Reading software intelligent search system based on natural language processing technology

Abstract

The invention discloses a reading software intelligent search system based on a natural language processing technology, and relates to the technical field of computer software and natural language processing, and the reading software intelligent search system comprises a user interaction interface which is used for receiving an original search request initiated by a user in a reading context; the query understanding engine is connected to the user interaction interface and used for carrying out reading scene-oriented deep semantic analysis on the original search request, and the deep semantic analysis at least comprises the steps of identifying an exploratory intention related to a reading behavior and sending the exploratory intention to the user interaction interface; and performing context completion on anaphora and omission in the query by using the real-time reading progress of the user and historical notes so as to output a structured query object. According to the method, a rollback strategy is triggered through consistency verification failure, the misleading generation result output risk is reduced, crossing from keyword matching to semantic understanding, from a static list to dynamic generation and from isolated search to situational dialogue is achieved, and intelligent reading search experience is provided for a user.

Description

Technical Field

[0001] This invention relates to the fields of computer software and natural language processing technology, and in particular to an intelligent search system for reading software based on natural language processing technology. Background Technology

[0002] With the widespread adoption of digital reading, the number of documents in reading software is growing rapidly, and users often have exploratory search needs with context during the reading process. Existing technologies mostly employ solutions such as keyword search, general semantic search, knowledge graph search, or personalized ranking and summary generation.

[0003] However, in reading scenarios, challenges remain: query ambiguity due to referential / omission issues, insufficient anchoring of chapter structure, crude utilization of behavioral signals, and a lack of traceable evidence and consistency constraints in generated content. For example, semantic retrieval schemes based on knowledge graphs can improve entity matching and templated query capabilities, but they typically lack strong constraints from the reading context and consistency verification mechanisms. Therefore, we propose an intelligent search system for reading software based on natural language processing technology. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an intelligent search system for reading software based on natural language processing technology. Through a closed-loop mechanism of "reading context completion - structural anchor index retrieval - evidence constraint generation - consistency verification and rollback", it significantly improves search accuracy and the credibility of generated content in reading scenarios.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: Intelligent search systems for reading software based on natural language processing technology include: The user interaction interface is used to receive the original search request initiated by the user in the reading context; The query understanding engine, connected to the user interaction interface, is used to perform deep semantic analysis of the original search request in a reading scenario. The deep semantic analysis includes at least: identifying exploratory intents related to reading behavior, and using the user's real-time reading progress and historical notes to complete the context of references and omissions in the query, so as to output a structured query object. A document semantic indexing and retrieval engine is used to construct and store a structural anchor semantic index. The structural anchor semantic index contains the hierarchical structure information of the document (including at least chapter / section / paragraph anchor identifiers) and knowledge graph entity anchors. Based on the structured query object, semantic similarity matching combined with the document structure is performed to obtain a preliminary retrieval result set. The personalized sorting and content generation engine connects to the query understanding engine and the document semantic indexing and retrieval engine. It is used to reorder the initial search result set based on user behavior profiles and generate summary or explanatory content that matches the query intent. The user behavior collection and profile modeling module is used to collect and analyze multi-dimensional in-depth behavior sequences of users during the reading process in real time, and to build and update dynamic user profiles. The personalized sorting and content generation engine is configured to introduce a Reading Context Evidence Constraint Architecture (RCS) when generating summary or explanatory content, wherein the RCS includes at least: (a) Must-Anchor Fields: Used to limit the chapter / paragraph anchor identifier, cited fragment ID, and evidence strength field that each conclusion / answer sentence in the generated content must be bound to; (b) Must-Hold Constraints: Consistencies in the original text that the generated content must not violate, wherein the consistency constraints include at least one or more of the following: consistency of concept definitions, consistency of character relationships, and consistency of timeline. Furthermore, the personalized sorting and content generation engine is configured to perform consistency checks on the generated content. When the consistency check fails, a rollback strategy is triggered. The rollback strategy includes at least: outputting only the retrieval evidence card, reducing the confidence level and prompting manual review, or requiring supplementary evidence before regenerating.

