AI auxiliary reading method and system based on dynamic partitioning knowledge classification self-verification

Abstract

The invention provides an AI auxiliary reading method and system based on dynamic partitioning knowledge classification self-verification, and relates to the technical field of artificial intelligence, the method comprises the following steps: carrying out dynamic semantic partitioning on a super-long text to obtain a plurality of initial text units with consistent internal semantics; performing topic model construction according to the initial text unit and a layered Dirichlet process to obtain a topic distribution tag; receiving a user query request, and extracting a retrieval text set according to the hierarchical retrieval strategy and a retrieval text corresponding to the user query request; inputting the constraint cue word extracted by the topic distribution tag corresponding to the retrieval text set, the retrieval text set and the retrieval text into a text generation model to output a to-be-calibrated text; if the error scoring function of the to-be-calibrated text and any initial text unit in the retrieval text set is smaller than a preset error threshold value, the to-be-calibrated text serves as a credible text to be output to a user. According to the invention, the credibility and accuracy of the answer can be enhanced.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an AI-assisted reading method and system based on dynamic block knowledge classification and self-verification. Background Art

[0002] With the rapid development of the internet and the ubiquity of mobile devices, digital reading has become an essential way for people to obtain information. China's high regard for the cultural industry has created a favorable environment for the development of the digital reading industry. The government has introduced a series of policies to encourage innovation in digital reading platforms and promote the upgrading of the digital publishing industry. Simultaneously, with the advancement of artificial intelligence (AI) technology, AI has become a key force driving industrial upgrading and economic transformation, demonstrating enormous potential for improving reading efficiency, content personalization, and social inclusion. Internationally, technology giants such as Google and Meta have launched intelligent reading assistants based on large models (such as DeepRead and MetaScholar), supporting features such as long-text summarization and multimodal interaction.

[0003] A Chinese patent with announcement number CN118394897A discloses a multi-document intelligent question-answering method and system based on a large language model, the method comprising: obtaining the content of various documents predicted by demand analysis; converting the content of the various documents into text blocks, and segmenting the text blocks to obtain segmented text blocks; applying vector retrieval to the multiple text blocks, and applying keyword retrieval to the segmented text blocks to obtain multiple target text blocks related to user questions; and using a large language model to generate answers to the user questions based on the multiple target text blocks, the user questions, and predefined prompt instruction templates. However, the retrieval strategy of the above-mentioned solution is single, only using single-layer vector retrieval and keyword matching, and no topic model is constructed, lacking domain-customized guidance, resulting in biased content and insufficient answer accuracy. Therefore, it is very necessary to provide an AI-assisted reading method and system based on dynamic block knowledge classification and self-verification to enhance the credibility and accuracy of answers. Summary of the Invention

[0004] In light of this, this paper proposes an AI-assisted reading method and system based on dynamic block knowledge classification and self-verification. By automatically segmenting extremely long texts according to semantic boundaries, the thematic unity and information coherence within each initial unit are ensured. Furthermore, a hierarchical Dirichlet process is introduced to model the topic distribution of each unit, enabling the automatic generation of topic labels and unit grading. This method effectively filters out possible deviations or hallucinatory outputs from the model, thereby enhancing the credibility and accuracy of answers.

[0005] The present invention provides an AI-assisted reading method based on dynamic block knowledge classification and self-verification, the method comprising:

[0006] Dynamically segment long text into semantic chunks to obtain multiple initial text units with consistent internal semantics;

[0007] Constructing a topic model based on the initial text units and a hierarchical Dirichlet process to obtain topic distribution labels corresponding to the initial text units, and performing unit grading on each initial text unit;

[0008] receiving a user query request, and extracting a search text set according to a hierarchical search strategy and a search text corresponding to the user query request, wherein the search text set includes a plurality of initial text units related to the search text;

[0009] Inputting the constraint prompt words extracted from the topic distribution label corresponding to the search text set, the search text set, and the search text corresponding to the user query request into a text generation model to output a text to be calibrated;

[0010] Based on the self-verification mechanism, the text to be calibrated is dynamically verified with each initial text unit in the search text set. If the error scoring function of the text to be calibrated and any initial text unit in the search text set is less than a preset error threshold, the text to be calibrated is output to the user as a credible text.

[0011] Based on the above technical solution, preferably, the dynamic semantic segmentation of the super-long text to obtain multiple initial text units with consistent internal semantics specifically includes:

[0012] Extracting block boundaries in the super-long text according to a logical boundary detection function, and dividing the super-long text into blocks one by one according to the block boundaries to obtain a plurality of initial text units;

[0013] A semantic consistency judgment is performed on each initial text unit based on a semantic coherence scoring function. If the semantic coherence score corresponding to the initial text unit is less than a preset semantic score, the initial text unit is segmented again until the semantic coherence score corresponding to each initial text unit in the super-long text is greater than or equal to the preset semantic score, and then a plurality of internally semantically consistent initial text units are output.

[0014] On the basis of the above technical solutions, preferably, a multi-level vectorized index at the chapter level, paragraph level and sentence level is written for each initial text, a chapter-level index is constructed based on the vector representation of BERT-Whitening post-processing, a paragraph-level index is constructed based on the HNSW algorithm, and a sentence-level index is supported based on the FAISS Flat index.

[0015] More preferably, the hierarchical retrieval strategy includes performing a preliminary screening based on the query vector and the chapter-level index, and then refining the candidate chapters obtained from the preliminary screening based on the paragraph index and / or sentence-level index, and finally returning the completed retrieval results to the corresponding initial text unit.

