Jacoi knowledge base self-adaptive question answering system based on knowledge graph

Abstract

The invention relates to the technical field of Jacoi question answering, and discloses a Jacoi knowledge base self-adaptive question answering system based on a knowledge graph. The system comprises five modules: a knowledge graph construction module for analyzing a Jacoi question bank text, extracting keywords and knowledge point entities and constructing association; the user portrait generation module collects historical answer data of a user, and generates a portrait vector containing knowledge point mastering degree weights and answer preference features in combination with knowledge graph nodes; the question intention recognition module receives a user question, extracts an intention keyword and a question focus, and maps the intention keyword and the question focus to a knowledge graph to determine a knowledge point range; the answer retrieval module traverses the knowledge graph to retrieve the answer analysis node with the highest association degree, and adjusts the detail degree of the answer in combination with the user portrait; and the feedback optimization module records the user satisfaction score and the subsequent answer performance, updates the answer preference characteristics and corrects the knowledge point mastering degree weight. The system can realize accurate and personalized Jiaoxian question answering, and is adaptive to individual differences of examinees.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of IELTS answering, in particular to an IELTS knowledge base adaptive answering system based on a knowledge graph. BACKGROUND

[0002] During IELTS preparation, the needs of examinees for answering are diverse and personalized. Current IELTS answering tools on the market mostly focus on fixed question banks and preset analysis, lacking effective adaptation to individual differences of examinees. Traditional answering systems usually use keyword matching to retrieve answers, which cannot accurately understand the true intention of examinee questions, often resulting in mismatched answers and questions, and examinees cannot obtain effective answers.

[0003] Different examinees have significant differences in knowledge point mastery, answering habits, and learning preferences. Some examinees may have weak knowledge of map questions in the listening section, while others have difficulty with judgment questions in the reading section. Some examinees answer quickly, but their accuracy needs improvement, while others are meticulous but time-consuming. However, existing answering systems do not fully consider these individual differences, and the provided answer analysis is often uniform and standardized, failing to adjust the level of detail and presentation of analysis according to the specific circumstances of examinees. For examinees with good knowledge, overly basic and lengthy analysis wastes valuable preparation time, while for examinees with weak foundation, concise analysis may not meet their understanding needs, making it difficult for them to overcome knowledge difficulties.

[0004] Existing answering systems lack effective feedback mechanisms, making it difficult to continuously track and optimize the learning process of examinees. The system cannot adjust settings related to answering preferences based on examinee satisfaction with answers, nor can it dynamically correct knowledge point mastery assessments based on subsequent answering performance, making it difficult for answering services to adapt to the learning progress of examinees, and making it difficult to achieve personalized learning guidance, thereby affecting the preparation efficiency and learning effectiveness of examinees. SUMMARY

[0005] The present application aims to provide an IELTS knowledge base adaptive answering system based on a knowledge graph to solve the problems raised in the background.

[0006] To achieve the above purpose, the present application provides an IELTS knowledge base adaptive answering system based on a knowledge graph, which comprises:

[0007] A knowledge graph construction module is used to analyze question texts and answer analysis texts in the IELTS question bank, extract keywords in the questions and knowledge point entities in the analysis, and construct knowledge graph nodes and edges based on semantic associations between keywords and knowledge point entities.

[0008] a user portrait generation module, which collects the correct answer rate, the answer time length and the knowledge point mastery degree in the user historical answer record, and generates a user portrait vector in combination with a knowledge graph node, the user portrait vector including a knowledge point mastery degree weight and an answer preference feature;

[0009] a question intention recognition module, which receives a user inputted question text for answering, extracts an intention keyword and a question focus in the question text through a semantic analysis engine, and maps the intention keyword to a knowledge graph node to determine a knowledge point range to which the question belongs;

[0010] an answer retrieval module, which traverses a knowledge graph edge according to the knowledge point range to which the question belongs, retrieves an answer analysis node with the highest correlation degree with the question focus, and adjusts the detail degree of the answer in combination with the knowledge point mastery degree weight in the user portrait vector;

[0011] a feedback optimization module, which is used for recording a user satisfaction score for the answer and a subsequent answer performance, updating the answer preference feature in the user portrait vector according to the satisfaction score, and correcting the knowledge point mastery degree weight according to the subsequent answer performance.

[0012] Preferably, the knowledge graph construction module comprises:

[0013] a question structure analysis unit, which is used for sentence processing and syntax tree analysis on the IELTS question text, and extracts a main component of the question stem and an interference item feature of the option;

[0014] a knowledge point entity extraction unit, which identifies a defining description, an example explanation and a comparison relationship from the answer analysis text, and marks a core term in the defining description as a knowledge point entity;

[0015] a semantic relationship labeling unit, which is used for calculating a context similarity between the main component of the question stem and the knowledge point entity, establishing a semantic correlation edge and labeling a relationship type according to a similarity threshold.

[0016] Preferably, the user portrait generation module comprises:

[0017] a time sequence behavior analysis unit, which arranges the user historical answer record in time sequence, and detects a correct answer rate change trend of the same knowledge point;

[0018] a weight dynamic calculation unit, which dynamically adjusts the knowledge point mastery degree weight according to the correct answer rate change trend and an answer time length decay coefficient;

[0019] a preference feature extraction unit, which counts the repeated error question types and the answer analysis click behaviors of the user, and generates an answer preference feature vector.

[0020] Preferably, the question intention recognition module comprises:

[0021] The question focus localization unit is used to perform dependency parsing on the question text input by the user and extract phrase components modified by interrogative pronouns and negative words;

[0022] The intent extension unit is used to perform multi-hop traversal of the question focus and knowledge graph nodes to extend synonym nodes and hierarchical concept nodes.

[0023] The knowledge point range filtering unit selects the top N knowledge points with the highest weight as the scope of the question based on the expanded set of nodes.

