Individualized health question-answering system based on reinforcement learning

Abstract

The invention relates to the technical field of machine question answering, in particular to a personalized health question answering system based on reinforcement learning, which comprises an input positioning module, an anaphora freezing module, a state injection module, a path screening module and a content splicing module. According to the method, the fuzzy region is marked by identifying semantic features such as subject separation, pointing missing and intention interruption, and unstable content is frozen and screened by combining semantic chain consistency, word order change and context distance judgment; the internal structural integrity and semantic closure of the statement are enhanced by means of subject-called structure reconstruction, verb-object combination recognition and modification chain judgment, and semantic tripping content is further eliminated through subject affiliation matching, upper and lower concept connection and logic consistency screening. Content recombination and linkage optimization are realized based on objective word positioning, tail word sequence adjustment and semantic ending complementation, and semantic undertaking stability, content coherence and reply pertinence in ambiguous input, structure disorder or multi-turn jump scenes are improved.

Description

Technical Field

[0001] This invention relates to the field of machine question answering technology, and in particular to a personalized health question answering system based on reinforcement learning. Background Technology

[0002] The field of machine question answering technology involves achieving intelligent question-answering interaction between humans and machines through natural language processing and artificial intelligence technologies. Its core aspects include understanding question intent, answer generation and matching, contextual reasoning, and knowledge fusion. Based on structured and unstructured data, this technology combines information retrieval, semantic understanding, and knowledge graphs to build automated response systems with semantic recognition and reasoning capabilities. These systems are widely used in scenarios requiring high levels of professional knowledge understanding and expression, such as medicine, law, and finance, forming a systematic research framework encompassing user modeling, multi-turn dialogue management, and knowledge-driven question answering. Traditional personalized health question answering systems, however, rely on user health questions as input and retrieve fixed answers from a pre-set health knowledge base through natural language matching. They typically use keyword matching and semantic similarity calculation to compare user input with existing health question-answer pairs, and then combine rule-based user interest modeling results to rank and select the retrieved answers. This approach cannot dynamically learn and adjust user interaction behavior, making it difficult to adapt to changes in user questions and updated feedback patterns. The personalized health question-answering system based on reinforcement learning addresses the aforementioned technical issues by employing reinforcement learning strategies to model users' historical behavioral data and interaction feedback signals in health question-answering tasks. It dynamically optimizes the question-answering selection path through a policy network and evaluates individual health question-answering matching strategies using a state value function, thereby completing the strategy adjustment and sequence decision generation for personalized question-answering tasks.

[0003] Existing technologies for medical question answering rely on a fixed knowledge base and static matching based on keywords and semantic similarity to process user input. This lacks a dynamic analysis mechanism for the internal structural relationships of questions. When there are ambiguous terms, structural breaks, or sentence jumps, the technology cannot identify and handle potential unclear points, making it difficult to adjust the semantic connection of sentences. When users use non-standard word order, lack clear subjects, or have multiple rounds of context switching, problems such as semantic connection failure, disjointed content, or response fragments deviating from the main point of the question often occur, resulting in poor response continuity, unclear main points, and interrupted interaction, affecting the question answering system's adaptability to complex language input. Summary of the Invention

[0004] To address the technical problems existing in the prior art, embodiments of the present invention provide a personalized health question-answering system based on reinforcement learning. The technical solution is as follows: On the one hand, a personalized health question-answering system based on reinforcement learning is provided, which includes: The input positioning module reads the question and answer content sequentially, segments it according to word order, identifies the structural type, determines whether it is separated from the subject based on the position, identifies missing reference through dependency, and determines the interruption of intent by comparing the end of the sentence. When all three conditions are met, it is marked as semantically ambiguous and a tag sequence is generated. The referential freezing module extracts semantically overlapping terms based on the marked sequence, compares semantic chains to determine consistency of direction, evaluates word order and context distance, and freezes and generates a list of frozen segments if there are chain breaks and sequence differences. Based on the frozen fragment list, the state injection module rearranges the word order of the unfrozen fragments, locates the intent position, reconstructs the subject-predicate structure, identifies verb-object combinations, modification chains, and isolated verbs, and generates a sequence of connecting fragments. The path filtering and reduction module traverses the connection segment sequence based on the connection segment sequence, and judges the semantic domain, topic orientation continuity and conceptual consistency. If there are logical jumps and semantic misalignments, a semantic continuous chain group is formed. The content splicing module reads the chain fragments based on the semantic continuous chain group, restores the structure according to the main narrative order, matches the main theme words, fills in the missing content, adjusts the order of the last words, and generates personalized health Q&A results.

[0005] As a further aspect of the present invention, the marker sequence includes subject detachment markers, missing pointer markers, and intent interruption markers; the frozen fragment list includes semantic chain interruption markers, word order consistency anomaly markers, and context distance anomaly markers; the connecting fragment sequence includes intent localization results, subject-verb structure reconstruction results, and semantic integrity recognition results; the semantic continuous chain group includes domain-consistent paragraphs, topic-pointing continuous paragraphs, and concept-connecting matching paragraphs; and the personalized health Q&A results include topic word matching content, duplicate item screening content, semantic ending completion content, and end word order adjustment content.

[0006] As a further aspect of the present invention, the input positioning module includes: The word order semantic segmentation submodule acquires the language content in the medical question and answer input, reads it from front to back in word order, and performs segmentation judgment based on the semantic coupling degree between words and the word class boundary coefficient. It performs segmentation on word groups whose boundary coefficient is less than the syntactic boundary benchmark value and generates a forward semantic segmentation sequence. The structure type identification submodule extracts the structural markers of terms in the sentence based on the forward semantic segmentation sequence, calculates the subject-predicate structure mapping weight, compares it with the subject occupancy threshold, identifies terms that are separated from the subject, and generates a semantic subject-separated marker matrix. The sentence intent interruption detection submodule calls the semantic detachment subject tag matrix, compares the structural continuity index between adjacent sentences, and combines the matching accuracy of missing directional terms to identify terms that simultaneously meet the three conditions and generate a tag sequence.

[0007] As a further aspect of the present invention, the reference freezing module includes: The context review submodule obtains the marked terms in the marked sequence, reads the original expression in the question-and-answer context, extracts semantically overlapping terms and intra-sentence positional relationships, calculates the consistency between expression mode and usage position, and generates a set of semantically overlapping expressions. The semantic chain detection submodule constructs a semantic chain of terms based on the set of semantically overlapping expressions, calculates the word order change value in adjacent rounds, filters links whose change value exceeds the word order stability threshold, and obtains a semantic chain break index table. The stability determination submodule calls the semantic chain break index table, compares the content distance of terms in the current context, identifies terms whose pointing consistency is lower than the stability benchmark, and generates a list of frozen fragments.