[0016] More preferably, the constraint prompt words extracted according to the topic distribution labels specifically include:

[0017] Predefine a plurality of knowledge categories related to auxiliary reading, wherein each knowledge category corresponds to a knowledge category feature vector, and the knowledge category feature vector includes a static initialization part and a dynamic update part;

[0018] The embedding representation of the topic distribution label corresponding to each initial text unit in the retrieved text set is sequentially compared with all knowledge category feature vectors through an attention mechanism, and the text semantic features related to the embedding representation and various knowledge features are respectively extracted, wherein the embedding representation includes token-level embedding, whole sentence-level embedding, and paragraph-level embedding;

[0019] The text semantic features are dynamically fused with the corresponding knowledge category feature vectors to obtain the constraint prompt words corresponding to the topic distribution labels.

[0020] More preferably, the method further comprises:

[0021] Performing topic clustering on text content based on a hierarchical Dirichlet process to generate a topic-content mapping library, and automatically associating different reading topics found in the user's reading history data with topics in the topic-content mapping library;

[0022] Generate dynamic interest tags based on the user's reading history data, display the subject interest weights to the user in a visual panel, and receive manual adjustment instructions from the user.

[0023] More preferably, the text generation model provides a multimodal interaction interface, which is used to perform one or more of voice command parsing, visual focus prompts, and customized text-to-speech synthesis.

[0024] In a second aspect of the present application, an AI-assisted reading system based on dynamic block knowledge classification and self-verification is provided, wherein the AI-assisted reading system includes a text acquisition module, a retrieval constraint module, and a text output module, wherein:

[0025] The text collection module is used to dynamically segment the super-long text into semantic blocks to obtain multiple initial text units with consistent internal semantics;

[0026] The retrieval constraint module is used to construct a topic model based on the initial text unit and the hierarchical Dirichlet process to obtain a topic distribution label corresponding to the initial text unit, and perform unit classification on each initial text unit, receive a user query request, and extract a retrieval text set based on a hierarchical retrieval strategy and the retrieval text corresponding to the user query request, wherein the retrieval text set includes a plurality of initial text units related to the retrieval text, input the constraint prompt words extracted from the topic distribution label corresponding to the retrieval text set, the retrieval text set, and the retrieval text corresponding to the user query request into a text generation model to output a text to be calibrated;

[0027] The text output module is used to dynamically verify the text to be calibrated and each initial text unit in the search text set based on a self-verification mechanism. If the error scoring function of the text to be calibrated and any initial text unit in the search text set is less than a preset error threshold, the text to be calibrated is output to the user as a credible text.

[0028] In a third aspect of the present application, an electronic device is provided, comprising a processor, a memory, a user interface and a network interface, wherein the memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory.

[0029] In a fourth aspect of the present application, a computer-readable storage medium is provided, on which a computer program is stored, and the computer program is executed by a processor to implement the steps of an AI-assisted reading method based on dynamic block knowledge classification and self-verification.

[0030] The AI-assisted reading method and system based on dynamic block knowledge classification and self-verification provided by the present invention has the following beneficial effects compared with the existing technology:

[0031] (1) By automatically segmenting ultra-long texts according to semantic boundaries, the unified theme and coherent information within each initial unit are guaranteed, and the context fragmentation caused by arbitrary equal-length segmentation is avoided. The block granularity can be adaptively adjusted according to the text structure and topic complexity, thereby improving the efficiency and accuracy of subsequent processing. In addition, a hierarchical Dirichlet process is introduced to model the topic distribution of each unit, realizing the automatic generation of topic labels and unit classification. The multi-level topic representation takes into account both the global topic framework and local details, providing richer and more discriminative semantic constraints for the retrieval and generation stages. At the same time, it can also search from top to bottom or bottom to top at the topic level according to user queries, giving priority to locating the most relevant units, significantly improving the text recall rate and guidance accuracy, and dynamically comparing and scoring the calibrated text with each retrieval unit, which can effectively filter out possible deviations or hallucinations of the model, thereby enhancing the credibility and accuracy of the answer.

[0032] (2) By simultaneously modeling the embedding of tokens, whole sentences, and paragraphs, we can take into account both details and global context, avoid information loss or noise interference caused by a single granularity, and use the attention mechanism to compare the embedding of topic distribution labels with the feature vectors of each knowledge category, dynamically calculate the matching weight, and ensure that the prompt word strictly fits the topic semantics of the current retrieval text. The static initialization part provides domain or general knowledge support, and the dynamic update part allows the model to be tuned in real time with the retrieval text and user interaction, effectively taking into account reliability and adaptability, and fusing the extracted text semantic features with the corresponding knowledge category vector at the vector level, so that the final prompt word retains the details of the retrieval context and incorporates the abstract concept of category knowledge. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0034] Figure 1 A flowchart of an AI-assisted reading method based on dynamic block knowledge classification and self-verification provided by the present invention;

[0035] Figure 2 This is a schematic diagram of the structure of the AI-assisted reading system provided by the present invention;

[0036] Figure 3 This is a schematic structural diagram of the electronic device provided by the present invention.

[0037] Explanation of the accompanying drawings: 1. AI-assisted reading system; 11. Text acquisition module; 12. Retrieval constraint module; 13. Text output module; 2. Electronic device; 21. Processor; 22. Communication bus; 23. User interface; 24. Network interface; 25. Memory. DETAILED DESCRIPTION

[0038] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0039] The present invention discloses an AI-assisted reading method based on dynamic block knowledge classification self-verification, referring to Figure 1 , the steps of the method include S1 to S5.

[0040] Step S1: Dynamically semantically segment an overlong text to obtain a plurality of initial text units with consistent internal semantics.

[0041] This step also includes steps S11 to S12.

[0042] Step S11 , extracting block boundaries in the super-long text according to a logical boundary detection function, and dividing the super-long text into blocks one by one according to the block boundaries to obtain a plurality of initial text units.

[0043] In this step, logical boundary detection and semantic coherence scoring are combined to split the extremely long text into logically self-consistent independent units to avoid semantic breaks caused by mechanical cutting, and the block boundaries are retroactively corrected according to the content of subsequent chapters (such as merging foreshadowing and echo paragraphs).