[0006] As a preferred technical solution of the present invention, the query understanding engine includes: an intent recognition unit, a semantic parsing and entity linking unit, and a query expansion and correction unit; wherein, the intent recognition unit adopts a Transformer-based sequence classification model to classify query intents into at least one of fact query, opinion exploration, content recommendation, comparative analysis, or task execution; wherein, the semantic parsing and entity linking unit is used to identify named entities in the query and link them to standard entity nodes in the system's built-in or external knowledge graph, while parsing the semantic role framework of the query; wherein, the query expansion and correction unit, based on the intent recognition results and linked entities, uses a thesaurus network and an associated concept graph to semantically expand the original query, and performs context completion for queries with referential or ellipsis based on the session history.

[0007] As a preferred technical solution of the present invention, the document semantic indexing and retrieval engine includes: a document deep encoder, a knowledge enhancement module, and a hierarchical index library.

[0008] As a preferred technical solution of the present invention, the document deep encoder encodes the title, body, paragraphs and metadata of the document based on a pre-trained language model to generate document-level and paragraph-level multi-granularity semantic vectors. The knowledge enhancement module introduces external knowledge graph information during the encoding process to enhance the representation of entities and concepts in the document. The hierarchical index library stores multi-granularity semantic vectors and supports multi-level retrieval based on the approximate nearest neighbor search algorithm. Each vector entry is bound to the corresponding chapter / section / paragraph anchor point identifier.

[0009] As a preferred technical solution of the present invention, the personalized sorting and content generation engine includes an intent-driven dynamic content planner, a dynamic summary generator, and a multi-dimensional sorting model. The dynamic content planner selects a generation mode based on the structural position distribution of the search results in the document hierarchy. The generation modes include at least explanatory summaries, comparative reviews, timeline summaries, or relationship graph summaries.

[0010] As a preferred technical solution of the present invention, the user behavior collection and profile modeling module is configured to: first, reorder the preliminary search results based on the deep interest focus and cognitive state vector mined from user behavior; second, dynamically generate integrated content aimed at bridging the gaps in user knowledge based on the reordered result set, the user's exploratory intent, and the structural position distribution of the search results in the document. The user behavior collection and profiling modeling module includes: A real-time behavior stream processor is used to capture users' search, click, reading time, page turning speed, underlining, note-taking, and sharing behaviors. The interest topic evolution model, based on long-term user behavior data, uses topic models or deep interest networks to extract users’ interest tags in different dimensions and the changing trends of their intensity over time. A short-term conversation context-aware unit is used to capture the sequence of behaviors within the current search session and construct a short-term contextualized preference vector.

[0011] As a preferred technical solution of the present invention, the user behavior collection and profile modeling module is used to construct short-term conversation context vectors and long-term interest topic evolution vectors, and input the constructed vectors as features into the multi-dimensional ranking model and dynamic content planner of the personalized ranking and content generation engine.

[0012] As a preferred technical solution of the present invention, the dynamic summary generator adopts an encoder-decoder architecture. Based on the intent and key entities in the structured query object, it extracts or generates a coherent summary text focused on the query point from the top-K retrieval result documents. The ranking factors of the multi-dimensional ranking model include at least: query-document semantic relevance score, document preference based on user's historical clicks and reading time, document timeliness weight, document authority weight, social collaborative filtering recommendation score, and context awareness.

[0013] As a preferred technical solution of the present invention, the consistency verification includes: extracting entities and relation assertions from the generated content and matching them with reference fragments bound to Must-Anchor Fields; determining that the consistency verification fails when the matching fails or conflicts with Must-Hold Constraints.

[0014] As a preferred technical solution of the present invention, an intelligent search method for reading software based on natural language processing technology includes the following steps: S1: Receives the original search request initiated by the user through the user interaction interface; S2: The query understanding engine performs deep semantic analysis on the original search request and generates a structured query object; S3: Document semantic indexing and retrieval engine performs structural anchor semantic retrieval based on structured query objects to obtain a preliminary retrieval result set; S4: The personalized ranking and content generation engine combines user behavior profiles to re-rank the initial search results set and generate auxiliary content under the RCS constraint of the reading context evidence constraint architecture. S5: Present the final sorted result list and generated auxiliary content to the user; when the consistency check fails, execute the rollback strategy.