[0024] Preferably, the answer retrieval module includes:

[0025] The relevance scoring unit is used to calculate the semantic similarity and path distance between the question focus and the answer parsing node, and generate a comprehensive relevance score;

[0026] The answer adaptation unit selects the basic, advanced, or extended version of the answer parsing text based on the knowledge point mastery weight in the user profile vector.

[0027] The presentation optimization unit sorts the terminology explanations and examples in the answers according to the answer preference characteristics.

[0028] Preferably, the feedback optimization module includes:

[0029] The satisfaction mapping unit is used to map user ratings to the weight values ​​of knowledge graph edges, thereby reducing the retrieval priority of low-rated answers.

[0030] The behavior correction unit is used to compare the user's subsequent answer performance with the expected level of mastery and to compensate for errors in the weighting of knowledge point mastery.

[0031] The profile update unit is used to merge newly generated answer preference features with historical feature vectors to generate an updated user profile.

[0032] Preferably, the knowledge graph construction module further includes:

[0033] Cross-question type related units are used to analyze different ways of testing the same knowledge point in listening, reading and writing questions, and to establish cross-question type related edges;

[0034] The difficulty labeling unit adds difficulty level tags to knowledge point nodes based on official question bank annotations and user answer statistics.

[0035] Preferably, the user profile generation module further includes:

[0036] The group comparison unit is used to perform difference analysis between individual user profiles and the characteristics of groups with the same score range, and to identify abnormal deviation characteristics;

[0037] The calibration suggestion unit generates knowledge point review suggestions based on the group comparison results and writes them into the user profile vector.

[0038] Preferably, the question intent recognition module further includes:

[0039] The ambiguity resolution unit is used to detect polysemous word conflicts in the question text and select the most likely semantic meaning by combining the user's historical answer records.

[0040] The follow-up question generation unit automatically generates selective questions within a limited scope when the focus of the question cannot be clearly mapped.

[0041] Preferably, the answer retrieval module further includes:

[0042] The vulnerability detection unit is used to periodically scan isolated nodes and contradictory edges in the knowledge graph, triggering a manual review process.

[0043] The version tracking unit is used to record the update history of the answer parsing node and return the corresponding version of the answer based on the usage time in the user profile.

[0044] Compared with the prior art, the beneficial effects of the present invention are:

[0045] By constructing a knowledge graph, the system analyzes the question text and answer explanation text in the IELTS question bank, extracts keywords and knowledge point entities, and establishes connections, breaking through the limitations of traditional Q&A systems that rely on single keyword matching. The semantic relationship structure of nodes and edges in the knowledge graph can more comprehensively and accurately present the IELTS knowledge system, making the system's organization of knowledge points more systematic. This provides structured knowledge support for subsequent accurate Q&A, allowing test takers to more clearly understand the connections between knowledge points when obtaining answers, rather than being limited to isolated solutions to individual questions.

[0046] The user profile generation module collects information such as answer accuracy, answer time, and knowledge point mastery from users' historical answer records. It then combines this information with knowledge graph nodes to generate user profile vectors that include knowledge point mastery weights and answer preference characteristics, achieving a precise characterization of individual student differences. This characterization is not simply a matter of information aggregation, but rather a deep integration of student learning data with their knowledge system. This allows the system to clearly understand each student's strengths and weaknesses, such as identifying which knowledge points a student has mastered well, which knowledge points they lack, and whether they prefer quick and concise explanations or detailed and in-depth explanations. This lays the foundation for providing personalized services in the future.

[0047] The question intent recognition module uses a semantic parsing engine to extract intent keywords and key questions from user question texts and maps them to knowledge graph nodes to determine the scope of knowledge points, effectively solving the problem of inaccurate intent understanding in traditional systems. The semantic parsing engine delves into the true needs behind the question, rather than merely matching surface-level text. Even if a candidate's question is not well-defined or is ambiguous, the system can accurately pinpoint the relevant knowledge area by mapping it to knowledge graph nodes, preventing mismatches between answers and questions and ensuring that candidates receive answers highly relevant to their questions.

[0048] The answer retrieval module traverses the knowledge graph edges based on the scope of knowledge points, searching for the answer analysis node with the highest relevance. It also adjusts the level of detail in the answer analysis by incorporating the knowledge point mastery weights from the user profile vector, achieving personalized adaptation of the answer analysis. For test-takers with a high level of knowledge point mastery, the system automatically simplifies the analysis, skipping basic parts and focusing directly on the key points of the question, saving test-takers time. For test-takers with weak knowledge points, the system increases the depth and breadth of the analysis, breaking down the knowledge points, problem-solving approaches, and techniques involved in the question in detail, helping test-takers gradually understand and master the difficult content. This ensures that test-takers at different levels can receive analysis services tailored to their needs.

[0049] The feedback optimization module records user satisfaction ratings for answers and subsequent performance on subsequent questions. These are used to update the answer preference features in the user profile vector and correct the weights of knowledge point mastery, forming a closed loop of continuous optimization. When a test-taker's satisfaction with an answer is low, the system can adjust the presentation style of subsequent explanations accordingly. For example, if a test-taker finds the explanation too brief, subsequent explanations for that test-taker will add more detail. Simultaneously, based on the test-taker's subsequent performance, the system can dynamically adjust its assessment of their knowledge point mastery. If a test-taker's accuracy rate on questions related to a particular knowledge point increases, the system will correspondingly increase the mastery weight of that knowledge point and reduce unnecessary detailed explanations; conversely, it will strengthen the explanation of that knowledge point. This allows the Q&A service to continuously adjust along with the test-taker's learning progress, consistently aligning with their learning status and providing more targeted learning support to help them prepare for exams efficiently. Attached Figure Description

[0050] Figure 1 This is a schematic diagram illustrating the working principle of the IELTS knowledge base adaptive Q&A system based on knowledge graphs as described in this invention.