[0044] Furthermore, the logical boundary detection function is used to identify potential block boundaries in the text (such as chapter titles and paragraph transition words). The expression of the logical boundary detection function is:

[0045]

[0046] Here, α represents the weight of explicit structural features. Explicit structures (such as titles, serial numbers, and paragraph separators) are strong segmentation signals. Experiments show that their accuracy is >85%. Text materials in digital reading have obvious explicit structural features. Represents the explicit structural feature detection function, with the current input text unit as the parameter. This function can be implemented by integrating existing technologies (such as pre-trained model + rule library); β represents the implicit semantic change weight, which refers to the weight of the semantic change degree of adjacent sentences on the block. Semantic mutation is the core indicator of the potential block boundary; Δ semantic (s i , s i+1 ) represents the semantic difference calculation function of adjacent unit content, which is calculated by BERT embedding cosine similarity, and its formula is expressed as Δ semantic (s i , s i+1 )=1-cos(E(s i ), E(s i+1 )), the parameters are the contents of two adjacent text units, s i Represents the content of the i-th initial text unit.

[0047] If BoundaryScore(s i )>τ, then the content s of the i-th initial text unit i The block boundary is then inserted, and the boundary block threshold τ is dynamically calibrated through the validation set to ensure a balance between recall and precision.

[0048] Step S12: Perform semantic consistency judgment on each initial text unit based on the semantic coherence scoring function. If the semantic coherence score corresponding to the initial text unit is less than the preset semantic score, the initial text unit is divided into blocks again until the semantic coherence score corresponding to each initial text unit in the super-long text is greater than or equal to the preset semantic score, and then output multiple internally semantically consistent initial text units.

[0049] In this step, the semantic coherence scoring function is used to ensure the semantic consistency within each block and control the block size. In the case of semantic inconsistency, the blocks are further split or adjusted. The expression of the semantic coherence scoring function is:

[0050]

[0051] in, represents the kth block, and the block includes several sentences {s1, s2, ..., s n}, cos(E(s i ), E(s i+1 ) represents the semantic similarity of adjacent sentences, and the semantic similarity of adjacent sentences is calculated based on the 768-dimensional vector generated by Sentence-BERT.

[0052] Furthermore, the blocks must meet the minimum length L min =3, to avoid fragmentation, and the maximum length is L max =1000 to prevent information overload. The closer the semantic similarity of adjacent sentences is to 1, the more coherent the sentences are (e.g., discussing the same experimental step consecutively in a paper). The chunking termination threshold θ is a preset semantic score. If the average chunk coherence is less than the chunking termination threshold θ, that is, the average chunk coherence is less than 0.6, re-chunking is triggered. The mean coherence of valid chunks is between 0.65 and 0.75. The default threshold is 0.6, which covers 95% of valid chunks.

[0053] Step S2: construct a topic model based on the initial text units and the hierarchical Dirichlet process to obtain topic distribution labels corresponding to the initial text units, and perform unit grading on each initial text unit.

[0054] In this step, topic modeling technology (HDP) is used to automatically extract topic distribution and key information from the initial text unit, thereby generating a concise summary or compressed version. The results of topic modeling can be stored and used for other model tasks. The topic distribution vector is used as part of the input to help the model better capture the potential structure and semantic information in the text.

[0055] Furthermore, to achieve efficient retrieval and contextual association for very long texts, this technology constructs a three-level index for chapters, paragraphs, and sentences. Once the index content is established, it is input into the model as auxiliary content when processing long texts, enhancing the model's understanding of text data. At the same time, a hierarchical screening mechanism is used to balance retrieval speed and accuracy during large-scale model inference and retrieval. Multi-level vectorized indexes at the chapter, paragraph, and sentence levels are written for each initial text. A chapter-level index is constructed based on the vector representation of BERT-Whitening post-processing, a paragraph-level index is constructed based on the HNSW algorithm, and a sentence-level index is supported based on the FAISSFlat index.

[0056] In one example, BERT-Whitening is used to encode the chapter summary to generate a 128-dimensional vector, which is then accelerated by the IVF_PQ index of FAISS to quickly locate relevant chapters (such as "Experimental Methods" or "Result Analysis"). The paragraph topic vector (HDP output) and the semantic vector (Sentence-BERT) are fused to construct a 256-dimensional hybrid vector, which is then indexed with the HNSW graph to accurately match topic-related paragraphs (such as "Model Training Steps" or "Drug Side Effect Description"). A 768-dimensional fine-grained vector is generated based on Sentence-BERT, and the retrieval accuracy is ensured by the Flat index of FAISS to support the question-answering system in directly locating key sentences (such as "The experimental sample size is 200 groups").

[0057] When the full-text model needs to be processed, the index content is used as an auxiliary input to strengthen the model's reasoning. The index level weight is automatically adjusted according to the large model query type. For example, academic retrieval focuses on the paragraph level, and the question-answering system focuses on the sentence level. The index and the original text are associated through pointer mapping to retain contextual information. For example, the retrieval results are accompanied by associated charts or formulas. This is accomplished by integrating innovative technologies with existing technologies. The logical boundary detection formula in the dynamic semantic segmentation process comprehensively considers the explicit structural features and implicit semantic change weights, and also dynamically calibrates the threshold through the verification set to ensure a balance between recall and precision. This is an innovative semantic segmentation design theory and technology in text segmentation processing. At the same time, existing technologies (such as topic modeling) are integrated for innovative expansion, ultimately realizing the model's intelligent understanding of complex texts and coherent understanding and efficient retrieval of texts at the 100,000-word level.

