vector lookup table, each discrete character is mapped to a continuous real-valued word vector of dimension 128. The dimension of the word vector ranges from 128 to 512, representing positive integers. Therefore, the dimension of the historical context feature matrix and the target context feature matrix output by the feature encoding component is unified to 128. ;in, To preset the cutoff radius, represents the dimension of the continuous real word vector after mapping a single character.

[0086] The association calculation module encapsulates the basic text similarity, historical context feature matrix, and target context feature matrix into structured alignment tuples. These structured alignment tuples are serialized and stored in the underlying memory system in key-value pair format. By constructing and storing these structured alignment tuples, the association calculation module transforms discrete character substitution and variation phenomena into a dense matrix representation containing contextual features. The structured alignment tuples are then pushed into the calculation queue within the association calculation module as direct input tensors for subsequent adaptive weight allocation and joint inference of pronunciation and semantics. The structured alignment tuples are used to assist in correcting the basic text similarity, and a decay penalty is applied when the feature matrix distance is too large.

[0087] Regarding the pre-training process of the continuous bag-of-words model using a static word vector lookup table, the system collects a massive collection of unlabeled ancient Chinese texts as training samples. The preprocessing logic uses a sliding window to extract context word sequences from the ancient text collection as input data for the model, and the center word at the center of the sliding window is used as the corresponding true label. The continuous bag-of-words model consists of an input layer, a projective hidden layer, and an output layer. The input layer receives the initial feature vectors of the context word sequences, the projective hidden layer performs average pooling on the initial feature vectors, and the output layer uses a normalized exponential function to calculate the predicted probability distribution of each word in the vocabulary as the center word. During the training phase, a multi-class cross-entropy loss function is used to calculate the error between the predicted probability distribution and the true label. Backpropagation combined with negative sampling is used to update the word vector weight parameters in the lookup table, continuously optimizing the continuous bag-of-words model's ability to represent the semantic features of single characters in Classical Chinese.

[0088] After determining the genre attribute of the target query text, the receiving and parsing module controls the subsequent data flow by outputting a Boolean routing flag.

[0089] The phonology verification module and the orthogonalized phonology matrix verification workflow are triggered based on the condition that the Boolean routing flag is equal to 1. When the receiving and parsing module determines that the target query text belongs to non-metrical texts such as prose, historical biographies, or the works of various philosophers, the Boolean routing flag output by the receiving and parsing module is equal to 0. For the processing branch where the Boolean routing flag is equal to 0, the system triggers the non-metrical text compatibility processing mechanism. In order to avoid the error caused by the computer's underlying layer forcibly performing matrix bit operations on plain text without tonal constraints, resulting in a mismatch between computational resource consumption and dimension, the system controls the underlying data flow to directly bypass the phonology verification module. The target query text and the corresponding numerical timestamp are directly transmitted to the reverse retrieval module and the association calculation module, entering the pure semantic and glyph dimension tracing and deduction link.

[0090] In the absence of a total baseline of metrical errors as a benchmark for subsequent weight derivation, the association calculation module includes a semantic density evaluation component. This component is designed to provide confidence weights for similarity matching of text without metrical constraints by measuring the information richness of the target query text. The semantic density evaluation component calculates the semantic density index of the target query text using the following formula: ; in, Represents semantic density index; This represents the total number of low-frequency content words and proper nouns contained in the target query text; Represents the total length of the target query text; This represents a smoothing factor to prevent the denominator from being zero.

[0091] To accurately extract the total number of low-frequency content words and proper nouns, the semantic density evaluation component internally calls the entity recognition interface. This interface uses a conditional random field model to perform part-of-speech tagging and entity extraction on the target query text. The semantic density evaluation component only counts words tagged as nouns, verbs, and place names / personal names, and then accumulates these counts to obtain the accurate total. When calculating the semantic density index, if the target query text is too short, resulting in a total character length of 0 or less than the system's minimum truncation length, the underlying division operation will trigger a fatal exception where the denominator is zero. To remedy this dead zone in the algorithm, the semantic density evaluation component introduces a smoothing factor, set to a constant of 1. The semantic density index ranges from 0 to 1, representing a continuous floating-point number. The business meaning characterizes the proportion of core semantic information in non-metrical text; a higher value indicates a denser concentration of proprietary historical referential information in the text.