[0008] As a further aspect of the present invention, the state injection module includes: The word order reconstruction submodule obtains the unfrozen content in the frozen fragment list, locates the position of the intended expression, identifies the subject fragment and the corresponding predicate part, rearranges the order of the subject and predicate fragments and completes the sentence structure adjustment, and generates the word order reconstruction result; The structure determination submodule calls the word order reconstruction result, detects semantic combination relationships word by word, determines whether it constitutes a verb-object collocation, a modification structure, and whether the verb exists independently, marks sentences that meet all the determination conditions, and obtains semantic structure consistency identifiers; The state access submodule filters structurally complete statements based on the semantic structure consistency identifier, accesses the language content into the state process in its original order and registers it in the current expression chain, and generates a sequence of connecting fragments.

[0009] As a further aspect of the present invention, the path filtering and reduction module includes: The paragraph traversal submodule obtains the language paragraphs in the sequence of connecting segments, reads each paragraph in order, extracts the semantic range between the current paragraph and the previous paragraph, judges whether the semantic attribution between the two paragraphs is consistent by classifying them by concept tags, and generates a semantic attribution comparison table. The domain consistency determination submodule calls the semantic attribution lookup table, extracts the topic pointing relationship of each paragraph phrase, analyzes the continuity value of the target pointing in the phrase chain, determines whether there is a break in the pointing chain and topic jump, and obtains the topic pointing consistency identifier. The semantic jump filtering submodule compares the first and last noun concept tags of each language segment with the theme pointing to the consistent identifier. If the concepts before and after do not match and there is no mapping relationship between concepts in the same domain, the content is removed from the language chain and a semantic continuous chain group is generated.

[0010] As a further aspect of the present invention, the content splicing module includes: The topic restoration submodule obtains the semantic continuous chain group, arranges it according to the topic order in the original question description, performs a comparison operation on the main words in each language segment, checks for duplicate and overlapping terms, and obtains a main word order mapping table. The semantic repair submodule calls the topic word order mapping table to identify missing semantic ending paragraphs and abnormal expressions of the last word position, inserts the corresponding conclusions and rearranges the order of the last words to generate a structurally adjusted paragraph sequence. The question-and-answer combination submodule adjusts the sequence of language segments according to the structure, combines language fragments in order of thematic clues, connects each segment to form a continuous expression chain, and establishes personalized health question-and-answer results.

[0011] As a further aspect of the present invention, the method of identifying missing references through dependency refers to judging whether there are unclear referents and missing subjects in a sentence based on the dependency relationship between words. The judgment of semantic domain, topical continuity and conceptual consistency refers to judging whether the semantics are coherent and consistent by analyzing the semantic category to which the words belong, topic coherence and conceptual logical relationship.

[0012] As a further aspect of the present invention, the subject occupancy threshold refers to the minimum weight standard for measuring a word as the subject in a sentence, and determining whether a word has deviated from the subject structure.

[0013] As a further aspect of the present invention, the word order stability threshold refers to the upper limit of the numerical value that measures the tolerance for word order changes in the semantic chain, and determines whether the semantic chain maintains structural stability.

[0014] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: In this invention, ambiguous areas are marked by identifying semantic features such as subject separation, missing reference, and interrupted intent. Unstable content is frozen and filtered by combining semantic chain consistency, word order change, and contextual distance judgment. The structural integrity and semantic closure of the sentence are enhanced by subject-predicate structure reconstruction, verb-object combination recognition, and modification chain judgment. Furthermore, semantically disjointed content is screened out by topic attribution matching, connection of concepts, and logical consistency. Content reorganization and connection optimization are achieved based on topic word location, end word order adjustment, and semantic ending completion. This improves the semantic continuity stability, content coherence, and response relevance in scenarios with ambiguous input, structural disorder, or multiple jumps. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a system schematic diagram of the present invention; Figure 2 This is a schematic diagram of the system framework of the present invention; Figure 3 This is a flowchart of the input positioning module in this invention; Figure 4 This is a flowchart of the referencing freeze module in this invention; Figure 5 This is a flowchart of the state injection module in this invention; Figure 6 This is a flowchart of the path filtering and reduction module in this invention; Figure 7 This is a flowchart of the content splicing module in this invention. Detailed Implementation

[0017] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0018] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0019] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0020] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0021] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0022] This invention provides a personalized health question-answering system based on reinforcement learning, such as... Figure 1-2 The diagram shown illustrates a personalized health question-answering system based on reinforcement learning. This system includes: The input positioning module continuously reads the language content appearing in the medical question-and-answer input, segments each sentence semantically from front to back according to word order, identifies the structural type of each word in the sentence, determines whether it has semantic detachment from the subject through positional placement rules, identifies whether there is missing direction through phrase dependency analysis, and identifies whether the intention between sentences is interrupted through sentence end structure comparison. For words that meet all three conditions, semantic fuzzy labeling is performed to generate a label sequence. The freezing module reviews each marked part in the marked sequence, reads the usage of words in the questions and answers, calls the expressions that semantically overlap with the words in the original context, checks whether they have pointed to the same medical content in the differentiated context through semantic chain comparison, calculates the consistency of the expression order in adjacent rounds through word order comparison, and judges the stability of the pointing through contextual content distance. If the semantic chain is broken or the sequence changes too much, it is marked as frozen and a list of frozen fragments is generated. The state injection module rearranges the expression order of the unfrozen content in the frozen fragment list. First, it locates the position of the intended expression in the sentence, reconstructs the position of the subject fragment and the following predicate part, and performs semantic integrity recognition on the word order reconstruction result word by word to identify whether a verb-object combination is formed, whether a reasonable modification chain is formed, and whether there is an isolated verb phenomenon in the sentence. If none of the above three items conflict, the current language content is connected to the state process to generate a sequence of connecting fragments. The path filtering module sequentially traverses the language segments formed in the sequence of connecting fragments, compares the semantic scope of the segment with the previous expression, judges the continuity of the topic orientation between the phrases used, and filters out cases where the concepts of the first and last nouns are inconsistent. If there are logical jumps or semantic misalignments, the segment is removed from the language chain and a semantic continuous chain group is generated. The content splicing module reads each language segment in the semantic continuous chain group, performs a structure restoration operation according to the theme order expressed in the original question description, matches the main words of each language segment, checks for overlapping or repeated words, introduces corresponding conclusions for paragraphs with missing semantic endings, rearranges the word order of expressions with incorrect end word positioning, and logically connects and combines all the languages ​​to generate personalized health Q&A results.