[0051] Figure 2 A flowchart illustrating the workflow of the knowledge graph construction module;

[0052] Figure 3 A flowchart illustrating how the problem intent recognition module works. Detailed Implementation

[0053] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] Please see Figure 1 This invention provides an adaptive Q&A system for IELTS knowledge base based on knowledge graphs. The system includes: a knowledge graph construction module, a user profile generation module, a question intent recognition module, an answer retrieval module, and a feedback optimization module. Specific implementation methods are as follows:

[0055] The knowledge graph construction module processes the question text and answer explanation text from the IELTS question bank. It uses natural language processing technology to parse the text content, extracting keywords from the questions and knowledge point entities from the explanations. Keywords include core terms from the question stem and features of the answer choices, while knowledge point entities are derived from the definitional descriptions and examples in the answer explanations. Based on semantic relationships, it constructs knowledge graph nodes and edges. Nodes represent knowledge points or question elements, and edges represent semantic relationships such as membership or causality. The user profile generation module collects users' historical answer records, including answer accuracy, answer time, and knowledge point mastery data. It analyzes this data to generate a user profile vector, which stores knowledge point mastery weights and answer preference features in numerical form. Knowledge point mastery weights reflect the user's proficiency with specific knowledge points, while answer preference features cover the types of questions the user frequently gets wrong and their answer explanation browsing habits. The question intent recognition module receives user-input question text, processes the text using a semantic parsing engine, and extracts intent keywords and question focus. Intent keywords are core words in the question that point to the knowledge point, while question focus is the specific part of the user's confusion. These are mapped to knowledge graph nodes to determine the scope of the knowledge point to which the question belongs. The answer retrieval module traverses the knowledge graph edges based on a defined range of knowledge points, retrieving the answer parsing node with the highest relevance to the question's focus. This relevance is calculated based on semantic similarity and path distance, and the level of detail in the answer is adjusted by incorporating the knowledge point mastery weights from the user profile vector. For example, more basic explanations are provided to users with low mastery levels. The feedback optimization module records user satisfaction ratings for answers and subsequent answering performance. Satisfaction ratings are used to update answering preference features in the user profile vector, while subsequent answering performance adjusts the knowledge point mastery weights, ensuring the system adapts to the user's learning progress. All modules are connected via data interfaces. The knowledge graph construction module outputs graph data for use by other modules, and the user profile generation module and feedback optimization module form a closed-loop optimization mechanism.

[0056] Example 1: See Figure 2In the process of constructing the knowledge graph, the question structure analysis unit first performs in-depth processing of the IELTS question text. Its sentence segmentation process breaks down the complete question content into independent semantic units based on punctuation and grammatical structure. For example, a complex long sentence may be broken down into several short clauses for subsequent analysis. Then, the syntax tree analysis tool performs syntactic analysis on each clause, identifying the main components of the question stem, such as the subject, predicate, and object, thereby extracting the core testing intent of the question. At the same time, this unit carefully analyzes each option in the question and extracts the characteristics of the distractors. These characteristics usually include subtle lexical differences between options, deliberately set logical traps, or common cognitive biases. The knowledge entity extraction unit focuses on processing the answer parsing text. It identifies the defining descriptions in the text through pattern matching and rule engines. These descriptions often contain marker words such as "refers to" or "defined as", which clearly point to a core concept. For examples in the parsing, the unit captures the content after introductory words such as "for example" or "likely" and associates them with the corresponding definitions. Comparison relationships are identified by analyzing the context connected by transition words such as "however" or "but". Finally, the unit marks the core terms in the defining descriptions as formal knowledge entities. These entities constitute the most basic nodes in the knowledge graph. The semantic relation annotation unit is responsible for establishing connections between question elements and knowledge point entities. It evaluates the strength of association by calculating the contextual similarity between keywords in the main components of the question stem and the labeled knowledge point entities in the semantic space. This similarity calculation not only relies on the surface matching of words, but also delves into the semantic level to consider their relevance. When the calculated similarity exceeds a preset threshold, the system will automatically establish a semantic association edge between the corresponding nodes and label this edge with a specific relationship type. For example, the "examine" relationship indicates that the question mainly tests this knowledge point, the "involves" relationship indicates that the knowledge point is indirectly used in the problem-solving process, and the "explain" relationship indicates that the knowledge point is the core basis of the question's answer.

[0057] The cross-question-type related unit is designed to break down the barriers between different question types in the IELTS test, such as Listening, Reading, and Writing. It systematically analyzes the content of questions under different question types, and finds different angles and forms of testing the same knowledge point. For example, a grammar point may appear in the Listening section as a pronunciation identification question, in the Reading section as a question to understand complex sentences, and in the Writing section as a question to apply correctly. This unit establishes cross-question-type related edges by comparing the core knowledge point entities mentioned in the analysis of different questions. This connection allows the system to understand the application scenarios of knowledge points from a more macro perspective and provides users with a more comprehensive learning view. The difficulty labeling unit adds difficulty level information to each knowledge point node in the knowledge graph. It combines two data sources for judgment. On the one hand, it may have difficulty labels or score range information of questions in the official question bank, which are relatively objective benchmarks. On the other hand, it comes from a large amount of anonymous answer data collected by the system, including statistical analysis of the average answer accuracy rate and the answer time of users for this knowledge point. By integrating these two types of information, this unit assigns a difficulty level label to each knowledge point node, such as "basic", "intermediate" or "advanced". This label dynamically reflects the complexity of the knowledge point itself and the general level of mastery of it by the user group.