[0092] To address the issue of limited matching dimensions in non-metrical texts due to the lack of structured phonological features, the association calculation module utilizes an internally integrated global semantic encoding neural network model. This model employs a bidirectional long short-term memory (LSTM) network architecture and includes an input layer, a forward hidden layer, a backward hidden layer, and an attention convergence layer. The preprocessing logic transforms the target query text into a sequence of basic character vectors, which is then input to the input layer. Each character vector in this sequence has a dimension ranging from 256 to 512, representing a positive integer and thus carrying rich character features. The forward and backward hidden layers process the basic character vector sequences in forward and backward word order, respectively, extracting hidden state vectors containing long-distance contextual dependencies. The attention convergence layer performs a weighted summation of all hidden state vectors, outputting a global semantic feature tensor of the target query text. This global semantic feature tensor, in business terms, represents the high-dimensional semantic aggregation state of the non-metrical text within the current spatiotemporal semantic slice database context, reflecting the overall historical narrative intent of the long prose piece.

[0093] Regarding the construction and training process of the global semantic encoding neural network model, the system collects paragraph pairing data from non-metrical ancient texts from various dynasties as training samples. The training samples are automatically constructed by the system based on historical citation relationships between documents, including the base text, positive sample segments that have citation relationships with the base text, and negative sample segments unrelated to the base text. The supervision label for the training samples is a Boolean-type citation relationship identifier. During the training phase, the triplet edge boundary loss function is used to calculate the error value of the network output. The formula for the triplet edge boundary loss function is: ; in, The output error represents the edge boundary loss function of the triplet; The feature vector representing the baseline text; The feature vector representing a positive sample segment; The feature vector representing a negative sample segment; This represents the measure function for calculating the Euclidean distance between two vectors; This represents the preset boundary threshold parameter; This represents the function that takes the maximum value, with preset boundary threshold parameters. The value of is a continuous floating-point number between [0.2, 0.5]. The model uses the backpropagation algorithm to calculate the gradient and uses a stochastic gradient descent optimizer to update the weight parameters of the bidirectional long short-term memory network. This joint training mechanism helps to narrow the vector distance between the benchmark text and the positive samples and widen the vector distance between the benchmark text and the negative samples, continuously optimizing the global semantic coding neural network model's ability to capture features of loosely structured non-metrical ancient texts.

[0094] The association calculation module triggers the reverse hypothesis deduction workflow based on the difference character pairs issued by the reverse retrieval module and the target query text.

[0095] In order to reproduce the phenomena of adaptation and alteration in the creation of ancient literature from the underlying data structure, the association calculation module is equipped with a text reconstruction component. The text reconstruction component locates the absolute position coordinates of the current variant word in the character sequence of the target query text. The association calculation module extracts the historical original words in the difference character pairs, and uses the historical original words to replace the current variant word at the corresponding absolute position coordinate in the target query text, generating the hypothetical restoration text.

[0096] During character replacement, if the character length of the original historical word is inconsistent with the character length of the current mutated word (e.g., a single word is transformed into a two-word word), the text reconstruction component triggers an index adaptive offset mechanism. The text reconstruction component calculates the length difference between the original historical word and the current mutated word, and uses the length difference to dynamically update the absolute position coordinates of all subsequent characters in the hypothetical restored text. Through the index adaptive offset mechanism, the associated calculation module helps to avoid the system risk of character overwriting errors or memory out-of-bounds errors when the underlying array is reorganized.

[0097] Based on the generated hypothetical restored text, the association calculation module calls the previously mentioned phonological verification module to re-extract the binary pronunciation attribute sequence of the hypothetical restored text. Specifically, the phonological verification module uses the reconstructed complete sentence structure to re-perform dynamic contextual polyphonic character disambiguation on the newly replaced historical original words to obtain the true pronunciation of the characters in the current virtual context, ensuring the rigor of pronunciation deduction. According to the physical layout structure of the hypothetical restored text, the phonological verification module transforms the binary pronunciation attribute sequence into a two-dimensional mathematical matrix form to generate the hypothesis matrix.