[0058] Step S3: receiving a user query request and extracting a retrieval text set according to a hierarchical retrieval strategy and the retrieval text corresponding to the user query request, wherein the retrieval text set includes a plurality of initial text units related to the retrieval text.

[0059] In this step, the hierarchical retrieval strategy includes preliminary screening based on the query vector and chapter-level index, and then fine-ranking the candidate chapters obtained from the preliminary screening based on the paragraph index and / or sentence-level index, and finally returning the completed retrieval results to the corresponding initial text unit.

[0060] Step S4: input the constraint prompt words extracted from the topic distribution label corresponding to the search text set, the search text set, and the search text corresponding to the user query request into a text generation model to output the text to be calibrated.

[0061] This step also includes steps S41 to S43.

[0062] Step S41 : pre-define a plurality of knowledge categories related to auxiliary reading, wherein each knowledge category corresponds to a knowledge category feature vector, and the knowledge category feature vector includes a static initialization part and a dynamic update part.

[0063] In this step, knowledge categories related to assisted reading are predefined, such as "medical diagnosis," "legal terms," and "literary analysis," with each corresponding to a feature vector. The feature vectors are divided into a static initialization portion and a dynamic update portion (the static initialization portion can be based on the domain's authoritative knowledge base, while the dynamic update portion can automatically adjust based on user behavior and new data, ensuring that the model both respects professional authority and adapts to new scenarios).

[0064] In step S42, the embedding representation of the topic distribution label corresponding to each initial text unit in the retrieved text set is sequentially compared with all the knowledge category feature vectors through the attention mechanism, and the text semantic features related to the embedding representation and various types of knowledge features are extracted respectively, where the embedding representation includes token-level embedding, whole sentence-level embedding, and paragraph-level embedding.

[0065] In this step, the attention mechanism extracts semantic features from the input text that are relevant to specific knowledge categories. For example, medical text focuses on entities such as "symptoms" and "drugs." Text features are dynamically integrated with knowledge category features to generate constraint prompts, limiting the scope of model generation. For example, in medical scenarios, only verified literature can be retrieved.

[0066] Step S43: Dynamically fuse the text semantic features with the corresponding knowledge category feature vector to obtain the constraint prompt words corresponding to the topic distribution label.

[0067] In this step, by simultaneously modeling the embedding of tokens, whole sentences, and paragraphs, we can take into account both details and the global context, avoid information loss or noise interference caused by a single granularity, and use the attention mechanism to compare the topic distribution label embedding with the feature vector of each knowledge category "point-to-point". The matching weight is dynamically calculated to ensure that the prompt word strictly fits the topic semantics of the current retrieval text. The static initialization part provides the "hard core" support of domain or general knowledge, and the dynamic update part allows the model to be tuned in real time with the retrieval text and user interaction, effectively balancing reliability and adaptability. The extracted text semantic features are fused with the corresponding knowledge category vector at the vector level (such as weighted summation, gating mechanism, etc.), so that the final prompt word retains the details of the retrieval context and incorporates the abstract concepts of category knowledge. The prompt word produced by the above process comes with a three-layer semantic traceability path of "topic → knowledge category → text details", which can not only accurately constrain the generation model to focus on the core content, but also facilitate users or subsequent modules to track and review the source of the prompt.

[0068] Furthermore, classification decisions constrain generation: Based on the fused features, the input text is judged to which knowledge category it belongs, and corresponding generation strategies are activated, such as strict fact verification and loose creative generation. The final output must conform to the logical and factual boundaries of the knowledge category, reducing "illusions" at the source.

[0069] In one example, 600 knowledge categories related to assisted reading are predefined (such as "medical diagnosis", "legal terms", and "literary analysis"), and each category corresponds to a feature vector. The feature vector construction can be expressed as:

[0070]

[0071] Among them, v k represents the final feature vector of the j-th category of knowledge, Represents a static feature vector based on an authoritative knowledge base (80,000 samples have been constructed), It represents the feature vector dynamically updated by user behavior. α represents the fusion weight of static and dynamic features. The default value is 0.5. The fusion weight of static and dynamic features is determined based on cross-validation experiments.

[0072] The input text is encoded by BERT to generate a word embedding matrix. The global features related to the knowledge category are extracted through the attention mechanism and can be expressed as:

[0073]

[0074] Among them, E i Represents the BERT embedding vector (dimension 768) of the ii-th word in the input text, q k Represents the trainable query vector (dimension 768) of the kk-th category of knowledge, hk Represents the aggregated semantic feature vector (dimension 768).

[0075] Calculate the dynamic weight β through the Sigmoid function k , integrating input features and knowledge category features, its expression is:

[0076]

[0077] Among them, σ represents the Sigmoid activation function, W represents the weight matrix (dimension 512×1), β k Indicates the new feature fusion ratio (range [0,1]).

[0078] The matching probability between the input text and the updated features is calculated based on cosine similarity, which can be expressed as:

[0079]

[0080] Among them, h global Represents the global semantic representation of the input text, generated by CLS token, Sim represents the cosine similarity function, that is, K represents the total number of predefined knowledge categories, and K=600.

[0081] Through attention mechanisms and feature fusion, the model can dynamically adjust feature weights for different categories. For example, in a medical text test, the model increased the weight of features in the "Disease Diagnosis" category, ensuring that generated content strictly adheres to authoritative literature and reduces fabricated conclusions. In literary contexts, dynamic features adapt to creative needs, allowing for plausible plot associations while avoiding confusion between the original work and derivative content. Embracing the "dual side" of illusions.

[0082] In step S5, the text to be calibrated and each initial text unit in the retrieval text set are dynamically verified based on the self-verification mechanism. If the error scoring function of the text to be calibrated and any initial text unit in the retrieval text set is less than the preset error threshold, the text to be calibrated is output to the user as a credible text.