[0058] In the specific process of entity extraction, the system constructs a domain dictionary and a set of rule templates to assist in identification. The domain dictionary contains high-frequency terms and core concepts from common subject areas in the IELTS test, while the rule templates define various language patterns to capture the context in which knowledge point entities appear. When processing a new answer analysis, the unit first performs basic part-of-speech tagging and named entity recognition to filter candidate terms. Then, it combines the rule templates and the dictionary for precise matching and filtering to ultimately determine the core terms that need to be labeled as knowledge point entities. For the establishment of semantic relationships, the system adopts a deep semantic matching method based on a pre-trained language model. It transforms the main components of the question stem and knowledge point entities into high-dimensional vector representations, and then calculates the cosine similarity between these two vectors as a measure of contextual similarity. This similarity value reflects the semantic closeness between the two and is the main basis for deciding whether to establish a connection edge and what type of relationship to label. The implementation of cross-question type associations relies on the normalization of knowledge point entities in different contexts. Even though different question types may express the same knowledge point differently, the system ensures accurate association through its core terminology and semantic IDs. The establishment of association edges not only considers the similarity of knowledge points but also analyzes the differences in the depth of examination, thus making the associated information richer and more comprehensive. Difficulty labeling is a continuously updated process. As more users use the system, user answer statistics will be continuously updated, and the difficulty level of knowledge points may also be dynamically adjusted to reflect the latest user learning status and changes in question difficulty.

[0059] Example 2: The implementation of the user profile generation module begins with the time-series behavior analysis unit systematically organizing the user's historical answer records. This unit arranges all the user's answer events into a continuous behavior sequence according to the order of timestamps. Each record contains fields such as question identifier, user answer, correct answer, time spent answering the question, and whether the answer result is correct. Through in-depth analysis of this sequence, the unit can detect the changing trend of the user's answer accuracy when facing a series of questions on the same knowledge point. This trend analysis not only focuses on the simple rise and fall of the accuracy rate, but also examines its fluctuation range and stability. For example, a user may have a low accuracy rate on a certain knowledge point in the early stage, but it shows a steady upward trend as the learning process progresses, or it may show an unstable situation with repeated fluctuations. The dynamic weight calculation unit dynamically adjusts the mastery weight of each knowledge point based on the accuracy trend output by the time-series behavior analysis unit, combined with the answering time factor. This unit introduces the concept of a time decay coefficient, where recent answering behavior has a greater impact on the weight than long-term behavior, allowing the user profile to more sensitively reflect the current learning status. Regarding the handling of answering time, the unit assesses the changes in the user's problem-solving efficiency. Quick and correct answers usually indicate improved proficiency, while excessive time consumption, even if answered correctly, may suggest uncertainty in understanding. Weight calculation is a dynamic process that comprehensively considers accuracy trends, answering timeliness, and time decay effects. The preference feature extraction unit analyzes users' historical behavioral data using statistical methods, focusing on identifying two typical patterns: the types of questions users repeatedly get wrong and the click behavior characteristics of users when viewing answer explanations. The statistical analysis of question types covers various formats of the IELTS test, such as multiple choice, true / false, fill-in-the-blank, and graph / chart questions. By analyzing the concentration of errors users make on different question types, it is possible to determine the user's potential weaknesses in problem-solving strategies or fixed mindsets. The click behavior of answer explanations records the details of the user's interaction with the system, such as whether the user tends to click to view detailed explanations, the length of time spent on the explanation page, and whether they frequently switch between the basic and extended versions of the answers. These behavioral data are quantified into feature values, which together constitute a multi-dimensional vector reflecting the user's learning habits and preferences.

[0060] The group comparison unit examines individual user profile data within a broader context. It extracts a group sample from the system's user database that is similar to the target user's current score range, calculates the average mastery weight of this group on each knowledge point, and identifies common answering preferences to form a group feature profile. Subsequently, the unit performs a point-by-point difference analysis between the individual user's profile vector and this group feature profile, calculating the degree of deviation in each feature dimension. This difference analysis helps identify certain abnormal deviations in a user's characteristics, such as a significantly lower mastery of a certain knowledge point than the average level of users at the same level, or a much higher preference for a certain question type than the group norm. These abnormal signals may reveal unique learning difficulties or cognitive characteristics of the user. The calibration suggestion unit generates targeted knowledge point review suggestions based on the differences found by the group comparison unit. If the system detects that a user's mastery weight of the knowledge point "relative clauses" is significantly lower than that of the group with the same score, and this knowledge point is marked as a high-frequency test point in the knowledge graph, then the unit will generate a suggestion to prioritize the review of this knowledge point. At the same time, it may recommend related practice questions or learning materials. These specific review suggestions will be transformed into structured data and directly written into the suggestion field of the user profile vector, thereby affecting the behavior of the subsequent answer retrieval module, so that it can consciously strengthen the explanation of weak links when providing Q&A.

[0061] In the specific implementation of time-series behavioral analysis, the system employs time series analysis techniques to capture patterns of mastery changes. For example, a sliding window algorithm is used to observe the difference between recent answer accuracy and earlier performance. The window size can be dynamically adjusted according to the learning stage; a smaller window may be used for new users to quickly respond to changes, while a larger window may be used for experienced users to observe long-term trends. The analysis results not only output a simple trend direction (rising / falling / stable) but also provide confidence or strength indices for the trend, providing more refined input for weight calculation. The dynamic weight calculation process avoids using fixed formulas and instead employs a combination of rule-based and lightweight machine learning models. The system pre-sets a series of heuristic rules to handle common patterns, such as "three consecutive correct answers with reduced time result in a significant increase in weight." Simultaneously, a regression model is trained based on a large amount of user data to comprehensively evaluate the combined impact of various behavioral characteristics on weight changes, making weight adjustments more closely reflect real learning patterns. Preference feature extraction not only focuses on macro-level statistical results but also emphasizes micro-level behavioral sequence patterns. For example, does a user habitually look at the options before reading the question stem, or does they immediately check the explanation after answering incorrectly? These behavioral sequences can be pattern-recognized using sequence modeling techniques such as Hidden Markov Models (HMMs) to extract deeper-level preference features. After feature vector normalization, behavioral indicators of different dimensions can be compared and calculated together. When selecting comparison groups, the group comparison unit employs a multi-dimensional matching strategy, considering not only the total score range but also factors such as the user's learning time and target score range to find a more comparable reference group. Difference analysis uses methods such as Z-scores to quantify the degree of deviation and sets reasonable thresholds to determine what constitutes "abnormal deviation," avoiding oversensitivity to normal fluctuations.