[0023] The marker sequence includes subject detachment markers, missing pointer markers, and intent interruption markers; the frozen fragment list includes semantic chain interruption markers, word order consistency anomaly markers, and context distance anomaly markers; the connecting fragment sequence includes intent localization results, subject-verb structure reconstruction results, and semantic integrity recognition results; the semantic continuous chain group includes domain-consistent paragraphs, topic-pointing continuous paragraphs, and concept-connecting matching paragraphs; and the personalized health Q&A results include topic word matching content, duplicate item screening content, semantic ending completion content, and end word order adjustment content.

[0024] Specifically, such as Figure 2 , 3 As shown, the input positioning module includes: The word order semantic segmentation submodule acquires the language content in the medical question and answer input, reads it from front to back in word order, and performs segmentation judgment based on the semantic coupling degree between words and the word class boundary coefficient. It performs segmentation on word groups whose boundary coefficient is less than the syntactic boundary benchmark value and generates a forward semantic segmentation sequence. When the word order semantic segmentation submodule acquires the language content from the medical question-and-answer input, the system first reads the words in the input text sequentially. For expressions like "the patient has had a persistent low-grade fever with a cough for the past week," the system calls the language model to perform part-of-speech tagging and labeling on the words. The resulting word order is "patient / noun, past week / adverb of time, persistent / adverb of verb, low-grade fever / noun, accompanied / verb, cough / noun." The reading order remains from front to back. Semantic coupling is calculated between every two words, and the word embedding corpus model is used to... Based on the similarity score of the terms, a basic word vector database was established from a general medical corpus. The semantic coupling score between "low-grade fever" and "cough" was 0.82, between "persistent" and "low-grade fever" it was 0.73, and between "last week" and "persistent" it was 0.51. However, the coupling score between "last week" and "patient" was only 0.28. According to the internally defined semantic coupling degree segmentation reference values, a high coupling range of 0.7 or higher, a medium range of 0.5–0.7, and a low coupling range of 0.3–0.5 were defined. Extremely low coupling is defined as below 0.3. Term pairs falling into this range will be prioritized as potential segmentation points. When making segmentation decisions, the system also needs to calculate the word class boundary coefficient. This coefficient is set with reference to the switching probability of common word class combinations in historical medical texts. For example, the boundary probability of a noun followed by an adverb is 0.25, and the probability of an adverb followed by a noun is 0.6. The boundary coefficient is set based on the product of the word class switching probability and the frequency of pauses in language rhythm. For example, the boundary coefficient between "in the past week" and "continuous" is 0.25 × 0.4 = 0. 1. The boundary coefficient between "low-grade fever" and "accompanied" is 0.6 × 0.8 = 0.48. The system sets the boundary segmentation benchmark value to 0.2. This value is based on the statistical analysis of segmentation positions in tens of thousands of medical corpora. The top 20% of all segmentation points with a boundary coefficient lower than 0.2 are selected as the benchmark value. Therefore, the boundary between "in the past week" and "persistent" is segmented, while the boundary between "low-grade fever" and "accompanied" is not segmented. The final forward semantic segmentation sequence is "patient", "in the past week", "persistent", "low-grade fever with cough".

[0025] The structure type identification submodule extracts the structural markers of terms in the sentence based on the forward semantic segmentation sequence, calculates the subject-predicate structure mapping weight, compares it with the subject occupancy threshold, identifies terms that are separated from the subject, and generates a semantic subject-separated marker matrix. The structure type identification submodule processes the forward semantic segmentation sequence. During the structure identification process, the grammatical role of each word in the original sentence is first determined. For example, "patient" is usually a subject-type word, "low-grade fever" and "cough" are often the object components of verb structures, and "persistent" plays a modifying role. The system uses a dependency graph to extract the subject-verb and verb-object structure paths between words. In the short sentence "patient has persistent low-grade fever", the path "patient → persistent → low-grade fever" is identified, where "patient" is at the beginning of the sentence, which is the basic point for locating the subject. The system then evaluates the structural mapping weight between each word and the subject. This weight value is determined by both the path length and semantic similarity. In the "patient → low-grade fever" path, the path length is 2, the semantic association score is 0.7, and the mapping weight is 0.35 according to the set rules. Regarding the word "chills," in the sentence "I experienced chills this morning," the subject is missing. By tracing back to the previous sentence "The patient has had a persistent low-grade fever and cough for nearly a week," it was found that the path between "chills" and the subject "patient" is 4 steps, with a semantic similarity score of 0.4. Therefore, its structural mapping weight is 0.1. The system sets the subject occupancy threshold to 0.25. This threshold is determined by statistically analyzing subject-verb matching samples from 10,000 medical Q&A entries, taking the top 30% of those with subject-verb mapping weights below 0.25 as the judgment reference. Therefore, "chills" is judged as a word semantically separated from the subject, and the corresponding position is marked as 1 in the tag vector. For example, the vector [0, 0, 0, 1, 0, 1] indicates that the 4th and 6th words are separated from the subject. Finally, a semantically separated-subject tag matrix is ​​generated for subsequent inter-sentence association judgment operations.