[0015] The beneficial effects of this invention are as follows: By supplementing context with reading progress and notes, the system improves the accuracy of referential resolution and intent recognition; through structural anchor semantic indexing, it achieves precise location of search results and chapters / paragraphs; through RCS's Must-Anchor and Must-Hold mechanisms, it ensures that generated summaries / answers have traceable evidence and suppresses conflicts with the original text; and by triggering a rollback strategy through consistency check failures, it reduces the risk of generating misleading results. This system achieves a leap from keyword matching to semantic understanding, from static lists to dynamic generation, and from isolated search to contextual dialogue, providing users with an intelligent reading and search experience. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the overall system architecture of the present invention; Figure 2 This is a schematic diagram of the system composition structure of the present invention; Figure 3 This is a flowchart of the query comprehension engine workflow of the present invention; Figure 4 This is a schematic diagram illustrating the document semantic index construction process of this invention; Figure 5 This is a flowchart of the personalized sorting and content generation engine workflow of the present invention; Figure 6 This is a flowchart of the search method of the present invention.

[0017] In the diagram: 100, User Interaction Interface; 200, Query Understanding Engine; 201, Intent Recognition Unit; 202, Semantic Parsing and Entity Linking Unit; 203, Query Expansion and Correction Unit; 300, Document Semantic Indexing and Retrieval Engine; 301, Document Deep Encoder; 302, Knowledge Enhancement Module; 303, Hierarchical Index Library; 400, Personalized Ranking and Content Generation Engine; 401, Dynamic Content Planner; 402, Dynamic Summary Generator; 403, Multi-Dimensional Ranking Model; 500, User Behavior Collection and Profile Modeling Module; 501, Real-Time Behavior Stream Processor; 502, Interest Topic Evolution Model; 503, Short-Term Conversation Context Awareness Unit. Detailed Implementation

[0018] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0019] Example 1, Example 1, refer to Figure 1-5 The intelligent search system for reading software based on natural language processing technology includes a user interaction interface 100, a query understanding engine 200, a document semantic indexing and retrieval engine 300, a personalized sorting and content generation engine 400, and a user behavior collection and profile modeling module 500.

[0020] Furthermore, the user interaction interface 100 is used to receive the original search request initiated by the user in the reading context. The user interaction interface 100 is a cross-modal interaction interface that supports at least one of the following interaction methods: natural language text input box, supporting multi-turn conversational search, voice recognition and input interface, graphical conditional filtering interface based on gesture or touch, and augmented reality interface, which captures physical books or environmental text through the device camera and uses it as search input. Specifically, the user interaction interface 100, as the system front end, receives diverse user input. It not only includes a traditional text search box, but also integrates a voice recognition module, an image recognition module (for OCR recognition of images or physical text), and a contextual search entry that can be directly triggered within the reading interface. This interface is also responsible for presenting the final search results in rich media formats (such as cards, summaries, and speech bubbles).

[0021] Furthermore, the query understanding engine 200 works in collaboration with the user behavior collection and profile modeling module 500. Specifically, when parsing a query, the query understanding engine 200 receives the current book identifier, recently read chapter, and high-frequency note keywords from the user profile modeling module 500 as contextual input to resolve pronouns and complete omissions.

[0022] Furthermore, the query understanding engine 200 connects to the user interaction interface 100 to perform deep semantic analysis of the original search request, tailored to the reading scenario. This deep semantic analysis includes at least: identifying exploratory intents related to reading behavior, and using the user's real-time reading progress and historical notes to complete the context of references and omissions in the query, outputting a structured query object. Further, it utilizes real-time reading progress for completion: for example, when a user searches for "What happened to him later?" while reading Chapter 30 of *Dream of the Red Chamber*, the system can automatically link "him" to the core character of the current chapter, "Jia Baoyu". Specifically, this is the key entry point for achieving "intelligent" functionality; the engine performs multi-level, fine-grained NLP analysis on the original query.