[0062] The generation of calibration suggestions requires close interaction with the knowledge graph. The unit queries the knowledge graph for upstream and downstream related knowledge points based on identified weak knowledge points, thus providing systematic review path suggestions rather than isolated points. For example, when suggesting a user review "relative clauses," it might simultaneously suggest reviewing their superordinate concept "clause" and subordinate concept "non-restrictive relative clauses." Once the content and priority of these suggestions are written into the user profile, a dynamic and personalized learning path graph is effectively constructed. The user profile vector itself uses a scalable data structure for storage and updating, containing both a static snapshot of current features and historical changes within a certain time window. This allows the system to not only make decisions based on the current state but also predict potential user needs by referencing recent trends. The entire generation and updating of the user profile is a continuously running background process, ensuring that users receive a personalized response based on their latest state each time they interact with the system.

[0063] Example 3: SeeFigure 3 The implementation of the question intent recognition module revolves around accurately understanding the natural language questions input by users. Its core workflow begins with the question focus localization unit's deep analysis of the question text. This unit uses dependency parsing technology to structurally decompose the user-input sentence, identifying the grammatical dependencies between its components, such as subject-verb, verb-object, and attributive-head relationships. Based on this, the unit focuses on locating phrases modified by interrogative pronouns or negative words. These modified components often directly point to the core content of the user's question and are candidate questions for focus. The intent expansion unit begins its work after the question focus is initially located. It uses the identified focus words or phrases as starting points and performs multi-hop traversal within a pre-constructed knowledge graph. The traversal follows the semantic edges of the graph, aiming to discover nodes semantically related to the initial focus. These connections include synonyms, hyponyms, and partial-whole relationships. Through multi-hop traversal, the system can expand to a set of related nodes, thereby gaining a more comprehensive understanding of the knowledge scope that the user's question may involve. The knowledge point scope filtering unit filters and sorts the expanded set of nodes. This unit calculates the semantic relevance of each node in the set relative to the initial question focus, while also considering the node's importance in the user's current learning context. Ultimately, it selects the top N knowledge points with the highest weights, forming the core scope of the question. This process maps the ambiguous user question to a relatively precise subgraph within the knowledge system. The ambiguity resolution unit specifically handles the inherent polysemy in natural language. This unit is activated when keywords in the question text correspond to multiple nodes with different concepts in the knowledge graph. It infers the most likely semantic meaning by analyzing the user's historical answer records. For example, if a user has recently frequently practiced questions related to "literature," the system will prioritize the literary imagery of "riverbank" rather than the financial meaning of "bank" when the question contains the polysemous word "bank." The follow-up question generation unit, as a backup mechanism, is activated when the system cannot confidently map the question focus to a specific knowledge point. Based on node information related to the ambiguous focus in the knowledge graph, it automatically generates one or more selective questions to guide the user to provide more specific information, thereby narrowing the query scope.

[0064] During the knowledge point selection process, the system employs a comprehensive scoring mechanism to determine the final weight of each node. The calculation method for this mechanism is as follows:

[0065] S node =α·R semantic +β·C context +γ·P user

[0066] Wherein: S nodeThis represents the overall screening score of a node in a knowledge graph. The higher the score, the greater the likelihood that the node will be included in the problem scope. R semantic C is a semantic relevance score calculated using graph embedding techniques. It quantifies the distance between the node and the initial focus of the question in the vector space. context It is the context importance score, which evaluates the centrality or hub role of a node in the current query context of the entire knowledge graph. user This is a personalized user score, reflecting the relevance of the node to the user's historical focus and weaknesses in their profile. α, β, and γ are configurable hyperparameters used to balance the weights of these three factors, and their values ​​are determined through offline tuning. This formula ensures that the selection process considers both the semantics of the question itself and the structural characteristics of knowledge, as well as the user's individual learning status.

[0067] The implementation of the interrogative focus localization unit relies on a pre-trained language analysis model that accurately labels the part of speech of each word in a sentence and the dependencies between words. The unit extracts the focus from the analysis results according to a set of predefined rules. For example, the rules might stipulate that if a noun phrase is directly modified by an interrogative pronoun, then the noun phrase is marked as an interrogative focus. For more complex sentence structures, such as interrogative sentences with clauses, the unit recursively analyzes the clause structure to ensure that nested focus information is not missed. The multi-hop traversal strategy of the intent extension unit is usually breadth-first. It starts from the initial node, first visits all directly connected one-hop neighbor nodes, then visits the neighbors of these neighbor nodes, i.e., two-hop nodes, and so on. The number of hops is usually limited to two to three to avoid introducing too much noise. After each hop, the system performs preliminary filtering on the collected nodes, eliminating obviously irrelevant types.