[0098] To prevent length changes caused by character replacement from triggering mismatch errors in the dimension tensors of the underlying matrix bitwise operations, the phonology verification module performs a forced dimension alignment operation on the hypothetical matrix based on the fixed column number dimension of the standard metrical matrix. When the number of columns in the hypothetical matrix is ​​greater than that in the standard metrical matrix, the last digit is truncated; when the number of columns in the hypothetical matrix is ​​less than that in the standard metrical matrix, placeholder padding based on null value identifiers is performed to ensure that the hypothetical matrix and the standard metrical matrix have completely consistent row and column dimensions.

[0099] After completing the forced dimension alignment, in order to extract the out-of-rule coordinate nodes of the hypothetical restored text, the logical operation unit configured inside the association calculation module performs Boolean bitwise operations on the hypothesis matrix. The association calculation module performs a bitwise XOR operation on the hypothesis matrix and the standard metrical matrix, and performs a bitwise logical AND operation on the result of the bitwise XOR operation with the mask matrix to generate the hypothesis out-of-rule error matrix, the formula of which is: ; in, The matrix representing the out-of-law error of the assumption; The hypothesis matrix generated by the text transformation represents the hypothesis. This represents the standard metrical matrix built into the system; A mask matrix representing the relaxed rules for marking; Represents the bitwise XOR operator for matrices; This represents the bitwise logical AND operator for matrices.

[0100] After obtaining the hypothesis prosody error matrix, the correlation calculation module sums up all the values ​​of all elements in the hypothesis prosody error matrix to calculate the total number of hypothesis prosody errors in the hypothesis-reconstructed text. The calculation formula is as follows: ; in, This represents the total number of errors in the assumed metrical pattern; Represents the matrix element summation operator, assuming the total number of metrical errors. The data type is a non-negative integer, and the business meaning represents the number of absolute metrical errors contained in the hypothetical restored text in the current virtual context.

[0101] To quantitatively evaluate the effect of historical original words on the restoration of the phonological structure of the target query text, the association calculation module retrieves the total number of basic prosodic errors cached in the early stage by the phonological verification module. The association calculation module calculates the difference between the total number of basic prosodic errors and the total number of assumed prosodic errors, generating a prosodic restoration gain index. The formula for calculating the prosodic restoration gain index is as follows: ; in, Represents the metrical repair gain index. The total number of basic metrical errors representing the target query text. The total number of hypothetical metrical errors in the restored text, and the metrical repair gain exponent Δ. The data type is integer.

[0102] To transform absolute error differences into continuous weight parameters suitable for subsequent integrated decision-making models, the correlation calculation module calculates normalized phonological confidence weights based on the metrical restoration gain index. The phonological confidence weight is equal to the maximum value between the metrical restoration gain index and zero, divided by the ratio of the total number of basic metrical errors to the sum of smoothing factors. The formula for calculating the phonological confidence weight is as follows: ; in, Represents the confidence weight of phonology; This represents the function that takes the maximum value. Smoothing factor represents a smoothing factor to prevent the denominator from being zero; smoothing factor The value of is set to a constant 1. Phonological confidence weight. The value of is a continuous floating-point number between [0, 1]. At the business logic level, if the prosody repair gain index is less than or equal to 0, it indicates that the original historical words did not bring about prosodic repair, and the phonological confidence weight is truncated to 0; if the prosody repair gain index is greater than 0, it indicates that the original historical words are superior to the current variant words in terms of tonal compliance, and the larger the value of the phonological confidence weight, the more sufficient the physical evidence for candidate documents is in the textualization of the target query.

[0103] For text sequence replacement operations, Boolean operation instruction scheduling of underlying matrices, and scalar arithmetic operations, those skilled in the art can implement them using standard central processing unit instruction sets and memory reallocation mechanisms. The related underlying data access and scheduling logic is well-known in the field and will not be elaborated here.

[0104] After completing the extraction of data features and error extrapolation for each independent dimension, and calculating the comprehensive association weight between the target query text and all candidate documents, the association calculation module transfers the underlying data flow to the graph generation module, triggering the graph construction and structured data output workflow.