[0026] The inter-sentence intent interruption detection submodule calls the semantic detachment subject tag matrix, compares the structural continuity index between adjacent sentences, and combines the matching accuracy of missing directional terms to identify terms that simultaneously meet the three conditions and generate a tag sequence. When processing adjacent sentence pairs, the inter-sentence intent interruption detection submodule first reads the semantic decoupling subject marker matrix corresponding to the preceding and following sentences. For example, the sentences "The patient has had a persistent low-grade fever and cough for nearly a week." and "This morning, chills appeared." generate marker vectors [0, 0, 0, 0, 0, 0] and [1, 0, 1] respectively. The system determines whether the beginning of the current sentence is a structural breakpoint. Then, it evaluates the structural continuity between the end of the preceding sentence and the beginning of the current sentence. This index is calculated by weighting semantic coupling score and word class transition score. For example, the semantic coupling score between the word "cough" at the end of the preceding sentence and "this morning" at the beginning of the current sentence is 0.25, and the word class transition score is 0.3. The weight coefficients are set to 0.6 and 0.4 respectively. The sum of the two values ​​yields a structural continuity index of 0.27. The system sets the inter-sentence intent interruption detection submodule as follows: The structural continuity threshold is 0.3. This threshold is determined by statistical analysis of 500 manually annotated medical question-and-answer pairs. The top 25% of the structural continuity scores for natural connections between sentences correspond to a value of 0.3. If the structural continuity index is lower than 0.3, and the first word of the current sentence, "this morning," does not form a structural path with the subject of the preceding text, and the subject-verb structure mapping weight between "chills" and the subject "patient" in the current sentence is only 0.1, which is lower than the subject occupancy threshold of 0.25, and it does not match common directional terms such as "cause," "lead to," and "appear," then all three conditions are met. The system marks the position of "chills" as an inter-sentence breakpoint, and finally generates a marker sequence of [0, 1], where each bit corresponds to the break judgment result of the adjacent sentence pair.

[0027] Specifically, such as Figure 2 , 4 As shown, the freezing module includes: The context review submodule obtains the marked terms in the marked sequence, reads the original expression in the question-and-answer context, extracts semantically overlapping terms and intra-sentence positional relationships, calculates the consistency between expression mode and usage position, and generates a set of semantically overlapping expressions. After the context review submodule obtains the marked terms in the marked sequence, the system first reads the original expression fragments related to the term in the original question-and-answer text. For example, for the marked "chills," it reviews the original expression "I felt chills this morning" and extracts the complete sentence structure corresponding to its context. Based on this, it calls the semantic overlap analysis process to compare the keywords in the original expression with the historical sentences in the question-and-answer context, extracting semantically overlapping terms one by one, such as "feeling cold," "feeling chills," and "shivering." By statistically analyzing the co-occurrence frequency of such terms in a large-scale medical question-and-answer corpus, terms with a co-occurrence frequency of no less than 3 times in three rounds of question-and-answer are selected as overlapping objects, and the actual position of these terms in the sentence is recorded. For example, "feeling cold" is the 3rd position in "felt cold last night," and "chills" is the 4th position in the current question "I felt chills this morning." The system calculates the position offset value within the sentence for each pair of overlapping terms, and further... The system compares whether their expressions have the same grammatical structure. In "I had a chill this morning" and "I felt cold last night," both are "time + verb + symptom noun" structures, which are consistent structures. The system records the grammatical structure consistency mark accordingly. Then, the system evaluates the total number of overlapping words and whether their positional offsets are concentrated. The consistency score is calculated based on two factors: first, whether the proportion of semantically similar words exceeds 60% of all candidate words; and second, whether the positional offset of these words in their respective sentences does not exceed 40% of the sentence length. In this example, there are 5 candidate words and 3 overlapping words, accounting for 60%. The maximum positional offset of the two groups of overlapping words is 2, and the sentence length is 6 words, accounting for 33.3%. This meets the two conditions set by the system, and the consistency score meets the standard. Finally, a set of semantically overlapping expressions consisting of "chills," "cold," and "chills" is generated, and the position and structural mark of each word in the sentence are recorded.

[0028] The semantic chain detection submodule constructs a semantic chain of terms based on the set of semantically overlapping expressions, calculates the word order change value in adjacent rounds, filters links whose change value exceeds the word order stability threshold, and obtains a semantic chain break index table. The specific formula for calculating word order change value is as follows: ; in, Representing the The word order change value between rounds, The semantic chain is numbered as The term in the first The word order of the wheel, Representative number is The term in the first The word order of the wheel, Representative number is The term in the first Semantic weight values ​​in the wheel, Representative number is The term in the first Semantic weight values ​​in the wheel, Represents the index of all terms involved in the calculation within the semantic chain. Representing the The average value of word order changes for all terms in the round.

[0029] Word order position , After parsing the semantic chain, the system assigns a natural number to the terms in each semantic unit. For example, in the following text structure: ["artificial intelligence", "model", "training", "parameters", "inference"], if "model" is the 2nd term in the previous round and the 4th term in the next round, then the corresponding term is... , ; Semantic weight The weighted average is obtained from the following formula: ; in, For word frequency, Inverse document frequency, Cosine similarity is used to measure the cosine similarity between a word and the target topic within the context. For example, the parameter for the word "model" in the k-th round is... , , If the maximum value is 8.5, then: ; Reference length Based on the maximum chain length of the semantic chain in this system, the value is 16.

[0030] Normalized average displacement The normalized position changes were calculated for each of the five term samples and then averaged, as follows: Table 6 Semantic Chain Sample Parameter Table: As shown in Table 6, a total of 5 groups of samples were collected.

[0031] Substitute into the formula to perform the complete calculation: Molecular calculation terms: ; ; calculate: ; Denominator calculation: ; ; ; ; Final calculation results: ; Results analysis and threshold comparison: The stability threshold for semantic chain break detection has been set to 0.22 (based on the mean plus standard deviation of the first 30 rounds), therefore the currently calculated... Below the threshold.

[0032] This result indicates that the semantic chain is in the first... If no significant structural changes occur in a round and the break index marker is not triggered, it can be determined to be a stable round state; this value will be input into the semantic chain monitoring module of the next round for fluctuation trend aggregation calculation.

[0033] The advantage of the formula lies in the introduction of a position normalization factor. To achieve comparability of fluctuations between different semantic chain lengths, normalized semantic position differences are coupled with dual-wheel semantic weights to effectively suppress interference from low-weight terms. Furthermore, overall variation and mean smoothing terms are introduced into the denominator. This has resulted in a technical solution that is sensitive to mutations and improves the detection accuracy of highly variable structures.