[0083] In this step, SelfChecker is pre-built. SelfChecker is a lightweight, small model used to verify hallucinations and dynamically verify and correct the output of the large model. This self-checking mechanism allows the large reading assistance model to identify logical contradictions, factual errors, or ambiguous expressions in the output after it produces a first draft. It then identifies potential errors and provides guidance for correction. This process is similar to the "review" process performed by humans after solving a problem or the "review" process after writing a piece. It can significantly improve the reliability and interpretability of the large language model for complex tasks. The small model will be designed with the unique characteristics of digital reading in mind to significantly improve answer accuracy and prevent misleading answers.

[0084] By automatically segmenting ultra-long texts according to semantic boundaries, we ensure that the internal themes of each initial unit are unified and the information is coherent, avoiding context fragmentation caused by arbitrary equal-length segmentation. The block granularity can be adaptively adjusted according to the text structure and topic complexity, improving the efficiency and accuracy of subsequent processing. In addition, we introduce a hierarchical Dirichlet process to model the topic distribution of each unit, realize the automatic generation of topic labels and unit grading, and the multi-level topic representation takes into account both the global topic framework and local details, providing richer and more discriminative semantic constraints for the retrieval and generation stages. At the same time, we can also search from top to bottom or bottom to top at the topic level according to user queries, give priority to locating the most relevant units, significantly improve the text recall rate and guidance accuracy, and dynamically compare and score the calibrated text with each retrieval unit, which can effectively filter out possible deviations or hallucinations of the model, thereby enhancing the credibility and accuracy of the answer.

[0085] Furthermore, through the self-check mechanism (Self-Check), potential errors in the output of the large model are quantified. The input of the SelfChecker small model includes the original output text T of the large model and the intermediate thought chain C during its generation process. The corresponding error scoring function can be expressed as:

[0086] S error (T, C) = λ1·S fact +λ2·S logic +λ3·S ambiguity

[0087]

[0088] S ambiguity =1-cos(f(T), f(clarification question))

[0089] Among them, S fact represents the factual error score (based on knowledge base retrieval verification), S logic represents the logical consistency score (detected by reasoning chain contradiction), S ambiguityThe fuzziness score is expressed (using semantic similarity analysis), and the f() text embedding function is used to clarify questions such as "Please specify the specific conditions." Based on the text characteristics of digital reading, the use of knowledge classification combined with a self-checking model can effectively eliminate illusions.

[0090] In one example, the method further includes: clustering the text content by topic based on a hierarchical Dirichlet process to generate a topic-content mapping library, and automatically associating different reading topics found in the user's reading history data with topics in the topic-content mapping library. Dynamic interest tags are generated based on the user's reading history data, and topic interest weights are displayed to the user in a visual panel, and manual adjustment instructions from the user are accepted.

[0091] Furthermore, by integrating the semantic analysis capabilities of the ultra-long text processing platform and combining it with multi-dimensional user behavior data, we can build "two-way interactive recommendations." Through conversational interaction, users can directly participate in optimizing the recommendation logic, breaking the "black box" model of traditional recommendation systems, improving user trust in recommendation results and reading efficiency, and achieving the dual goals of accurate recommendations and transparent explanations. The technical solution for this is as follows:

[0092] For topic-driven user interest modeling and dynamic optimization, the ultra-long text processing base uses the hierarchical Dirichlet process (HDP) to cluster massive amounts of content into topics (such as "deep learning algorithms" and "medical imaging diagnosis") to form an interpretable topic-content mapping library. Users' long-term interests are automatically associated with corresponding topics (such as "deep learning algorithms" accounting for 70%) through their reading behavior (such as continuous attention to the "neural network optimization" chapter), and dynamic interest tags are generated in combination with real-time behavior (such as frequent searches for "overfitting solutions" in the past 7 days). Users can view and manually adjust interest weights (such as lowering the priority of "traditional algorithms") through the visualization panel. The system realizes dynamic optimization of the interest model through a time decay model (historical behavior weights decrease monthly) and a feedback loop (labeling and collection trigger topic weight increases).

[0093] For conversational interaction and transparent recommendation logic, users define their needs in natural language (e.g., "Recommend a method to address vanishing gradients, excluding complex formulas"). The model analyzes key requirements and proactively clarifies ambiguous ones (e.g., "Are we more focused on model efficiency or generalization?"). Recommendations are extracted from a database based on topic matching (e.g., "Optimization Algorithms") and behavioral data (e.g., historical annotations of "Residual Connections"). The model also generates readable reasoning for the recommendation (e.g., "You have annotated relevant paragraphs, and this method ranks among the top 3 for addressing vanishing gradients"). Users can rate the recommendations (e.g., "Helpful" or "Irrelevant") or directly modify the logic (e.g., "Prioritize literature from the past three years"). Based on feedback, the system downgrades low-quality content in real time (e.g., "L1 Regularization" is weighted by -20%) and adjusts the topic strategy, forming a closed loop of "demand → recommendation → feedback → iteration."

[0094] Topic mapping enables multi-scenario adaptation. For academic scenarios, it recommends content with a complete logical chain (e.g., "overfitting → regularization → data augmentation"), allowing users to trace content by topic. For fragmented scenarios, it provides summaries and voice broadcasts (TTS technology) for quick browsing on mobile devices. During the cold start phase, new users quickly identify their interests through topic popularity recommendations (e.g., "Recent Top Ten Hot Topics in AI") and conversation guidance (e.g., "Are you interested in model optimization?"). New content enters the candidate pool based on topic similarity (e.g., matching "graph neural networks" with "deep learning"). User feedback determines its long-term weight.

[0095] Personalized recommendations are centered around topic modeling, multi-level indexing, and two-way dialogue interaction within our foundation for processing very long texts. These techniques significantly improve the comprehensiveness and accuracy of recommended content. Users define their needs and influence the recommendation logic through dialogue, breaking with the traditional one-way recommendation model. This framework creates a transparent and precise recommendation experience with deep user engagement, making recommendations more explicit and results more relevant to real-world scenarios, significantly improving reading efficiency.