[0105] The evolution of ancient texts is often not a simple linear inheritance, but may involve multiple sources converging or diverging. To perform global topological aggregation of discrete source text fragments, the association calculation module is equipped with a weight fusion component. This component dynamically and linearly combines various similarity indicators between the target query text and candidate documents based on the Boolean routing flags output by the receiving and parsing module. The association calculation module calculates the comprehensive association weight between the target query text and candidate documents using the following formula: ; in, Represents the overall correlation weight; Represents the basic text similarity; Represents the confidence weight of phonology; Represents semantic density index; Represents a Boolean routing flag; Preset weight hyperparameters representing the basic glyph dimensions; Preset weight hyperparameters representing the phonological dimension; Preset weight hyperparameters representing semantic dimensions. , and The values ​​of are all continuous floating-point numbers between [0, 1].

[0106] When calculating the comprehensive association weight, the system employs strict weight normalization constraints. When the Boolean routing flag is equal to 1, it indicates that the current text is subject to metrical constraints, and this condition is satisfied. + =1; When the Boolean routing flag is equal to 0, it indicates that the current text has no metrical constraints, and at this time it satisfies + =1. Through the dynamic weight allocation mechanism, the correlation calculation module helps to keep the final value of the comprehensive correlation weight always constrained within the closed interval [0, 1], thereby avoiding the system dead zone caused by underlying floating-point overflow or downstream comparison logic failure.

[0107] After obtaining the comprehensive association weights between the target query text and all candidate documents, the graph generation module instantiates the target query text and historically retrieved candidate documents as discrete nodes in the graph, and instantiates the evolution paths with derivative relationships between documents as directed edges in the graph. The graph generation module uses the numerical timestamp attribute of each discrete node to constrain the direction. The connection direction of the directed edges is forcibly defined as pointing from the historical node with the smaller timestamp value to the subsequent node with the larger timestamp value. Since the historical timeline has an objective monotonically increasing physical property, the timestamp-based directed edge connection mechanism eliminates the loops caused by time reversal at the physical topology level, making the graph data structure generated by the graph generation module conform to the strict characteristics of a directed acyclic graph. The weight values ​​of the directed edges are the comprehensive association weights calculated between the corresponding nodes.

[0108] Faced with multiple potential evolution paths within a directed acyclic graph (DAG), selecting the globally optimal solution becomes a key issue. To address this, the graph generation module integrates a path ranking neural network model. This model employs a multilayer perceptron architecture and includes an input layer, two fully connected hidden layers, and an output layer. The input layer receives preprocessed fixed-dimensional path feature vectors. The preprocessing logic traverses the DAG using a depth-first search algorithm, extracting complete unidirectional paths from the furthest historical node to the target query text node. It then sequentially concatenates the equivalent feature tensors of each discrete node along the complete unidirectional path with the corresponding directed edge weights to generate a variable-length path feature matrix.

[0109] To transform the variable-length path feature matrix into a fixed-dimensional tensor compatible with multilayer perceptrons, the preprocessing logic performs adaptive average pooling on the variable-length path feature matrix, inputting the output fixed-dimensional path feature vector of size into the input layer. The size of the vector is a positive integer between [128, 512]. Two fully connected hidden layers use a linear rectified function for non-linear activation to extract long-distance semantic evolution features hidden in multi-hop paths. The output layer uses a logistic regression function to output the confidence probability distribution of the current unidirectional path as a true historical path. The business meaning of the confidence probability distribution represents the rationality of the unidirectional path in the overall verification logic.

[0110] Regarding the construction and training process of the path ranking neural network model, the system collects ancient literary allusion transmission relationship trees manually verified by domain experts as training samples. Positive sample labels are defined as time-series node paths in historical documents that have actually been cited or adapted, while negative sample labels are defined as random node splicing paths that are legal on the timeline but do not have actual adaptation relationships. During the training phase, the binary cross-entropy loss function is used to calculate the error between the confidence probability distribution and the positive and negative sample labels. The path ranking neural network model uses the backpropagation algorithm to calculate the gradient and adopts an adaptive moment estimation optimizer to update the weight parameters of the fully connected hidden layers, continuously optimizing the path ranking neural network model's ability to identify and rank complex multi-hop tracing paths.