[0023] Furthermore, the document semantic index and retrieval engine 300 is used to construct and store a structural anchor semantic index. This index contains hierarchical structural information of the document (at least including chapter / section / paragraph anchor identifiers) and knowledge graph entity anchors. Based on structured query objects, it performs semantic similarity matching combined with document structure to obtain a preliminary retrieval result set. The document semantic index and retrieval engine 300 also stores a deep semantic vector index of documents within the reading software. This stored deep semantic vector index contains hierarchical structural information of the document and knowledge graph entity anchors, used for semantic similarity matching combined with document structure based on structured query objects. Further, "hierarchical structural information and entity anchors" refers to the fact that in vectorized representation, not only is textual semantics encoded, but hierarchical information such as "chapter," "section," and "paragraph" is also incorporated through positional encoding or special markers. Furthermore, knowledge graph IDs are linked to entities mentioned in the text, giving the index "structure awareness" and "knowledge awareness" capabilities. Specifically, this is the core of the system's retrieval, distinguishing it from traditional inverted indexes.

[0024] Furthermore, the personalized ranking and content generation engine 400 is connected to the query understanding engine 200 and the document semantic indexing and retrieval engine 300 to reorder the preliminary retrieval result set based on user behavior profiles and generate summary or explanatory content that matches the query intent; specifically, the engine transforms the preliminary semantic retrieval results into what the end user sees.

[0025] Furthermore, the personalized ranking and content generation engine 400 is configured to introduce a Reading Context Evidence Constraint Architecture (RCS) when generating summary or explanatory content. The RCS includes at least: (a) Must-Anchor Fields: Used to limit the chapter / paragraph anchor identifier, cited fragment ID, and evidence strength field that each conclusion / answer sentence in the generated content must be bound to; (b) Must-Hold Constraints: Consistency constraints that the generated content must not violate. Consistency constraints include at least one or more of the following: consistency of concept definition, consistency of character relationships, and consistency of timeline. Consistency verification includes: extracting entity and relationship assertions from the generated content and matching them with the reference fragments bound to Must-Anchor Fields; if the match fails or conflicts with Must-Hold Constraints, the consistency verification is deemed to have failed. Furthermore, the personalized sorting and content generation engine 400 is configured to perform consistency checks on the generated content. When the consistency check fails, a rollback strategy is triggered. The rollback strategy includes at least: outputting only the retrieval evidence card, reducing the confidence level and prompting manual review, or requiring supplementary evidence before regenerating.

[0026] Furthermore, the user behavior collection and profile modeling module 500 is used to collect and analyze multi-dimensional in-depth behavioral sequences of users during the reading process in real time, and to build and update dynamic user profiles; specifically, this module is the data foundation for the system to achieve personalization.

[0027] Furthermore, the query understanding engine 200 includes: The intent recognition unit 201 uses a Transformer-based sequence classification model to classify query intent into at least one of fact query, opinion exploration, content recommendation, comparative analysis, or task execution. In practice, a fine-tuned BERT or similar Transformer model is used to classify queries into fine-grained intent categories, such as "find specific facts", "seek explanation / definition", "compare two concepts", "request recommendations for similar content", "perform an action (such as jumping to a certain chapter)", etc. Different intents will trigger different subsequent processing flows.

[0028] The semantic parsing and entity linking unit 202 is used to identify named entities in the query and link them to standard entity nodes in the system's built-in or external knowledge graph, while parsing the semantic role framework of the query. In practical applications, sequence labeling models (such as BiLSTM-CRF) are used to identify entities such as names of people, places, book titles, and technical terms in the query. Then, entity linking technology is used to link these surface-level entities to standardized nodes in the "reading domain knowledge graph" constructed by the system. For example, "Jobs" is linked to the entity "Steve Jobs," and associated with nodes such as his works and related biographies. Simultaneously, dependency parsing and semantic role labeling are performed to understand the relationships between actions, subjects, and objects in the query.

[0029] The query expansion and correction unit 203 expands the semantics of the original query based on the intent recognition results and linked entities using a thesaurus and a related concept graph, and completes the context for queries with references or omissions based on the session history. In practical use, based on the identified intent and entities, semantic expansion is performed using a thesaurus and a concept hyponym graph. For example, a query for "good books on machine learning" can be expanded to "(machine learning OR deep learning OR artificial intelligence) AND (good books OR recommendations OR classic works)". More importantly, the system maintains session-level context, enabling it to resolve and complete queries such as "Who is its author?" (referring to the book mentioned above).