[0096] In one example, AI technology can be integrated and applied to actual products, enabling features like voice control, AI audiobooks, AI reading companions, adaptive testing, and academic seminars. By deeply integrating TTS (text-to-speech), voiceprint cloning, and voice commands with visual interfaces, multi-sensory collaborative interaction can be achieved, encompassing text, voice, and graphics.

[0097] For simultaneous voice-text feedback, when an interaction is triggered by a voice command (such as "Explain the principle of vanishing gradients"), the system will simultaneously generate a voice broadcast and a text summary. For example, while the voice answer "Gradient vanishing is usually caused by derivative decay in deep network backpropagation..." is displayed, the screen will display a text explanation and a schematic diagram (such as the backpropagation chain rule diagram). Key points are dynamically displayed through highlighting and animation, achieving dual-modal learning of "voice guidance + visual focus", enhancing user understanding and reading immersion.

[0098] For personalized broadcast reading through voiceprint cloning, users can record personalized voiceprints (such as choosing the "academic explanation" or "easy broadcast" style). The system combines voiceprint cloning technology to generate customized voice, and broadcasts through real-time text to speech to adapt to different scenarios and needs (such as driving scenarios, visually impaired users).

[0099] For real-time AI interaction, you can create AI virtual characters, discuss the content of digital reading works with them, and receive personalized reading suggestions and guidance. This also enables functions such as AI reading companionship and AI quizzes.

[0100] Furthermore, the text generation model provides a multimodal interaction interface, which is used to perform one or more of voice command parsing, visual focus prompts, and customized text-to-speech synthesis.

[0101] In a feasible example, the upper limit of long text processing of the large model is a single document ≥ 100,000 words, and the context association accuracy is ≥ 90% (based on the SQuAD long text question and answer dataset test); the interaction and information retrieval response delay of the auxiliary reading system is ≤ 200ms; the auxiliary reading system can support modular access and is compatible with 80% of digital reading platforms; the information distortion rate of large model reasoning is ≤ 5%, and the general large model hallucination rate is generally between 12% and 15% (based on the TruthfulQA dataset test); through the combination of large models and personalized recommendations, the knowledge coverage rate is increased by more than 30%, and the reading efficiency is improved by 25%; the reading platform supports full voice command use, and has intelligent functions such as AI book boy, AI debate, AI quiz, and AI podcast reading.

[0102] Through preprocessing, long texts are segmented into smaller, meaning-based sections, and content is extracted from these sections to reduce semantic redundancy in long input. Compared to existing technical solutions, this approach is expected to increase the contextual accuracy of large models in assisted reading scenarios from approximately 70% to ≥90%, and improve inference cost efficiency by 15%. Existing "hallucinations" eliminate multiple reliance on manual rules. This project, "KnowledgeCFunc Corpus + SelfCheck Self-Verification," implements automated verification through a modular toolchain, reducing manual intervention costs. Model inference strategies based on this approach can effectively reduce hallucinations. The "hallucination" generation rate is expected to drop from approximately 15% with existing technologies to ≤5%, while the content accuracy rate will reach ≥95%. This technical solution is innovative and pioneering in the market. The reading method uses AI technology to support voice command operations and intelligent interaction, meeting the needs of visually impaired users and driving scenarios. The project integrates innovative AI technology modules, such as the AI-based "Opinion Debate Arena," where users can act as "referees" or "debaters" and engage in offensive and defensive dialogues with AI. The system assesses logical vulnerabilities in real time (e.g., using IBM Debater technology), helping users gain a deep understanding of complex topics, addressing the pain point of "passively accepting information" in academic reading, and enhancing critical thinking skills. Based on the shortcomings of AI capabilities in assisted reading scenarios in the digital field, the project explores new technical solutions to break through the bottleneck of general large models. It provides new ideas and practical solutions for addressing the long-text processing capabilities of large models, demonstrating industry-leading capabilities in assisted reading.

[0103] Based on the above method, the embodiment of the present application discloses an AI-assisted reading system based on dynamic block knowledge classification self-verification, referring to Figure 2 The AI-assisted reading system 1 includes a text acquisition module 11, a search constraint module 12, and a text output module 13, wherein:

[0104] The text collection module 11 is used to dynamically segment the super-long text into semantic blocks to obtain multiple initial text units with consistent internal semantics;

[0105] The retrieval constraint module 12 is used to construct a topic model based on the initial text units and the hierarchical Dirichlet process to obtain the topic distribution labels corresponding to the initial text units, and to perform unit classification on each initial text unit, receive a user query request, and extract a retrieval text set based on the hierarchical retrieval strategy and the retrieval text corresponding to the user query request, wherein the retrieval text set includes multiple initial text units related to the retrieval text, input the constraint prompt words extracted from the topic distribution labels corresponding to the retrieval text set, the retrieval text set, and the retrieval text corresponding to the user query request into the text generation model to output the text to be calibrated;

[0106] The text output module 13 is used to dynamically verify the text to be calibrated and each initial text unit in the search text set based on a self-verification mechanism. If the error scoring function of the text to be calibrated and any initial text unit in the search text set is less than a preset error threshold, the text to be calibrated is output to the user as a credible text.

[0107] In one example, the text acquisition module 11 is used to extract block boundaries in an ultra-long text according to a logical boundary detection function, and to successively block the ultra-long text according to the block boundaries to obtain multiple initial text units; a semantic consistency judgment is performed on each initial text unit based on a semantic coherence scoring function; if the semantic coherence score corresponding to the initial text unit is less than a preset semantic score, the initial text unit is block-divided again until the semantic coherence score corresponding to each initial text unit in the ultra-long text is greater than or equal to the preset semantic score, and then multiple internally semantically consistent initial text units are output.