[0111] After outputting the confidence probability distribution, the graph generation module sorts all complete unidirectional paths in the directed acyclic graph (DAG) in descending order based on the magnitude of the confidence probability distribution values. The graph generation module extracts the top N path data and the complete topological structure information of the DAG, and encapsulates them into a multidimensional structured output file. The positive integer N ranges from [1, 5]. The multidimensional structured output file adopts a lightweight data exchange format. The multidimensional structured output file contains a set of node objects and a set of edge objects. The set of node objects clearly records the text fragments, numerical timestamps, and genre attributes of each discrete node; the set of edge objects clearly records the difference character pairs, comprehensive association weights, and evolution directions between discrete nodes. The multidimensional structured output file is finally transmitted to the external business terminal via the system bus, providing end users with a transparent and traceable spatiotemporal evolution topology network of ancient literary allusions.

[0112] For the specific pathfinding logic of the depth-first search algorithm, the forward propagation matrix multiplication operation of the multilayer perceptron, and the serialization encoding process of the lightweight data exchange format, those skilled in the art can refer to the standard graph theory algorithm library and system input / output protocol for implementation. The relevant underlying algorithm implementations are well-known technologies in this field and will not be elaborated here.

[0113] Specific application examples: To further verify the effectiveness of the intelligent question - answering system proposed in this invention for ancient literature teaching in solving problems such as cross - era semantic drift, computational ambiguity induced by the evolution of polyphonic pronunciation, and inversion errors prone to occur in multi - hop traceability in the context of tracing the origin of ancient literature allusions, this embodiment constructs an application scenario containing five hundred thousand volumes of ancient book literature data and elaborates in conjunction with the data Figure 3 shown in Figure 4 the appendix.

[0114] In the application scenario of this embodiment, the intelligent question - answering system for ancient literature teaching traces the origin of the allusion for the target query text "The setting sun and the solitary wild duck fly together; the autumn waters merge with the vast sky in one color". First, the intelligent question - answering system for ancient literature teaching uses the receiving and parsing module to analyze the text sequence structure and selects the author Wang Bo and the work name "Preface to the Tengwang Pavilion" from the pre - installed literary knowledge graph in the system. The system operating environment is configured as follows: There is a graph database configured at the bottom layer of the data isolation module deployed in the cloud data center. The graph database stores the lexical interpretation data and pronunciation attribute data of past dynasties, and both the lexical interpretation data and pronunciation attribute data of past dynasties are marked with time - dimension identifiers; the receiving and parsing module is configured with a deep - learning model architecture of a bidirectional long - short - term memory network connected to a conditional random field, which can perform basic word segmentation and词性标注 (pos tagging) on the input natural language; the cloud system bus is configured to respond to the query feature data packets transmitted from the front - end at the millisecond level.

[0115] Implementation of spatio - temporal semantic isolation and genre routing: Regarding the problems of age inversion and interference from later dictionary annotations commonly existing in ancient book question - answering, the intelligent question - answering system for ancient literature teaching first extracts and maps the creation dynasty time of Wang Bo through the receiving and parsing module into a numerical timestamp of 675 (representing 675 AD).

[0116] During the operation of the intelligent question - answering system for ancient literature teaching, the receiving and parsing module synchronously determines that the genre attribute of the target query text is a ci - fu (rhapsody), meeting the set metrical requirements. The receiving and parsing module then generates a response, assigns the boolean routing flag bit as 1, and thus activates the downstream phonetic verification module.

[0117] After receiving the numerical timestamp of 675, the data isolation module uses the view isolation component to dynamically generate a spatio - temporal semantic slice database in the underlying graph database. The view isolation component forcibly filters the literature annotations with historical time attributes greater than 675. The data isolation module constructs a data access barrier through the forced filtering mechanism, eliminating the temporal error caused by cross - era dictionary data for the semantic deduction of the target query text.

[0118] Implementation of reverse retrieval and phonetic self - adaptation association: After the dynamic loading of the spatio-temporal semantic slice database is completed, the intelligent question-answering system for ancient literature teaching uses the feature words contained in the target query text to capture potential historical allusion documents. The reverse retrieval module first identifies the feature words and performs a retrieval within a limited time axis range.

[0119] The intelligent question-answering system for ancient literature teaching sets the preset word frequency threshold of the word frequency and inverse document frequency algorithm to 0.3. Within the time constraint range (<675), the reverse retrieval module successfully recalls the candidate text "The falling flowers fly with the mushroom canopies, and the willows share the same color with the spring flags" created by Yu Xin during the Southern and Northern Dynasties (the creation timestamp corresponds to around 504 AD) from the spatio-temporal semantic slice database.