[0030] Furthermore, the document semantic indexing and retrieval engine 300 includes: The Document Deep Encoder 301 encodes the title, body, paragraphs, and metadata of a document based on a pre-trained language model, generating document-level and paragraph-level multi-granular semantic vectors. In practical applications, it uses contrastive learning pre-trained models such as Sentence-BERT and SimCSE to encode each document and its internal paragraphs, even important sentences, generating high-dimensional semantic vectors. These vectors can capture "semantic similarity" rather than "lexical overlap." The knowledge enhancement module 302 introduces external knowledge graph information during the encoding process to enhance the representation of entities and concepts in the document. In actual use, during the encoding process, it identifies entities in the document and obtains information such as the entity's description, attributes, and related entities from the knowledge graph. This knowledge information is then integrated into the vectorized representation of the document with special tags, making the document representation more informative. The external knowledge graph introduced by the knowledge enhancement module 302 is a domain-specific reading knowledge graph. This graph contains entities such as books, people, concepts, and schools of thought, and defines the relationships between entities in the reading context, such as "belongs to a school of thought," "originates from a work," and "historical period." The hierarchical index library 303 stores multi-granular semantic vectors and supports multi-level retrieval based on the approximate nearest neighbor search algorithm. In actual use, an efficient vector database (such as Faiss, Milvus) is built to store document-level and paragraph-level vectors. During retrieval, the structured query object is first encoded as a vector, and then the set of semantically most relevant documents and paragraphs is quickly found through the approximate nearest neighbor (ANN) search.

[0031] Furthermore, the personalized sorting and content generation engine 400 includes: The multi-dimensional ranking model 403 has ranking factors that include at least: query-document semantic relevance score, document preference based on user history clicks and reading time, document timeliness weight, document authority weight, social collaborative filtering recommendation score, and context awareness. In practical applications, a weighted ranking function is designed, with input features including: ① cosine similarity (semantic relevance) between the query vector and the document vector; ② historical preference scores for the document's topic and author in the user profile; ③ document freshness (publication / update time); ④ document authority (source publisher, author reputation, number of citations / shares); ⑤ "Similar users also like" score based on collaborative filtering. The model (such as LambdaMART) is trained offline using historical interaction data to maximize click-through rate, reading completion rate, and other objectives; ⑥ context-aware input features include at least: real-time cognitive focus vector in the user profile, query intent category, semantic similarity between the document / paragraph vector and the query vector, and the popularity weight of the document / paragraph annotated (underlined / noted) by users with similar cognitive states. "Users with similar cognitive states" refers to other users found through clustering algorithms who have similar underlining patterns, note question types, and reading jump paths to the current user.

[0032] The dynamic summary generator 402 adopts an encoder-decoder architecture. Based on the intent and key entities in the structured query object, it extracts or generates a coherent summary text focused on the query point from the top-K retrieval result documents. The dynamic summary generator 402 also includes a knowledge-enhanced content generator: according to the strategy of the content planner, it calls the corresponding generation model. During the generation process, it uses the retrieved document fragments as constraints and selectively introduces descriptive information of relevant entities in the knowledge graph to enhance the accuracy and information content of the generated content. The intent-driven dynamic content planner 401 determines the form and focus of the final output content based on the exploratory intent categories in the structured query objects and the topic and structural distribution of the preliminary search results set. The planning strategies include: generating an explanatory summary when the intent is "conceptual understanding" and the results are concentrated in a single area; generating a comparative overview when the intent is "viewpoint comparison" and the results are scattered in multiple places; and generating a timeline or relationship diagram when the intent is "plot summary".

[0033] In practical use, this component is activated for "explanation" or "summary" intents, or when the top results involve multiple documents. It employs generative models such as BART and T5, extracting key information from relevant documents based on the query, and generating a concise, coherent, and comprehensive summary that directly addresses the query, presenting it directly to the user and saving them the trouble of poring over multiple documents.