[0034] The stability determination submodule calls the semantic chain break index table, compares the content distance of terms in the current context, identifies terms whose pointing consistency is lower than the stability benchmark, and generates a list of frozen fragments. After obtaining the semantic chain break index table, the stability determination submodule reads the marked link break terms and their context information. For example, "chills" is identified as a break point in the third-round sentence "chills occurred this morning." The system first extracts the text distance between this term and the corresponding semantic term "chills" in the previous round. This distance is defined as the difference in the relative positions of the two terms in their respective sentences multiplied by a conventional factor of the time span between sentences. The position difference is 2, the time span is one day, and the corresponding factor is 1.5, so the content distance is 3.0. Subsequently, the system evaluates whether there is a valid semantic path connection between "chills" and the subject term "patient" in the current context. By searching the verb components and their subject-object pairing structures in the current sentence, it is found that the sentence "chills occurred this morning" lacks bridging words connecting the subject and symptoms, such as "causes," "leads to," and "feels," thus lacking logical connection. The system identifies missing semantic connections and records this missing behavior as "no-direction path". Next, it evaluates the sum of the semantic similarity score and structural distance score between "chills" and "patient". If this sum is lower than a set stable baseline value, it is considered inconsistent in direction. This stable baseline value is set at 0.25, derived from the lower limit of the minimum confidence interval corresponding to the mean semantic and structural scores of subjects and corresponding symptoms forming a direct logical relationship in a manually annotated question-and-answer set. In this example, the semantic score between "chills" and "patient" is 0.1, the structural path length is 4, and the total score after conversion is 0.12, significantly lower than 0.25. Therefore, the segment containing "chills this morning" is marked as unstable content and added to the frozen segment list. This list ultimately records the sentence containing the term "chills", the disconnection of "patient" as the subject, and the round index information of this segment in the original question-and-answer set.

[0035] Specifically, such as Figure 2 , 5 As shown, the state injection module includes: The word order reconstruction submodule obtains the unfrozen content from the frozen fragment list, locates the position of the intended expression, identifies the subject fragment and its corresponding predicate, rearranges the order of the subject and predicate fragments and completes the sentence structure adjustment, and generates the word order reconstruction result; After obtaining the unfrozen content from the frozen fragment list, the word order reconstruction submodule first performs a core structure identification operation on each piece of language content. During the identification process, the system locates the intention expression position in the sentence, that is, identifies the words in the sentence that carry the main action or state. For example, in the sentence "Today, a chill suddenly appeared," "appeared" is identified as the main predicate verb and is thus identified as the intention expression position. Subsequently, the system further identifies the subject fragment and its corresponding predicate part, extracting the subject candidate word "chill" and its connected predicate verb "appeared" through dependency relation analysis. During the identification process, non-core parts such as the time adverbial "today" and the manner adverbial "suddenly" are excluded, and these subordinate structures are marked to retain their reconstructed relative positions. Then, the system adjusts the order of the subject-predicate structure according to preset word order rules. These rules are defined based on the general Chinese word order structure "subject + adverbial + predicate + object." If any is found... If the original word order is inconsistent with the structure, it will be rearranged. For example, if the original sentence structure is "time + subject + adverbial + predicate", it will be adjusted to either "time + adverbial + predicate + subject" or "time + adverbial + subject + predicate" in a logical order that conforms to the context. After determining the subject-predicate rearrangement, the system completes the sentence structure adjustment operation, performs syntactic rule matching on the rearranged sentence, and judges whether there are omissions of modal particles, voice errors, abnormal order of function words, etc. If there is no clear conjunction between the subject and predicate in the original sentence, the original compressed structure will be maintained without introducing new words. For example, "chills appeared this morning" will be reconstructed as "chills appeared this morning", while "the patient was diagnosed with fever" will not be adjusted in word order because it contains a passive structure. The system will automatically exclude such sentence structures according to the preset list of structures that do not undergo word order adjustment in the structure type table, and finally form a new sentence structure that meets the structural requirements, and use it as the result of word order reconstruction.

[0036] The structure determination submodule calls the word order reconstruction results, detects semantic combination relationships word by word, determines whether it constitutes a verb-object collocation, a modification structure, and whether the verb exists independently, marks sentences that meet all the determination conditions, and obtains semantic structure consistency identifiers; After receiving the word order reconstruction result, the structure determination sub-module analyzes the structure of each sentence one by one. The system first scans each sentence character by character to identify the词性 and dependency relationship of each lexical item in the sentence, and confirms whether there is an effective semantic combination relationship between the lexical items. For example, in the sentence "Suddenly there is a chill today", "出现" is a verb and "寒战" is a noun. The system identifies it as a typical verb-object structure and judges by detecting whether there is a nominal object following the verb "出现". This judgment method is based on whether there is a structure where a verb is followed by a noun and the noun has no subject independently as the judgment condition. If it is satisfied, the verb-object structure is established. Then the system identifies the modification relationship where "突然" modifies "出现", and judges whether the modification structure is established by analyzing whether there is an attachment relationship in the词性 combination "adverb + verb". Finally, it checks whether the verb "出现" exists independently in the sentence, that is, whether there is a lack of subject or object structure complementation. If nothing is connected after "出现", it is marked as "verb exists independently". After making these three judgments on the word order reconstruction result of each sentence, the system judges whether the sentence meets the requirements according to the set structure consistency standard. This standard stipulates that all three structural conditions need to be met to be recognized as semantic structure consistency. This judgment standard is set based on the statistical results of a large-scale structurally annotated corpus. More than 90% of the standard medical sentence patterns meet the above three conditions in the question-and-answer structure. Therefore, the system takes the full establishment of the three structural judgments as the only standard for structural consistency. If only any two of them are satisfied, it is marked as partially consistent in structure. If only one or none of them is satisfied, it is marked as inconsistent in structure. After the structure determination is completed, the system marks a "√" beside the sentence with consistent structure and summarizes its index information to prepare for the subsequent screening operation, and finally obtains the semantic structure consistency identifier.

[0037] The status access sub-module filters out the sentences with complete structures according to the semantic structure consistency identifier, accesses the language content into the status process in the original order and registers it in the current expression chain, generating a sequence of衔接 fragments; After reading the structure consistency identifier sequence, the state access submodule sequentially filters the statements marked as structurally consistent. The system accesses the language state process sequentially according to the order of the statements in the original text. Before accessing, the system first verifies whether the statement has continuity with the existing content in the current expression chain in terms of theme or time span. The judgment criteria are: 1. Whether the current statement uses the subject of the previous statement. If the subject of the previous statement is "patient" and the current statement has no subject, then the subject is inherited by default; 2. Whether the current statement introduces new time adverbs or obvious time breaks, such as words like "today," "this morning," and "early morning." If so, the system marks it as time progression; 3. Whether the predicate verb in the current sentence and the predicate of the previous sentence form a continuous relationship. If at least two of the three criteria are "yes," the sentence is considered a connecting statement and can be formally added to the state process. For example, the sentence "I suddenly felt chills today" and the previous sentence "I had a slight cough yesterday" both meet the system's judgment criteria in terms of subject consistency, time progression, and action intensification. Therefore, it is registered as a valid structural statement in the state process. The system then classifies and marks the state content of the statement as "symptom intensification," categorizes it as "chill symptoms," and records its position in the expression chain. For example, if it is in the 7th sentence, it is marked as "state process node 7." Finally, the system outputs a complete sequence of connecting fragments and updates the logical and semantic association information between each node in the state chain.