[0120] The reverse retrieval module uses the dynamic programming algorithm to extract the current variant words (rosy clouds, solitary wild duck, autumn waters, long sky) in the target query text and the historical original words (flowers, mushroom canopies, willows, spring flags) in the candidate text to form a differential character pair. The phonetic verification module calls the pronunciation attribute data in the spatio-temporal semantic slice database, converts the target query text into a target matrix, and calculates the total number of basic metrical errors through bit operations.

[0121] Combined with the hypothetical text generated by the association calculation module, the phonetic verification module re-performs matrix conversion and Boolean operations on the hypothetical text to obtain the total number of hypothetical errors. The association calculation module uses the formula to invert the difference between the total number of basic metrical errors and the total number of hypothetical errors. The result shows that the historical original words repair the potential out-of-metre nodes caused by the current variant words. The association calculation module combines the basic text similarity to assign weight parameters and calculates that the comprehensive association weight between the target query text and the candidate text is as high as 0.89.

[0122] Verification of the multi-dimensional fusion traceability performance: Based on the calculated comprehensive association weight parameters, the graph generation module establishes the candidate text as the origin node of the allusion, constructs a directed acyclic graph and integrates and outputs structured parsing data.

[0123] To visually present the correction accuracy and operation efficiency, this embodiment introduces a conventional deep semantic question-answering system (using a pure text tensor global retrieval method) without a configured time isolation mechanism and a phonetic verification module as a comparison reference.

[0124] Refer to Appendix Figure 3 ,Appendix Figure 3Comparative results show that although conventional deep semantic question answering systems introduce a global vector encoding model, the lack of mandatory time axis constraints results in a cross-era false alarm rate as high as 18.0%. This leads to a high risk of misjudging Ming and Qing dynasty references as sources of Tang dynasty poetry when faced with a large number of ancient book citations, seriously affecting the historical logic compliance of the output results. The method of this invention, based on an intelligent question answering system for ancient literature teaching, improves the source tracing accuracy to 94.5% through the dual constraints of the homophony verification module of the view isolation component, while precisely reducing the cross-era false alarm rate to 0.0%, achieving non-misleading source tracing in the logical dimension.

[0125] See attached document Figure 4 When the total amount of ancient literature data at the bottom layer expands to 500,000 volumes, conventional deep semantic question answering systems experience an almost linear increase in system computational load due to the full tensor comparison performed in the global graph database, with response time soaring to 4.8 seconds. In the method of this invention, the data isolation module uses numerical timestamps to dynamically lock a small range of spatiotemporal semantic slice databases, transforming the massive global retrieval into lightweight computation within local time slices. The two curves show a significant difference at the coordinate point of 500,000 volumes of data, and the response time of this invention is stably maintained at around 1.2 seconds.

[0126] Summary of application examples: This application example verifies the effectiveness of an intelligent question-and-answer system based on ancient literature teaching in tracing the origins of massive amounts of ancient texts.

[0127] In terms of data filtering, the data isolation module uses the numerical timestamp of the target query text to build an access barrier, successfully isolating massive amounts of redundant and interfering data from later generations, laying a deterministic spatiotemporal semantic foundation for matrix deduction in downstream modules.

[0128] In terms of source tracing computation, an orthogonalized phonological matrix is ​​constructed through a phonological verification module and bitwise operations are performed. Based on the intelligent question-answering system for ancient literature teaching, the system successfully uncovers the ancient prosody inheritance relationship hidden behind discrete word variations by utilizing the deductive logic of hypothetical texts. Compared with conventional pure semantic search methods, the intelligent question-answering system for ancient literature teaching keeps the response latency under high-concurrency retrieval conditions at an extremely low level and significantly improves the accuracy of association, solving the core business problem of the difficulty in quantifying and tracing the complex historical evolution relationship of long texts.