[0034] Furthermore, the user behavior collection and profile modeling module 500 is used to construct short-term conversation context vectors and long-term interest topic evolution vectors, and inputs the constructed vectors as features into the multi-dimensional ranking model 403 and dynamic content planner 401 of the personalized ranking and content generation engine 400. The user behavior collection and profiling module 500 is configured as follows: First, based on the deep interest focus and cognitive state vector mined from user behavior, the preliminary search results are reordered; second, based on the reordered result set, the user's exploratory intent, and the structural location distribution of the search results in the document, integrated content aimed at bridging the gaps in the user's knowledge is dynamically generated; "exploratory intent" can include categories such as "concept understanding", "viewpoint comparison", "causal tracing", and "plot summary", which are different from the "navigation", "transaction", and "information" intents of general search; "deep interest focus and cognitive state vector" includes not only topic interests, but also "depth of understanding", "points of confusion", and "attention preferences" inferred by analyzing the sentiment tendency of underlined content, the question sentence structure of notes, and the jump pattern; "structural location distribution" refers to the system's consideration of whether the information source is concentrated in a certain chapter or scattered in various places when generating a summary or answer, thereby deciding whether to generate "in-depth explanation" or "overview".

[0035] The user behavior collection and profiling modeling module 500 includes: The real-time behavior stream processor 501 is used to capture user search, click, reading time, page turning speed, underlining, note-taking, and sharing behaviors. In actual use, it collects all explicit and implicit user feedback within the app in a non-intrusive manner: search terms, click results, time spent on a page, page turning speed, underlining, highlighting, adding notes, sharing, rating, etc.

[0036] Interest Topic Evolution Model 502, based on long-term user behavior data, uses topic models or deep interest networks to extract users’ interest tags in different dimensions and the changing trends of their intensity over time. In practical applications, Dynamic Topic Model (DTM) or Transformer-based sequence model is used to analyze long-term user behavior sequences and construct a dynamically changing interest profile, which includes several interest tags (such as "science fiction", "investment and finance", "Python programming") and their weights that decay or increase over time.

[0037] The short-term conversation context-aware unit 503 is used to capture the sequence of behaviors within the current search session and construct a short-term contextualized preference vector. In actual use, the user's behavior sequence in the current search session (such as consecutive searches for "black hole", "singularity", and "Hawking radiation") is recorded to form a short-term, focused context vector, which is used to influence the ranking of search results in the current session.

[0038] Example 2, refer to Figure 6This embodiment is an optimization based on Embodiment 1, specifically: an intelligent search method for reading software based on natural language processing technology, including the following steps: S1: Receive the original search request initiated by the user through the user interaction interface 100; S2: The query understanding engine 200 performs deep semantic analysis on the original search request and generates structured query objects; S3: Document Semantic Indexing and Retrieval Engine 300 performs semantic retrieval based on structured query objects to obtain a preliminary retrieval result set; S4: Personalized sorting and content generation engine 400 combines the current user's behavioral profile to reorder the initial search results set and generate auxiliary content; S5: Present the final sorted results list and generated auxiliary content to the user through the user interaction interface 100.

[0039] Furthermore, in step S2, “deep semantic analysis” specifically includes classifying the query intent and determining the dominant search intent, performing named entity recognition and linking and associating entities in the query with the knowledge base, and expanding and completing the query based on intent and entities and in conjunction with the session history.

[0040] Furthermore, in step S4, "generating auxiliary content" includes at least one of the following: when the search results involve multiple documents, generating a comparative summary that integrates information from multiple documents; when the query is an interrogative sentence, generating a direct and concise answer; and generating recommended items such as "you may also like to know" or "related in-depth reading" based on the query intent and user profile.

[0041] Furthermore, the above method also includes step S0: the build and update phase, which further includes: S01: Offline processing and reading software document library, which constructs a semantic index of documents through document depth encoder 301; S02: Continuously collect and process user behavior data to update the user interest topic evolution model 502; S03: Based on user feedback and behavioral data, perform online learning and optimization of the models in the query understanding engine 200 and the ranking models in the personalized ranking and content generation engine 400.

[0042] In summary: 1. The search accuracy of this invention is effectively improved: Through deep semantic understanding and knowledge graph-enhanced indexing, the system can capture the deep semantic relationship between queries and documents, and can accurately match even if the wording is different, greatly reducing "missed detections" and "false detections".