[0038] Specifically, such as Figure 2 , 6 As shown, the path filtering and reduction module includes: The paragraph traversal submodule obtains language paragraphs in the sequence of connected fragments, reads each paragraph in order, extracts the semantic range between the current paragraph and the previous paragraph, judges whether the semantic attribution between the two paragraphs is consistent by classifying them by concept tags, and generates a semantic attribution comparison table. After the paragraph traversal submodule obtains the language paragraphs in the sequence of connected fragments, the system traverses and reads the language content of each paragraph according to the original dialogue order, identifying the main keywords in each paragraph, including the main symptom description, diagnostic behavior, and disease course time. For example, if paragraph 1 is "The patient developed a cough and low-grade fever last night" and paragraph 2 is "Suddenly developed chills today," the system identifies the keywords "cough" and "low-grade fever" in paragraph 1 and "chills" in paragraph 2. The system then calls the concept tag database to perform semantic tag matching on the above terms, classifying "cough" as "respiratory symptoms," "low-grade fever" as "abnormal body temperature," and "chills" as "cold reaction," and establishing a semantic range set for each. Each set records the tag assignment results of the terms appearing in the paragraph. Then, the system compares the semantic range of the current paragraph with that of the previous paragraph. If the tags in the two sets belong to the same first-level semantic domain, such as both belonging to the "basic clinical symptoms" domain, the system will detect the error. If there is no overlap in the tags or only non-core tags match, such as one tag belonging to "physical signs" and the other to "diagnosis and treatment behavior", then it is recorded as "semantic attribution inconsistent". When judging whether the attribution is consistent, the system uses the tag overlap as the judgment index, that is, the number of tag intersections in the two sets divided by the number of tag unions. The system sets the semantic attribution judgment threshold to 0.5, which is derived from the statistical analysis of the proportion of tag intersections in the inter-segment coherence in multi-turn question and answer in standard medical corpus. The first 50% of the tags have an overlap of greater than or equal to 0.5, so this value is set as the judgment basis. For example, if the first and second paragraphs have a total of 1 tag and a total of 2 tags, then the overlap is 0.5, which meets the condition. Finally, the system records the attribution judgment result between each pair of paragraphs and generates a complete semantic attribution comparison table. This table contains information such as paragraph number, keyword items, tag set, tag intersection, and attribution judgment result, which is used for subsequent domain consistency processing.

[0039] The domain consistency determination submodule calls the semantic attribution lookup table, extracts the topic pointing relationship of each paragraph phrase, analyzes the continuity value of the target pointing in the phrase chain, determines whether there is a break in the pointing chain and topic jump, and obtains the topic pointing consistency identifier; After the domain consistency determination submodule calls all paragraph pairs judged as having "semantic consistency" in the semantic attribution lookup table, the system extracts the semantic pointing path of the core paragraph phrases in each paragraph. The so-called paragraph phrases are the phrases that best represent the semantic theme of the paragraph, such as "cough" and "chills". The system analyzes whether there is a unified pointing topic among these phrases, such as "infection symptoms" and "cold pathological state". If the paragraph phrases can be traced to the same topic branch in the semantic chain, the system records it as "continuous pointing". If the tracing results are assigned to different topic domains, such as one to "signs and symptoms" and the other to "treatment and intervention", the system records it as "pointing break". Then the system calculates the topic distance value in the chain of the two phrases, that is, records the step length of the shortest semantic path between the two phrases through the concept graph structure. The shorter the path, the higher the continuity of the theme. The system sets the continuity judgment threshold to a maximum path length of no more than 3 steps. The threshold setting is based on the analysis of common semantic theme shift structures in clinical literature. Paths exceeding 3 steps generally no longer constitute a valid pointing relationship. For example, the path between "cough → respiratory infection → clinical symptoms" and "chills → cold reaction → clinical symptoms" has 3 steps and meets the continuity condition. However, the path between "cough → respiratory infection → diagnosis → intervention → treatment" and "chills → cold reaction" exceeds 5 steps and is therefore broken. The system calculates the theme path value for all paragraph chains and compares it with the threshold. If the path length does not exceed 3, the system marks it as "theme consistent"; if it exceeds 3, it marks it as "theme inconsistent". Finally, the theme pointing to the consistent markers for all paragraphs is used for semantic filtering.

[0040] The semantic jump filtering submodule compares the first and last noun concept tags of each language segment with the consistent identifier of the topic. If the concepts before and after do not match and there is no mapping relationship between concepts in the same domain, the content is removed from the language chain and a semantic continuous chain group is generated. The semantic jump filtering submodule performs semantic filtering operations based on the generated topic-alignment identifiers. The system extracts the first and last words of each paragraph. The filtering logic is as follows: if the current paragraph is marked as "topic inconsistent," the concept tags corresponding to the first and last words are further extracted, and the same-domain concept mapping table is called to check whether there is an acceptable cross-concept correspondence. For example, in the paragraph "Frequent chills today, need to continue observation," the first word "chills" is tagged as "chill reaction," and the last word "observation" is tagged as "medical behavior." The two do not belong to the same semantic domain, and there is no acceptable mapping pair. The system determines that this paragraph is a semantic jump paragraph and removes it. The mapping judgment is based on a structured concept correspondence table, which... The system includes 900 pairs of concept mapping items, sourced from a medical semantic database and manually annotated results. The system uses this table as a benchmark for tag matching; if no match is found, it is considered a semantic jump. The system's judgment logic is: if a paragraph has an inconsistent theme, different first and last word tags, and no mapping path, then that paragraph must be removed. Removal does not delete the original sentence content; instead, it marks the paragraph as a "jump" in the semantic chain, and subsequent processing modules will no longer call its content. This design allows the system to retain and filter record information without destroying the original records. Finally, the system constructs a new chain from all the retained paragraphs, arranges them in the original sentence order to generate a semantic continuous chain group, and records their mapping positions in the original dialogue.