[0108] In one example, multi-level vectorized indexes at the chapter, paragraph, and sentence levels are written for each initial text. A chapter-level index is constructed based on the vector representation of BERT-Whitening post-processing, a paragraph-level index is constructed based on the HNSW algorithm, and a sentence-level index is supported based on the FAISS Flat index.

[0109] In one example, the hierarchical retrieval strategy includes preliminary screening based on the query vector and chapter-level index, and then fine-ranking the candidate chapters obtained from the preliminary screening based on the paragraph index and / or sentence-level index, and finally returning the completed retrieval results to the corresponding initial text unit.

[0110] In one example, the retrieval constraint module 12 is used to predefine a variety of knowledge categories related to auxiliary reading, wherein each knowledge category corresponds to a knowledge category feature vector, and the knowledge category feature vector includes a static initialization part and a dynamic update part; the embedding representation of the topic distribution label corresponding to each initial text unit in the retrieval text set is sequentially compared with all the knowledge category feature vectors through the attention mechanism, and the text semantic features related to the embedding representation and each type of knowledge feature are extracted respectively, wherein the embedding representation includes token-level embedding, sentence-level embedding and paragraph-level embedding; the text semantic features are dynamically fused with the corresponding knowledge category feature vector to obtain the constraint prompt words corresponding to the topic distribution label.

[0111] In one example, the method further includes:

[0112] The text content is clustered based on the hierarchical Dirichlet process to generate a topic-content mapping library, and different reading topics found in the user's reading history data are automatically associated with the topics in the topic-content mapping library;

[0113] Generate dynamic interest tags based on user reading history data, display topic interest weights to users in a visual panel, and receive manual adjustment instructions from users.

[0114] In one example, the text generation model provides a multimodal interaction interface for performing one or more of voice command parsing, visual focus prompts, and customized text-to-speech synthesis.

[0115] See Figure 3 , is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application. Figure 3 As shown, the electronic device 2 may include: at least one processor 21 , at least one network interface 24 , a user interface 23 , a memory 25 , and at least one communication bus 22 .

[0116] The communication bus 22 is used to realize the connection and communication between these components.

[0117] The user interface 23 may include a display screen (Display) and a camera (Camera). Optionally, the user interface 23 may also include a standard wired interface and a wireless interface.

[0118] The network interface 24 may optionally include a standard wired interface or a wireless interface (such as a WI-FI interface).

[0119] The processor 21 may include one or more processing cores. The processor 21 utilizes various interfaces and lines to connect various parts of the entire server. By running or executing instructions, programs, code sets, or instruction sets stored in the memory 25, and calling data stored in the memory 25, the processor 21 performs various server functions and processes data. Optionally, the processor 21 may be implemented in at least one hardware form selected from digital signal processing (DSP), field-programmable gate array (FPGA), and programmable logic array (PLA). The processor 21 may integrate one or a combination of a central processing unit (CPU), a graphics processing unit (GPU), and a modem. The CPU primarily processes the operating system, user interface, and application programs; the GPU is responsible for rendering and drawing the content to be displayed on the display; and the modem is used to handle wireless communications. It is understood that the modem may not be integrated into the processor 21 and may be implemented separately on a single chip.

[0120] Among them, the memory 25 may include a random access memory (Random Access Memory, RAM) and may also include a read-only memory (Read-Only Memory). Optionally, the memory 25 includes a non-transitory computer-readable storage medium. The memory 25 can be used to store instructions, programs, codes, code sets or instruction sets. The memory 25 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as a touch function, a sound playback function, an image playback function, etc.), instructions for implementing the above-mentioned various method embodiments, etc.; the data storage area may store data involved in the above-mentioned various method embodiments, etc. The memory 25 may also be optionally at least one storage device located away from the aforementioned processor 21. As Figure 3 As shown, the memory 25 as a computer storage medium may include an operating system, a network communication module, a user interface module, and an application program of an AI-assisted reading method based on dynamic block knowledge classification and self-verification.

[0121] exist Figure 3 In the electronic device 2 shown, the user interface 23 is mainly used to provide an input interface for the user and obtain data input by the user; and the processor 21 can be used to call an application stored in the memory 25 that uses an AI-assisted reading method based on dynamic block knowledge classification and self-verification. When executed by one or more processors, the electronic device executes one or more methods in the above embodiments.

[0122] A computer-readable storage medium stores instructions, which, when executed by one or more processors, cause the computer to execute one or more methods in the above-mentioned embodiments.

[0123] It should be noted that for the aforementioned method embodiments, for simplicity of description, they are all expressed as a series of action combinations, but those skilled in the art should be aware that this application is not limited by the order of the actions described, because according to this application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required for this application.

[0124] In the above embodiments, the description of each embodiment has its own focus. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0125] In the several embodiments provided in this application, it should be understood that the disclosed devices can be implemented in other ways. For example, the device embodiments described above are merely schematic, such as the division of units, which is only a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some service interface, and the indirect coupling or communication connection of devices or units can be electrical or other forms.

[0126] Units described as separate components may or may not be physically separate, and components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0127] In addition, the functional units in the various embodiments of the present application may be integrated into a single processing unit, or each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0128] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable memory. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a memory and includes several instructions for enabling a computer device (which can be a personal computer, server or network device, etc.) to execute all or part of the steps of the various embodiments of the present application. The aforementioned memory includes various media that can store program codes, such as USB flash drives, mobile hard drives, magnetic disks or optical disks.

[0129] The above are merely exemplary embodiments of the present disclosure and are not intended to limit the scope of the present disclosure. That is, any equivalent changes and modifications made in accordance with the teachings of the present disclosure are still within the scope of the present disclosure. After considering the disclosure of the specification and the truth of practice, those skilled in the art will easily think of other embodiments of the present disclosure. This application is intended to cover any variations, uses, or adaptive changes of the present disclosure, which follow the general principles of the present disclosure and include common knowledge or customary techniques in the art that are not described in the present disclosure.