[0068] The core of the ambiguity resolution unit is a context-based disambiguation algorithm. This algorithm extracts contextual features from the current question text and historical interest features from the user profile, forming a feature vector. This vector is then matched against the feature representation of each candidate semantic node in the knowledge graph, calculating the matching degree. Finally, the semantic node with the highest matching degree is selected. The entire disambiguation process aims to be completed quickly and seamlessly for the user. The follow-up question generation unit emphasizes the naturalness and effectiveness of the interaction. The follow-up questions it generates are typically in the form of multiple-choice questions, such as "Do you want to ask about financial knowledge or geographical knowledge about 'bank'?". The options directly originate from the concepts pointed to by different edge connections of polysemous word nodes in the knowledge graph, ensuring that each option corresponds to a clear semantic path existing in the knowledge base. The entire question intent recognition module operates as a pipeline system, with each unit processing input sequentially. The output of the previous unit serves as the input of the next unit. If, at the end of the pipeline, the system's confidence in understanding the user's intent is still below a certain threshold, it will trigger the follow-up question generation unit to interact with the user a second time, rather than returning a potentially inaccurate answer. This design balances the efficiency of automatic processing with the accuracy of results, ensuring that the system can robustly handle various complex and ambiguous user inputs.

[0069] Example 4: The implementation of the answer retrieval module focuses on locating and dynamically adapting the most suitable answer content from the knowledge graph. The relevance scoring unit first processes the question focus and candidate knowledge point range output by the question intent recognition module. This unit calculates the semantic similarity between the question focus and each answer parsing node in the knowledge graph. The semantic similarity is obtained by comparing the vector representations of the question focus text and the summary text of the answer parsing node. At the same time, the unit calculates the shortest path distance from the question focus node to the answer parsing node in the knowledge graph. The shorter the path distance, the more direct the concept association usually is. Finally, the unit generates a comprehensive relevance score, which is a weighted combination of semantic similarity and path distance. A high score means that the answer parsing node is highly relevant to the user's question. The answer adaptation unit determines the level of detail and depth of knowledge returned based on the knowledge point mastery weights stored in the user profile vector. For knowledge points with low mastery weights, the unit selects a basic version of the answer explanation text. The basic version focuses on the core definition and direct application of the concept, avoiding the introduction of complex background knowledge. For knowledge points with medium mastery weights, the unit triggers an advanced version of the answer. The advanced version adds typical examples, common variations, and analysis of common mistakes on top of the basic version. When the system detects that the user has a high mastery weight for a certain knowledge point, it provides an extended version of the answer. The extended version may include the historical origin of the knowledge point, comparison and analysis with other concepts, and advanced applications in academic scenarios. The presentation optimization unit is responsible for personalizing the specific content organization of the selected answer. It refers to the answer preference characteristics in the user profile. For example, if the characteristics show that the user tends to understand the terminology first and then look at the examples, the unit will sort the answer paragraphs, placing the terminology explanations first and the example explanations later. Or, if the user has a history of frequently clicking on charts and graphs, the unit will try to embed relevant diagrams or structural diagrams in the answer.

[0070] The vulnerability detection unit operates periodically as a quality monitoring component. It scans the entire knowledge graph structure using predefined rules to look for patterns that may pose problems. For example, isolated nodes are those nodes without any incoming or outgoing edges. These may be caused by data extraction errors or the fact that new knowledge points have not been fully associated. Conflicting edges are those edges that have semantic conflicts. For example, the same knowledge point node may be marked as "examined" in question type A and also as "not examined" in question type A. When such problems are detected, the unit records the problem details and triggers a manual review process, marking the problematic nodes and edges as "pending review" and notifying the knowledge base administrator. The version tracking unit maintains the complete update history of the answer parsing nodes. Each modification to the node content is recorded with a timestamp, modifier, and content change summary. When an answer needs to be provided to a user, the unit compares it with the last learning time or the last contact time of the knowledge point recorded in the user profile, and then selects an answer version from the version history that has not undergone significant content changes since the user's learning time. This mechanism ensures that the knowledge received by the user is continuous with the cognitive framework they have previously established, avoiding confusion caused by sudden changes in the answer definition.

[0071] Consider a specific example, referring to Table 1, where a user asks: "In IELTS Writing Task 2, how do you distinguish the usage of 'advantage' and 'benefit'?" The question intent recognition module has identified the focus as "distinguishing the usage of 'advantage' and 'benefit'" and determined it falls within the "vocabulary discrimination" knowledge point. The relevance scoring unit begins searching for relevant answer analysis nodes in the "vocabulary discrimination" subgraph. Assume there are three relevant nodes in the knowledge graph: Node A (Title: Basic Definitions of Advantage and Benefit), Node B (Title: Semantic Weight Comparison of Advantage and Benefit), and Node C (Title: Examples of Collocations of Advantage and Benefit in Academic Writing). The unit calculates the relevance between the question focus and each node.

[0072] Table 1: Relevance Scoring of Answer Analysis Nodes

[0073]

[0074] The relevance scoring unit employs an internal algorithm that maximizes the contribution of semantic similarity, but this contribution diminishes with increasing path distance. According to the table above, node A receives the highest score because it directly answers the basic definition needed for "distinguishing" and has the shortest path. The system queries the user profile and finds that the user's mastery of "vocabulary discrimination" knowledge points has a weight of 0.4 (range 0-1, lower values ​​indicate weaker mastery). Therefore, the unit decides not to directly provide node A (basic definition), which has the highest relevance, as the user may need comparison and examples to deepen their understanding. The unit ultimately selects node B (semantic weight comparison) as the core answer because it also has high relevance and its content depth is more suitable for users with moderate mastery. Simultaneously, the unit extracts the most crucial definition sentence from node A and selects a typical collocation example from node C, combining them to form the final answer. The presentation optimization unit checks the user profile and finds that the user's answer preference characteristics show that when reviewing vocabulary answers, the user has a 70% probability of clicking to view content in the form of a "comparison table." Therefore, the unit formats the original text retrieved from node B, transforming the comparative descriptions into a clear table. The vulnerability detection unit may discover during periodic scans that a node related to "digital literacy" in the graph is isolated and not yet associated with any writing or reading question type. The unit will mark this node and add it to the review queue. Version tracking records show that the example sentences for "benefit" in node B were updated three months ago, and the current user learned the relevant knowledge points last month. Therefore, the unit will ensure that the version returned to the user is from three months ago to maintain the continuity of the user's learning experience, unless the new version involves critical bug fixes.