[0041] Specifically, such as Figure 2 , 7 As shown, the content splicing module includes: The topic restoration submodule obtains semantic continuous chain groups, arranges them according to the topic order in the original question description, performs a comparison operation on the main words in each paragraph, checks for duplicate and overlapping terms, and obtains a main word order mapping table. After obtaining the semantic continuum, the topic reconstruction submodule first calls the topic parsing results of the original question to extract the main keywords and their order. For example, if the question is "The patient has had a persistent fever for the past three days and is now experiencing chills, is it possible that he / she has contracted influenza?", the system identifies the main topic order as "fever," "chills," and "influenza infection." Next, the system performs topic word extraction on each paragraph in the semantic continuum. For the paragraph "body temperature fluctuated to 38.5℃ and did not subside overnight," the topic word "body temperature fluctuation" is extracted and categorized as "fever." For the paragraph "sudden onset of chills accompanied by muscle aches in the early morning," the topic word "chills" is extracted. For the paragraph "preliminary exclusion of bacterial infection," the topic word "infection" is extracted and its semantic attribute is determined to be "excluded." Subsequently, the system establishes a mapping relationship between the topic words and the original question topic. If the topic words in a paragraph completely correspond to the original topic words, the system marks them as such. This is recorded as "precise mapping". If the topic word is a synonym or similar term of the original topic, such as "elevated body temperature" and "fever", it is marked as "synonymous attribution". The system performs a cross-check on the topic words of all paragraphs to identify whether the same topic keywords appear repeatedly in different paragraphs. For example, if "chills" appears repeatedly in two paragraphs as "frequent chills at night" and "worsening chills", the system retains the paragraph where it first appears and records the duplicate as "semantic redundancy". This judgment is based on the term similarity being greater than 0.85, which is considered a duplicate. This threshold comes from the clustering results of synonyms in the medical semantic lexicon. The lower limit of the top 20% of the average similarity is selected as the screening benchmark. Finally, the system outputs a topic word order mapping table, which includes fields such as: paragraph number, extracted topic word, corresponding topic position, whether it is a duplicate, whether it is a synonym, and attribution label, which are used for reference in the next module.

[0042] The semantic repair submodule calls the topic word order mapping table to identify missing semantic ending paragraphs and abnormal expressions of the last word position, inserts the corresponding concluding remarks and rearranges the order of the last words to generate a sequence of structurally adjusted paragraphs. After the semantic repair submodule calls the topic word order mapping table, it reads the structural information segment by segment. For each segment, it judges whether there is any expression loss or semantic incompleteness in the ending word. The system first checks the part of speech and semantic structure of the sentence-ending word. For example, in the sentence "sudden chills in the early morning," the ending word "chills" is a noun. The system judges that the sentence does not reflect the state evolution, treatment measures, or future inferences. According to the semantic segment integrity judgment standard, it is recorded as "semantic ending missing." Subsequently, based on the statistical results of common ending collocations of the topic word "chills" in medical corpora, the system selects the most suitable ending sentence, such as "it is recommended to record changes in body temperature," and performs semantic ending completion. The judgment standard for the completion operation is: if the ending of the topic segment does not contain any two of the following: time adverbs, action verbs, or medical advice expressions, then the ending is completed. The system initiates completion when two missing elements are found. This threshold is set by analyzing the standard expression structures in 1000 doctor responses. Over 85% of the complete responses contain at least two of the three elements mentioned above at the end. After completion, the sentence is checked for word order. If the inserted content causes grammatical disorder, for example, the sentence "It is recommended to record changes in body temperature" is rearranged to "It is recommended to record changes in body temperature." The word order rearrangement rule is: first the verb, then the object, and then the modifier. The completed sentence "Sudden chills in the early morning, it is recommended to record changes in body temperature" is recorded by the system as a successfully structured paragraph. Finally, the system outputs a sequence of structured paragraphs, with marked fields including the original sentence content, inserted content, rearrangement status, and main idea position, which are then used by the question-and-answer combination module.

[0043] The question-and-answer combination submodule adjusts the sequence of language segments according to the structure, combines language fragments in order of thematic clues, connects each segment to form a continuous expression chain, and establishes personalized health question-and-answer results. After receiving the structured paragraph sequence, the question-and-answer combination submodule rearranges all paragraphs according to the topic order information marked in the topic word order mapping table. This ensures that the combined expression conforms to the semantic development logic of the original question. For example, if the original question's topic order is "fever → chills → influenza infection," the system sets the paragraph "body temperature fluctuates to 38.5℃, not subsided overnight" as paragraph 1, "sudden chills in the early morning, it is recommended to record body temperature changes" as paragraph 2, and "preliminary exclusion of bacterial infection" as paragraph 3. After arrangement, the system begins to combine the content of each paragraph into a continuous expression chain. To avoid abrupt transitions between paragraphs, the system calls the list of connecting logic words and inserts connecting statements between paragraphs, such as "subsequently," "in addition," "considering," etc. The selection logic for inserting words is: if the topic words between paragraphs are of the same type of symptom, then insert "subsequently," "in addition," "in addition," etc. Subsequently, if the main idea of ​​the following paragraph is an extension or inference of the phenomenon in the preceding paragraph, then "considering" is inserted. For example, "subsequently" is inserted when connecting the paragraphs "body temperature fluctuation" and "chills", and "currently preliminarily judged" is inserted when connecting the paragraphs "chills" and "infection ruled out". The system sets the threshold for selecting connecting words to be that the semantic similarity of the main words between paragraphs is not less than 0.6. The judgment is based on the median value of the average semantic similarity distribution of the main word pairs before and after the connecting words in 1000 sets of semantically continuous medical Q&A. Finally, a complete Q&A result is generated: "Body temperature fluctuated to 38.5℃, did not subside at night, and then suddenly chills occurred in the early morning. It is recommended to record the changes in body temperature. Currently, it is preliminarily judged that bacterial infection has been ruled out." The system assembles this Q&A chain into a continuous semantic chain, marks it as a personalized health Q&A result, and attaches paragraph index, main word order and structural completeness tags.