[0043] 2. This invention provides a highly personalized search experience: search is no longer "one-size-fits-all". The system combines the user's unique reading history, interests and preferences and real-time context to customize the most relevant search results ranking and content recommendations for each user, making search a personalized content discovery engine; 3. The interactive method of this invention is intelligent and user-friendly: It supports multi-turn dialogue, referential resolution, and natural language question answering. Users can express their needs in a more natural and casual way, which lowers the threshold for using search. In addition, the multi-modal entry point (voice, image) further expands the interaction scenarios.

[0044] 4. The results of this invention exhibit high value density: The system not only provides a list of links, but also proactively generates summaries, directly answers questions, and recommends deeply related content, moving the information integration process forward and effectively improving the efficiency of users obtaining effective information, providing an "answer as a service" experience.

[0045] 5. The system of this invention has self-evolution capability: through continuous user behavior collection and online model learning, the system's query understanding, sorting and generation models can be continuously optimized to adapt to changes in user group language habits and interests, and have long-term vitality and adaptability.

[0046] 6. The results of this invention present a qualitative change from "information provision" to "knowledge stitching": Because the content generation engine is driven by exploratory intent and dynamically plans based on the structural distribution of the results, the system can proactively generate integrated content such as "comparative summaries", "contextual analysis", and "concept explanations" that directly address users' knowledge gaps, upgrading the search experience into a personalized reading aid and knowledge companion.

[0047] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An intelligent search system for reading software based on natural language processing technology, characterized in that: include: The user interaction interface (100) is used to receive the original search request initiated by the user in the reading context; The query understanding engine (200), connected to the user interaction interface (100), is used to perform deep semantic analysis of the original search request in the context of reading. The deep semantic analysis includes at least: identifying exploratory intents related to reading behavior, and using the user's real-time reading progress and historical notes to complete the context of the references and omissions in the query, so as to output a structured query object. Document semantic indexing and retrieval engine (300) is used to construct and store structural anchor semantic index, which contains hierarchical structure information of the document (at least including chapter / section / paragraph anchor identifiers) and knowledge graph entity anchors, and performs semantic similarity matching based on the structured query object in combination with the document structure to obtain a preliminary retrieval result set; A personalized sorting and content generation engine (400) is connected to a query understanding engine (200) and a document semantic indexing and retrieval engine (300) to reorder the initial retrieval result set based on user behavior profiles and generate summary or explanatory content that matches the query intent. The User Behavior Collection and Profile Modeling Module (500) is used to collect and analyze the multi-dimensional deep behavior sequence of users during the reading process in real time, and to build and update dynamic user profiles. The personalized sorting and content generation engine (400) is configured to introduce a Reading Context Evidence Constraint Architecture (RCS) when generating summary or explanatory content, wherein the RCS includes at least: (a) Must-Anchor Fields: Used to limit the chapter / paragraph anchor identifier, cited fragment ID, and evidence strength field that each conclusion / answer sentence in the generated content must be bound to; (b) Must-Hold Constraints: Consistencies in the original text that the generated content must not violate, wherein the consistency constraints include at least one or more of the following: consistency of concept definitions, consistency of character relationships, and consistency of timeline. Furthermore, the personalized sorting and content generation engine (400) is configured to perform consistency checks on the generated content. When the consistency check fails, a rollback strategy is triggered. The rollback strategy includes at least: outputting only the retrieval evidence card, reducing the confidence level and prompting manual review, or requiring supplementary evidence before regenerating.

2. The intelligent search system for reading software based on natural language processing technology according to claim 1, characterized in that, The query understanding engine (200) includes: an intent recognition unit (201), a semantic parsing and entity linking unit (202), and a query expansion and correction unit (203); wherein, the intent recognition unit (201) adopts a Transformer-based sequence classification model to classify query intent into at least one of fact query, opinion exploration, content recommendation, comparative analysis, or task execution; wherein, the semantic parsing and entity linking unit (202) is used to identify named entities in the query and link them to standard entity nodes in the system's built-in or external knowledge graph, while parsing the semantic role framework of the query; wherein, the query expansion and correction unit (203) performs semantic expansion of the original query based on the intent recognition results and linked entities using a thesaurus and associated concept graph, and performs context completion for queries with referencing or omission based on the session history.