Claims

1. An intelligent question-and-answer system for teaching ancient literature, characterized in that, include: The receiving and parsing module is used to obtain the target query text and extract the creation dynasty timestamp and genre attribute, and generate a boolean routing flag based on the genre attribute; The data isolation module is used to receive the numerical timestamp corresponding to the creation dynasty timestamp, store the pronunciation attribute data with historical time attributes, set the historical time attribute being less than or equal to the numerical timestamp as a forced filtering condition, and dynamically generate a spatiotemporal semantic slice database. The phonetic verification module is used to call the spatiotemporal semantic slice database to convert the target query text into a target matrix when the Boolean routing flag is in the enabled state. Based on the genre attribute, it reads the corresponding standard metrical matrix and mask matrix from the system's built-in rule base, performs a bitwise XOR operation on the target matrix and the standard metrical matrix, and performs a bitwise logical AND operation on the result of the bitwise XOR operation with the mask matrix to obtain and output the total number of basic metrical errors. The reverse retrieval module is used to extract candidate text from the spatiotemporal semantic slice database and output it, using the creation dynasty timestamp as a time filtering parameter. The association calculation module is used to calculate the comprehensive association weight between the target query text and the candidate text by combining the total number of basic metrical errors; The graph generation module is used to construct a topological path based on the comprehensive association weight, generate graph structure data with nodes and connections, and output the system parsing results.

2. The intelligent question-and-answer system for teaching ancient literature according to claim 1, characterized in that, The receiving and parsing module includes a text preprocessing component, which adopts a deep learning model architecture of bidirectional long short-term memory network connected to conditional random field to convert the target query text into a word vector sequence and output word segmentation and part-of-speech tagging sequence. The receiving and parsing module extracts the author's name or work title from the target query text, compares and extracts historical data in the system's preset literary knowledge graph, maps the historical data into a numerical timestamp and outputs it as the timestamp of the creation dynasty. The receiving and parsing module determines the genre attribute of the target query text based on the classification tags returned by the literary common sense graph, or by statistically analyzing the number of words, sentence symmetry, and rhyme distribution characteristics of the target query text when the graph classification tags are missing. The receiving and parsing module inputs the extracted genre attributes into the genre classification mapping table pre-built by the system for Boolean logic judgment and generates the Boolean routing flag.

3. The intelligent question-and-answer system for teaching ancient literature according to claim 1, characterized in that, The spatiotemporal semantic slice database is equipped with a weighted priority disambiguation strategy. When loading the spatiotemporal semantic slice database, the data isolation module calculates the authority score of the document through the built-in document scoring neural network model and loads the data record with the highest authority score into the active cache area of ​​the spatiotemporal semantic slice database for the phonological verification module to call. The document scoring neural network model adopts a multilayer perceptron architecture. The input data includes the time difference between the document's publication date and the present, the number of times the document has been cited in subsequent academic papers, and the document's collation version level.

4. The intelligent question-and-answer system for teaching ancient literature according to claim 2, characterized in that, The phonology verification module extracts the glyph codes of single characters in the word segmentation and part-of-speech tagging sequence, combines the word group text where the single character is located with the part-of-speech tag to generate a joint query primary key, and obtains the pronunciation attribute data corresponding to the joint query primary key from the spatiotemporal semantic slice database; If the pronunciation attribute data is not obtained, the phonology verification module triggers the semantic disambiguation neural network model. The semantic disambiguation neural network model adopts a Transformer encoder architecture based on a self-attention mechanism, outputs the context dynamic feature vector of the single character, and calculates the cosine similarity between the context dynamic feature vector and the baseline feature vector of the historical example sentence to obtain the pronunciation attribute data.

5. The intelligent question-and-answer system for teaching ancient literature according to claim 1, characterized in that, The reverse retrieval module uses term frequency and inverse document frequency algorithms to extract words in the target query text whose weight distribution values ​​exceed a preset term frequency threshold as specified feature words, and obtains a candidate text set containing the specified feature words; The reverse retrieval module uses a dynamic programming algorithm to construct an edit distance matrix, calculates the minimum number of single-character editing operations required to convert the candidate text into the target query text, and uses the minimum number of single-character editing operations as the absolute edit distance. The absolute edit distance is used to calculate the basic text similarity, and the difference character pairs between the target query text and each candidate text in the candidate text set are extracted and output to the association calculation module.