[0130] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. An AI-assisted reading method based on dynamic block knowledge classification and self-verification, characterized in that: The method comprises: Dynamically segment long text into semantic chunks to obtain multiple initial text units with consistent internal semantics; Constructing a topic model based on the initial text units and a hierarchical Dirichlet process to obtain topic distribution labels corresponding to the initial text units, and performing unit grading on each initial text unit; receiving a user query request, and extracting a search text set according to a hierarchical search strategy and a search text corresponding to the user query request, wherein the search text set includes a plurality of initial text units related to the search text; Inputting the constraint prompt words extracted from the topic distribution label corresponding to the search text set, the search text set, and the search text corresponding to the user query request into a text generation model to output a text to be calibrated; Based on the self-verification mechanism, the text to be calibrated is dynamically verified with each initial text unit in the search text set. If the error scoring function of the text to be calibrated and any initial text unit in the search text set is less than a preset error threshold, the text to be calibrated is output to the user as a credible text.

2. The AI-assisted reading method based on dynamic block knowledge classification and self-verification as claimed in claim 1 is characterized in that: The dynamic semantic segmentation of the overlong text to obtain multiple initial text units with consistent internal semantics specifically includes: Extracting block boundaries in the super-long text according to a logical boundary detection function, and dividing the super-long text into blocks one by one according to the block boundaries to obtain a plurality of initial text units; A semantic consistency judgment is performed on each initial text unit based on a semantic coherence scoring function. If the semantic coherence score corresponding to the initial text unit is less than a preset semantic score, the initial text unit is segmented again until the semantic coherence score corresponding to each initial text unit in the super-long text is greater than or equal to the preset semantic score, and then a plurality of internally semantically consistent initial text units are output.

3. The AI-assisted reading method based on dynamic block knowledge classification and self-verification as claimed in claim 1, characterized in that: Multi-level vectorized indexes at the chapter, paragraph, and sentence levels are written for each initial text. The chapter-level index is built based on the vector representation of BERT-Whitening post-processing, the paragraph-level index is built based on the HNSW algorithm, and the sentence-level index is supported based on the FAISS Flat index.

4. The AI-assisted reading method based on dynamic block knowledge classification and self-verification as claimed in claim 1 is characterized in that: The hierarchical retrieval strategy includes performing a preliminary screening based on the query vector and the chapter-level index, and then refining the candidate chapters obtained from the preliminary screening based on the paragraph index and / or sentence-level index, and finally returning the completed retrieval results to the corresponding initial text unit.

5. The AI-assisted reading method based on dynamic block knowledge classification and self-verification as claimed in claim 1 is characterized in that: The constraint prompt words extracted based on the topic distribution labels include: Predefine a plurality of knowledge categories related to auxiliary reading, wherein each knowledge category corresponds to a knowledge category feature vector, and the knowledge category feature vector includes a static initialization part and a dynamic update part; The embedding representation of the topic distribution label corresponding to each initial text unit in the retrieved text set is sequentially compared with all knowledge category feature vectors through an attention mechanism, and the text semantic features related to the embedding representation and various knowledge features are respectively extracted, wherein the embedding representation includes token-level embedding, whole sentence-level embedding, and paragraph-level embedding; The text semantic features are dynamically fused with the corresponding knowledge category feature vectors to obtain the constraint prompt words corresponding to the topic distribution labels.

6. The AI-assisted reading method based on dynamic block knowledge classification and self-verification as claimed in claim 1, characterized in that: The method further comprises: Performing topic clustering on text content based on a hierarchical Dirichlet process to generate a topic-content mapping library, and automatically associating different reading topics found in the user's reading history data with topics in the topic-content mapping library; Generate dynamic interest tags based on the user's reading history data, display the subject interest weights to the user in a visual panel, and receive manual adjustment instructions from the user.

7. The AI-assisted reading method based on dynamic block knowledge classification and self-verification as claimed in claim 1, characterized in that: The text generation model provides a multimodal interaction interface, which is used to perform one or more of voice command parsing, visual focus prompts, and customized text-to-speech synthesis.

8. An AI-assisted reading system based on dynamic block knowledge classification and self-verification, characterized by: The AI-assisted reading system (1) comprises a text acquisition module (11), a search constraint module (12) and a text output module (13), wherein: The text collection module (11) is used to dynamically segment the super-long text into semantic blocks to obtain a plurality of initial text units with consistent internal semantics; The retrieval constraint module (12) is used to construct a topic model based on the initial text unit and the hierarchical Dirichlet process to obtain the topic distribution label corresponding to the initial text unit, and perform unit classification on each initial text unit, receive a user query request, and extract a retrieval text set according to a hierarchical retrieval strategy and the retrieval text corresponding to the user query request, wherein the retrieval text set includes a plurality of initial text units related to the retrieval text, and input the constraint prompt words extracted from the topic distribution label corresponding to the retrieval text set, the retrieval text set and the retrieval text corresponding to the user query request into a text generation model to output a text to be calibrated; The text output module (13) is used to dynamically verify the text to be calibrated and each initial text unit in the search text set based on a self-verification mechanism. If the error scoring function of the text to be calibrated and any initial text unit in the search text set is less than a preset error threshold, the text to be calibrated is output to the user as a credible text.

9. An electronic device, characterized in that: The electronic device (2) comprises a processor (21), a memory (25), a user interface (23) and a network interface (24), wherein the memory (25) is used to store instructions, the user interface (23) and the network interface (24) are used to communicate with other devices, and the processor (21) is used to execute the instructions stored in the memory (25) so that the electronic device (2) executes the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images