[0075] Example 5: The implementation of the feedback optimization module constitutes a closed loop of system self-improvement and personalized adjustment. The satisfaction mapping unit is responsible for processing users' direct feedback on the provided answers, usually presented in the form of ratings from 1 to 5 stars. When a user submits a rating, this unit starts working, mapping the numerical rating to the weight adjustment of a specific edge in the knowledge graph. For example, a semantic edge connecting the user's question focus to the final presented answer node. If the user gives a 5-star rating, the weight of this edge will receive a positive boost, increasing the priority of this answer node being retrieved in similar queries in the future. Conversely, if the user gives a 1-star or 2-star low rating, the weight of the edge related to this answer will be correspondingly weakened, reducing its chance of being recommended later. This mapping is not a simple linear relationship. For a medium rating such as 3 stars, the weight adjustment may be very small to avoid unnecessary frequent fluctuations in the system's recommendation strategy. The behavior correction unit focuses on longer-term and indirect user feedback, namely the user's subsequent answering performance. This unit sets up an observation window to continuously track the user's answering performance on questions related to the knowledge points involved in that answer for a period of time after the user receives and learns a certain answer. The unit compares the user's actual performance with the expected mastery predicted by the system based on the user's original profile. If the user's actual performance is significantly better than expected, it indicates that the previous answer may have been very effective or that the user has achieved a breakthrough through other means, and the system will appropriately increase the weight of the mastery of that knowledge point. If the actual performance is lower than expected, it may indicate that the answer was not fully understood or that the knowledge point itself is more complex, and the system will perform a downward weight adjustment. The profile update unit is the final integration point of feedback information. It is responsible for fusing the emerging answering preference features generated by this interaction with the historical feature vector stored in the user profile. Emerging features may include the user's preference for the presentation format of this answer, rating behavior patterns, etc. The fusion process is not a simple replacement, but uses a weighted average or time decay-based fusion algorithm to make recent behavior have a greater impact on the profile, but will not completely cover long-term formed behavior patterns, thereby generating an updated user profile vector that better reflects the user's latest state.

[0076] Consider a concrete example to illustrate the entire process. A user, while practicing the True / False / Not Given questions in IELTS Reading, encountered a question about the impact of climate change on bird migration and asked, "How do I distinguish between F and NG in the reading?" After intent recognition and answer retrieval, the system ultimately provides an answer analysis node containing a decision flowchart that details the key steps and signal words for distinguishing between False and Not Given.

[0077] After reading the answer, the user gave it a 3-star rating (satisfaction score). The satisfaction mapping unit began processing, mapping the neutral 3-star rating to a small negative adjustment factor, such as 0.95. Then, in the knowledge graph, it found the "answer" edge connecting the user's question focus ("difference between F and NG") to the provided answer node ("True / False decision flowchart"), multiplying its current weight by 0.95. This means that for similar queries in the future, the system will still consider this answer, but its priority will be slightly lower than other alternative answer nodes with higher ratings. A few days later, the behavior correction unit initiated a long-term evaluation of the effectiveness of this Q&A session. Over the following week, the system recorded 10 reading comprehension questions with "False" and "Not Given" options completed by the user. The data in the observation window showed that the user's accuracy rate on these 10 questions was 70%, while the system predicted an accuracy rate of approximately 50% based on the user's previous profile (whose mastery of "True / False" questions was relatively low). The actual performance significantly exceeded expectations. Based on this, the behavior correction unit judged that the previously provided flowchart answer may have had a positive effect on the user, or that the user had effectively learned by combining the answer. Therefore, the unit made a positive error compensation to the weight of the user profile's mastery of the "true or false" knowledge point, increasing the weight value from the original 0.5 to 0.65. This correction reflects the user's real progress in the ability of such questions.

[0078] The profile update unit performed the update operation last, noticing a new answer preference feature: during this interaction, the user hovered the mouse over the flowchart section for a much longer time than other text descriptions when viewing the answer, and the user did not click the link to switch to the plain text version of the answer. The unit interpreted this behavior as a clear preference for "visualized" answers. At the same time, the user's rating behavior (giving 3 stars) was also recorded as a mild feedback style. The unit fused these new features with the historical feature vector. The historical features showed that the user did not have a particular preference for chart-based answers in the past, but the strong interaction signal this time increased the weight of the feature "preference for visual explanations" in the updated profile vector.

[0079] Another example might involve handling negative feedback. Suppose the system provides a user with a grammatical answer about the present perfect tense, but the answer is too brief, and the user gives it a 1-star rating. The satisfaction mapping unit will significantly reduce the weight of the edge connecting the question and the brief answer, potentially causing the answer to be "shelved" in the future. Subsequently, in subsequent exercises, the user performs poorly on a series of questions involving the present perfect tense, with an accuracy rate far below the system's expectations. The behavior correction unit will recognize this situation. It will not only lower the weight of the user's mastery of the present perfect tense but may also adjust the assessment of mastery of the broader knowledge point of "tense" in a correlated manner. The profile update unit will record the user's dissatisfaction with the "overly brief answer" and, when integrating new features, will strengthen the user's preference for "detailed explanations," thereby influencing the selection strategy of the answer adaptation unit in the future.

[0080] The entire feedback optimization process enables the system to move beyond a static, one-size-fits-all response model, transforming it into a dynamic system that learns from each interaction and adapts to individual user differences. User profiles are no longer fixed templates but evolving digital mirrors that accurately reflect a user's current knowledge state and learning habits throughout the learning journey. The knowledge graph is no longer a rigid database; its internal connections are continuously recalibrated based on the effectiveness of user feedback, ensuring that the most valuable answers are delivered most smoothly. This closed-loop optimization mechanism is the core of the system's true self-adaptation, making machine-human teaching interactions continuous and growth-oriented.