6. The intelligent question-and-answer system for teaching ancient literature according to claim 5, characterized in that, The association calculation module receives the difference character pair, extracts the historical original words and the current variant words from the difference character pair, and expands the preset truncation radius to the left and right to generate a historical context window and a target context window. The association calculation module calls the static word vector lookup table pre-trained on the unlabeled corpus of ancient books using the continuous bag-of-words model architecture, and maps the characters in the historical context window and the target context window into historical context feature matrices and target context feature matrices.

7. The intelligent question-and-answer system for teaching ancient literature according to claim 6, characterized in that, The association calculation module replaces the corresponding current variant words in the target query text with the historical original words to generate hypothetical restored text and sends it to the phonology verification module. The phonology verification module receives the hypothesis reconstruction text, converts the hypothesis reconstruction text into a hypothesis matrix, and returns it to the association calculation module; The logical operation unit configured inside the correlation calculation module performs Boolean bitwise operations on the hypothesis matrix, performs a bitwise XOR operation on the hypothesis matrix and the standard metrical matrix, and performs a bitwise logical AND operation on the result of the bitwise XOR operation with the mask matrix to obtain the total number of hypothesis metrical errors. The association calculation module calculates the difference between the total number of basic metrical errors and the total number of assumed metrical errors to generate a metrical repair gain index. The metrical repair gain index is used to calculate the phonological confidence weight. The association calculation module uses the phonological confidence weight to calculate the comprehensive association weight between the target query text and the candidate text.

8. The intelligent question-and-answer system for teaching ancient literature according to claim 1, characterized in that, When the Boolean routing flag is off, the target query text bypasses the phonetic verification module and is transmitted to the reverse retrieval module and the association calculation module; The association calculation module is equipped with a semantic density evaluation component. The semantic density evaluation component calls the entity recognition interface to extract the total number of low-frequency content words and proper nouns in the target query text, and calculates the semantic density index by combining the total character length of the target query text and the smoothing factor. The association calculation module calls a global semantic encoding neural network model with a bidirectional long short-term memory network architecture to output the global semantic feature tensor of the target query text, which is used to reflect the overall historical narrative intent of the long essay.

9. The intelligent question-and-answer system for teaching ancient literature according to claim 1, characterized in that, It also includes a graph generation module, which instantiates the target query text and candidate documents hit in the historical search as discrete nodes in the graph, and uses the numerical timestamp attribute of the discrete nodes to construct a directed acyclic graph for directional constraints. The graph generation module integrates a path ranking neural network model with a multilayer perceptron architecture. It traverses the directed acyclic graph to extract complete unidirectional paths, and concatenates the equivalent feature tensors of the discrete nodes on the complete unidirectional paths with the comprehensive association weights to generate a variable-length path feature matrix. After performing an adaptive average pooling operation on the path feature matrix, it is input into the path ranking neural network model, outputs a confidence probability distribution, and sorts the results in descending order according to the confidence probability distribution. The results are then packaged into a multidimensional structured output file.

10. An intelligent question-and-answer method for teaching ancient literature, characterized in that: An intelligent question-and-answer system for teaching ancient literature, as described in any one of claims 1-9, comprises the following steps: Obtain the target query text and extract the creation dynasty timestamp and genre attribute. Generate a Boolean routing flag based on the genre attribute and output the creation dynasty timestamp and the Boolean routing flag. Receive the numerical timestamp corresponding to the creation dynasty timestamp, and based on the stored pronunciation attribute data with historical time attributes, set the historical time attribute being less than or equal to the numerical timestamp as a mandatory filtering condition, and dynamically generate a spatiotemporal semantic slice database. When the received Boolean routing flag is in the enabled state, the spatiotemporal semantic slice database is invoked to convert the target query text into a target matrix. Based on the genre attribute, the corresponding standard metrical matrix and mask matrix are read from the system's built-in rule base. The target matrix and the standard metrical matrix are XORed, and the result of the XOR operation is ANDed with the mask matrix to obtain the total number of basic metrical errors and output it. Using the creation dynasty timestamp as a time filtering parameter, candidate texts are extracted and output from the spatiotemporal semantic slice database; Receive the total number of basic metrical errors and the candidate text, and calculate the comprehensive association weight between the target query text and the candidate text based on the total number of basic metrical errors; Based on the calculated comprehensive association weights, a topological path is constructed, generating graph structure data with nodes and connections, and the system parsing results are output.

Images