3. The intelligent search system for reading software based on natural language processing technology according to claim 1, characterized in that, The document semantic indexing and retrieval engine (300) includes: a document deep encoder (301), a knowledge enhancement module (302), and a hierarchical index library (303).

4. The intelligent search system for reading software based on natural language processing technology according to claim 3, characterized in that, The document deep encoder (301) encodes the document's title, body, paragraphs, and metadata based on a pre-trained language model, generating document-level and paragraph-level multi-granular semantic vectors. The knowledge enhancement module (302) introduces external knowledge graph information during the encoding process to enhance the representation of entities and concepts in the document. The hierarchical index library (303) stores multi-granular semantic vectors and supports multi-level retrieval based on the approximate nearest neighbor search algorithm. Each vector entry is bound to the corresponding chapter / section / paragraph anchor point identifier.

5. The intelligent search system for reading software based on natural language processing technology according to claim 1, characterized in that, The personalized ranking and content generation engine (400) includes an intent-driven dynamic content planner (401), a dynamic summary generator (402), and a multi-dimensional ranking model (403). The dynamic content planner (401) selects a generation mode based on the structural position distribution of the search results in the document hierarchy. The generation mode includes at least explanatory summaries, comparative reviews, timeline summaries, or relationship graph summaries.

6. The intelligent search system for reading software based on natural language processing technology according to claim 1, characterized in that, The user behavior collection and profile modeling module (500) is configured to: first, reorder the preliminary search results based on the deep interest focus and cognitive state vector mined from user behavior; second, dynamically generate integrated content aimed at bridging the gaps in user knowledge based on the reordered result set, the user's exploratory intent, and the structural position distribution of the search results in the document. The user behavior collection and profiling modeling module (500) includes: A real-time behavior stream processor (501) is used to capture user search, click, reading time, page turning speed, underlining, note-taking, and sharing behaviors; The interest topic evolution model (502) is based on long-term user behavior data and uses topic models or deep interest networks to extract users’ interest tags in different dimensions and their intensity trends over time. A short-term conversation context-aware (503) is used to capture the sequence of behaviors within the current search session and construct a short-term contextualized preference vector.

7. The intelligent search system for reading software based on natural language processing technology according to claim 1, characterized in that, The user behavior collection and profile modeling module (500) is used to construct short-term conversation context vectors and long-term interest topic evolution vectors, and input the constructed vectors as features into the multi-dimensional ranking model (403) and dynamic content planner (401) of the personalized ranking and content generation engine (400).

8. The intelligent search system for reading software based on natural language processing technology according to claim 5, characterized in that, The dynamic summary generator (402) adopts an encoder-decoder architecture. Based on the intent and key entities in the structured query object, it extracts or generates a coherent summary text focused on the query point from the top-K retrieval result documents. The ranking factors of the multi-dimensional ranking model (403) include at least: query-document semantic relevance score, document preference based on user history clicks and reading time, document timeliness weight, document authority weight, social collaborative filtering recommendation score, and context awareness.

9. The intelligent search system for reading software based on natural language processing technology according to claim 1, characterized in that, The consistency check includes: extracting entities and relation assertions from the generated content and matching them with reference fragments bound to Must-Anchor Fields; when the matching fails or conflicts with Must-Hold Constraints, the consistency check is deemed to have failed.

10. An intelligent search method for reading software based on natural language processing technology, applied to the system described in any one of claims 1-9, characterized in that, Includes the following steps: S1: Receives the original search request initiated by the user through the user interaction interface; S2: The query understanding engine performs deep semantic analysis on the original search request and generates a structured query object; S3: Document semantic indexing and retrieval engine performs structural anchor semantic retrieval based on structured query objects to obtain a preliminary retrieval result set; S4: The personalized ranking and content generation engine combines user behavior profiles to re-rank the initial search results set and generate auxiliary content under the RCS constraint of the reading context evidence constraint architecture. S5: Present the final sorted result list and generated auxiliary content to the user; when the consistency check fails, execute the rollback strategy.

Images