[0044] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A personalized health question-answering system based on reinforcement learning, characterized in that, The system includes: The input positioning module reads the question and answer content sequentially, segments it according to word order, identifies the structural type, determines whether it is separated from the subject based on the position, identifies missing reference through dependency, and determines the interruption of intent by comparing the end of the sentence. When all three conditions are met, it is marked as semantically ambiguous and a tag sequence is generated. The referential freezing module extracts semantically overlapping terms based on the marked sequence, compares semantic chains to determine consistency of reference, evaluates word order and context distance, and freezes and generates a list of frozen segments if there are semantic chain breaks or sequence differences. Based on the frozen fragment list, the state injection module rearranges the word order of the unfrozen fragments, locates the intent position, reconstructs the subject-predicate structure, identifies verb-object combinations, modification chains, and isolated verbs, and generates a sequence of connecting fragments. The path filtering and reduction module traverses the connection segment sequence based on the connection segment sequence, and judges the semantic domain, topic orientation continuity and conceptual consistency. If there are logical jumps and semantic misalignments, a semantic continuous chain group is formed. The content splicing module reads the chain fragments based on the semantic continuous chain group, restores the structure according to the main narrative order, matches the main theme words, fills in the missing content, adjusts the order of the last words, and generates personalized health Q&A results.

2. The personalized health question-answering system based on reinforcement learning according to claim 1, characterized in that: The marker sequence includes subject detachment markers, missing pointer markers, and intent interruption markers; the frozen fragment list includes semantic chain interruption markers, word order consistency anomaly markers, and context distance anomaly markers; the connecting fragment sequence includes intent localization results, subject-verb structure reconstruction results, and semantic integrity recognition results; the semantic continuous chain group includes domain-consistent paragraphs, topic-pointing continuous paragraphs, and concept-connecting matching paragraphs; and the personalized health Q&A results include topic word matching content, duplicate item screening content, semantic ending completion content, and end word order adjustment content.

3. The personalized health question-answering system based on reinforcement learning according to claim 1, characterized in that: The input positioning module includes: The word order semantic segmentation submodule acquires the language content in the medical question and answer input, reads it from front to back in word order, and performs segmentation judgment based on the semantic coupling degree between words and the word class boundary coefficient. It performs segmentation on word groups whose boundary coefficient is less than the syntactic boundary benchmark value and generates a forward semantic segmentation sequence. The structure type identification submodule extracts the structural markers of terms in the sentence based on the forward semantic segmentation sequence, calculates the subject-predicate structure mapping weight, compares it with the subject occupancy threshold, identifies terms that are separated from the subject, and generates a semantic subject-separated marker matrix. The sentence intent interruption detection submodule calls the semantic detachment subject tag matrix, compares the structural continuity index between adjacent sentences, and combines the matching accuracy of missing directional terms to identify terms that simultaneously meet the three conditions and generate a tag sequence.

4. The personalized health question-answering system based on reinforcement learning according to claim 1, characterized in that: The reference freezing module includes: The context review submodule obtains the marked terms in the marked sequence, reads the original expression in the question-and-answer context, extracts semantically overlapping terms and intra-sentence positional relationships, calculates the consistency between expression mode and usage position, and generates a set of semantically overlapping expressions. The semantic chain detection submodule constructs a semantic chain of terms based on the set of semantically overlapping expressions, calculates the word order change value in adjacent rounds, filters links whose change value exceeds the word order stability threshold, and obtains a semantic chain break index table. The stability determination submodule calls the semantic chain break index table, compares the content distance of terms in the current context, identifies terms whose pointing consistency is lower than the stability benchmark, and generates a list of frozen fragments.

5. The personalized health question-answering system based on reinforcement learning according to claim 1, characterized in that: The state injection module includes: The word order reconstruction submodule obtains the unfrozen content in the frozen fragment list, locates the position of the intended expression, identifies the subject fragment and the corresponding predicate part, rearranges the order of the subject and predicate fragments and completes the sentence structure adjustment, and generates the word order reconstruction result; The structure determination submodule calls the word order reconstruction result, detects semantic combination relationships word by word, determines whether it constitutes a verb-object collocation, a modification structure, and whether the verb exists independently, marks sentences that meet all the determination conditions, and obtains semantic structure consistency identifiers; The state access submodule filters structurally complete statements based on the semantic structure consistency identifier, accesses the language content into the state process in its original order and registers it in the current expression chain, and generates a sequence of connecting fragments.

6. The personalized health question-answering system based on reinforcement learning according to claim 1, characterized in that: The path filtering module includes: The paragraph traversal submodule obtains the language paragraphs in the sequence of connecting segments, reads each paragraph in order, extracts the semantic range between the current paragraph and the previous paragraph, judges whether the semantic attribution between the two paragraphs is consistent by classifying them by concept tags, and generates a semantic attribution comparison table. The domain consistency determination submodule calls the semantic attribution lookup table, extracts the topic pointing relationship of each word group, analyzes the continuity value of the target pointing in the word group chain, determines whether there is a break in the pointing chain and topic jump, and obtains the topic pointing consistency identifier; The semantic jump filtering submodule compares the first and last noun concept tags of each language segment with the theme pointing to the consistent identifier. If the concepts before and after do not match and there is no mapping relationship between concepts in the same domain, the content is removed from the language chain and a semantic continuous chain group is generated.

7. The personalized health question-answering system based on reinforcement learning according to claim 1, characterized in that: The content splicing module includes: The topic restoration submodule obtains the semantic continuous chain group, arranges it according to the topic order in the original question description, performs a comparison operation on the main words in each language segment, checks for duplicate and overlapping terms, and obtains a main word order mapping table. The semantic repair submodule calls the topic word order mapping table to identify missing semantic ending paragraphs and abnormal expressions of the last word position, inserts the corresponding conclusions and rearranges the order of the last words to generate a structurally adjusted paragraph sequence. The question-and-answer combination submodule adjusts the sequence of language segments according to the structure, combines language fragments in order of thematic clues, connects each segment to form a continuous expression chain, and establishes personalized health question-and-answer results.

8. The personalized health question-answering system based on reinforcement learning according to claim 1, characterized in that: The method of identifying missing referents through dependency refers to judging whether there are unclear referents and missing subjects in a sentence based on the dependency relationship between words. The judgment of semantic domain, topical continuity and conceptual consistency refers to judging whether the semantics are coherent and consistent by analyzing the semantic category to which the words belong, topic coherence and conceptual logical relationship.

9. The personalized health question-answering system based on reinforcement learning according to claim 3, characterized in that: The subject occupancy threshold refers to the minimum weight standard for measuring a word as the subject in a sentence, and is used to determine whether a word has deviated from the subject structure.

10. The personalized health question-answering system based on reinforcement learning according to claim 4, characterized in that: The word order stability threshold refers to the upper limit of the numerical value that measures the tolerance for word order changes in the semantic chain, and determines whether the semantic chain maintains structural stability.

Images