[0081] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0082] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An IELTS knowledge base adaptive question-answering system based on knowledge graphs, characterized in that, The system includes: The knowledge graph construction module is used to parse the question text and answer analysis text in the IELTS question bank, extract keywords from the questions and knowledge point entities from the analysis, and construct knowledge graph nodes and edges based on the semantic relationship between keywords and knowledge point entities. The user profile generation module collects the user's answer accuracy, answer duration and knowledge point mastery from the user's historical answer records, and generates a user profile vector by combining knowledge graph nodes. The user profile vector includes knowledge point mastery weights and answer preference features. The question intent recognition module receives the question text input by the user, extracts the intent keywords and question focus in the question text through the semantic parsing engine, and maps the intent keywords to knowledge graph nodes to determine the knowledge point range to which the question belongs; The answer retrieval module traverses the knowledge graph edges based on the knowledge point range to which the question belongs, retrieves the answer parsing node with the highest relevance to the question's focus, and adjusts the level of detail of the answer by combining the knowledge point mastery weights in the user profile vector; The feedback optimization module records users' satisfaction ratings for answers and their subsequent answering performance. It updates the answering preference features in the user profile vector based on the satisfaction rating and adjusts the weights of knowledge point mastery based on subsequent answering performance.

2. The IELTS knowledge base adaptive question-answering system based on knowledge graphs according to claim 1, characterized in that, The knowledge graph construction module includes: The question structure analysis unit is used to perform sentence segmentation and syntax tree analysis on IELTS question text, and extract the main components of the question stem and the features of the distractors in the options. The knowledge point entity extraction unit identifies definitional descriptions, illustrative examples, and comparative relationships from the answer analysis text, and marks the core terms in the definitional descriptions as knowledge point entities; The semantic relationship annotation unit is used to calculate the contextual similarity between the main components of the question stem and the knowledge point entities, and to establish semantic association edges and annotate the relationship types based on the similarity threshold.

3. The IELTS knowledge base adaptive question-answering system based on knowledge graphs according to claim 2, characterized in that, The user profile generation module includes: The temporal behavior analysis unit arranges users' historical answer records in chronological order and detects the changing trend of the correct answer rate for the same knowledge points. The weighted dynamic calculation unit dynamically adjusts the weight of knowledge point mastery based on the trend of answer accuracy and the answer time decay coefficient. The preference feature extraction unit counts the types of questions that users repeatedly get wrong and their click behavior on answer explanations, and generates a question preference feature vector.

4. The IELTS knowledge base adaptive question-answering system based on knowledge graphs according to claim 3, characterized in that, The question intent recognition module includes: The question focus localization unit is used to perform dependency parsing on the question text input by the user and extract phrase components modified by interrogative pronouns and negative words; The intent extension unit is used to perform multi-hop traversal of the question focus and knowledge graph nodes to extend synonym nodes and hierarchical concept nodes. The knowledge point range filtering unit selects the top N knowledge points with the highest weight as the scope of the question based on the expanded set of nodes.

5. The IELTS knowledge base adaptive question-answering system based on knowledge graphs according to claim 4, characterized in that, The answer retrieval module includes: The relevance scoring unit is used to calculate the semantic similarity and path distance between the question focus and the answer parsing node, and generate a comprehensive relevance score; The answer adaptation unit selects the basic, advanced, or extended version of the answer parsing text based on the knowledge point mastery weight in the user profile vector. The presentation optimization unit sorts the terminology explanations and examples in the answers according to the answer preference characteristics.

6. The IELTS knowledge base adaptive question-answering system based on knowledge graphs according to claim 5, characterized in that, The feedback optimization module includes: The satisfaction mapping unit is used to map user ratings to the weight values ​​of knowledge graph edges, thereby reducing the retrieval priority of low-rated answers. The behavior correction unit is used to compare the user's subsequent answer performance with the expected level of mastery and to compensate for errors in the weighting of knowledge point mastery. The profile update unit is used to merge newly generated answer preference features with historical feature vectors to generate an updated user profile.

7. The IELTS knowledge base adaptive question-answering system based on knowledge graphs according to claim 6, characterized in that, The knowledge graph construction module also includes: Cross-question type related units are used to analyze different ways of testing the same knowledge point in listening, reading and writing questions, and to establish cross-question type related edges; The difficulty labeling unit adds difficulty level tags to knowledge point nodes based on official question bank annotations and user answer statistics.

8. The IELTS knowledge base adaptive question-answering system based on knowledge graphs according to claim 7, characterized in that, The user profile generation module also includes: The group comparison unit is used to perform difference analysis between individual user profiles and the characteristics of groups with the same score range, and to identify abnormal deviation characteristics; The calibration suggestion unit generates knowledge point review suggestions based on the group comparison results and writes them into the user profile vector.

9. The IELTS knowledge base adaptive question-answering system based on knowledge graphs according to claim 8, characterized in that, The question intent recognition module also includes: The ambiguity resolution unit is used to detect polysemous word conflicts in the question text and select the most likely semantic meaning by combining the user's historical answer records. The follow-up question generation unit automatically generates selective questions within a limited scope when the focus of the question cannot be clearly mapped.

10. The IELTS knowledge base adaptive question-answering system based on knowledge graphs according to claim 9, characterized in that, The answer retrieval module also includes: The vulnerability detection unit is used to periodically scan isolated nodes and contradictory edges in the knowledge graph, triggering a manual review process. The version tracking unit is used to record the update history of the answer parsing node and return the corresponding version of the answer based on the usage time in the user profile.

Images