Expressway service area service level evaluation method and system based on knowledge reasoning

Abstract

The invention discloses a knowledge reasoning-based expressway service area service level evaluation method and system. The method comprises the steps of preprocessing acquired multi-source evaluation data to form a structured evaluation event transaction set; encoding text evidence in the transaction set into a vector, performing field adaptation by adopting unsupervised self-consistent contrast learning, and forming a synonym item set in combination with cosine similarity and threshold clustering; mining candidate rules from the synonym item set, and calculating quality scores of the candidate rules; converting the candidate rule into an executable inference rule model, and generating a rule file; converting the acquired to-be-processed fact data to generate a standardized fact, inputting the standardized fact into a rule engine, completing efficient mode matching based on a Rete network, and triggering a rule action according to a priority and a conflict resolution strategy; and aggregating results of the plurality of rules, and outputting a final evaluation grade and an interpretable link. According to the method, the rule stability and cross-body consistency can be improved, and contradictions or repeated triggering can be reduced.

Description

Technical Field

[0001] This invention relates to the field of service level evaluation technology for highway service areas, and specifically to a service level evaluation method and system for highway service areas based on knowledge reasoning. Background Technology

[0002] In the field of service level evaluation for highway service areas, traditional evaluation methods have many limitations. On the one hand, evaluation data sources are scattered, covering various types of information such as historical rating records, indicator data, questionnaire feedback, platform reviews, and internal inspection records. The credibility, timeliness, and information content of data from different sources vary greatly, and there are problems such as duplicate descriptions and inconsistent expressions, making data integration difficult and hindering the formation of standardized evaluation inputs. On the other hand, evaluation rules mostly rely on manual construction, which not only consumes a lot of manpower but is also susceptible to subjective factors, resulting in insufficient rule stability and poor consistency in evaluations across different service area entities. At the same time, manually formulated rules cannot fully cover complex and diverse evaluation scenarios, easily leading to problems such as rule contradictions and duplicate triggers. Furthermore, the evaluation results lack a clear and interpretable chain, making it impossible to effectively trace the scoring basis, thus affecting the credibility and auditability of the evaluation results.

[0003] Furthermore, traditional evaluation methods have weak semantic understanding capabilities for text-based evaluation data, making it difficult to accurately identify synonyms and antonyms, resulting in inaccurate mapping of evaluation indicators and further reducing the scientific rigor and accuracy of service level evaluation. Summary of the Invention

[0004] In order to solve the above-mentioned technical problems, this application proposes the following technical solution: In a first aspect, embodiments of this application provide a method for evaluating the service level of highway service areas based on knowledge reasoning, including: Acquire multi-source evaluation data and preprocess the multi-source evaluation data to form a structured evaluation event transaction set; The textual evidence in the evaluation event transaction set is encoded into vectors using BERT, and unsupervised self-consistent contrastive learning is used for domain adaptation. Cosine similarity and adaptive threshold clustering are combined to form a synonym set. Candidate rules are mined from the synonym set using indicator category constraints and weighted statistics, and the quality score of each candidate rule is calculated based on consistency verification. The candidate rules are made executable and standardized, which converts them into executable reasoning rule models containing priority and mutual exclusion attributes, and generates rule files that can be loaded by the rule engine. After semantic alignment and structural transformation of the acquired fact data of the service area to be evaluated to generate standardized facts, the data is input into the rule engine loaded with the rule file. Efficient pattern matching is performed based on the Rete network, and rule actions are triggered according to priority and conflict resolution strategies. Finally, the scores or deductions of multiple rules are aggregated at the indicator level and the overall level to output the final evaluation level and interpretable link.

[0005] In one possible implementation, the step of acquiring multi-source evaluation data and preprocessing the multi-source evaluation data to form a structured evaluation event transaction set includes: Multi-source evaluation data is acquired by combining web page data collection with incremental database synchronization. The credibility weights of the multi-source evaluation data from different sources are labeled; Duplicate descriptions from the same evaluation subject within a set time window are deduplicated and merged. Identify negative words and degree words in the evaluation text, and generate corresponding polarity reversal coefficients and intensity coefficients based on the identification results; By mapping the star rating and the polarity value obtained from sentiment analysis to a unified score range, and generating evidence labels corresponding to the indicator system, a structured evaluation event transaction set is formed.

[0006] In one possible implementation, the textual evidence in the evaluation event transaction set is encoded into vectors using BERT, and unsupervised self-consistent contrastive learning is used for domain adaptation. Cosine similarity and adaptive threshold clustering are then combined to form a synonym set, including: Each piece of textual evidence in the evaluation event transaction set is input into BERT for encoding to obtain a vector; pass Word fragment masking generates two views of the same text. , and respectively encoded to obtain , ; The , As positive sample pairs, different evidence texts are used as negative samples. Combining cosine similarity and temperature coefficient, InfoNCE loss is used for unsupervised self-consistent contrastive learning optimization. After optimization, the merging threshold is adaptively determined using intra-cluster similarity, and the calculation formula is as follows: ; in, The merging threshold, , The first The mean and standard deviation of pairwise similarity within each cluster. This is the threshold adjustment coefficient; After merging, a set of synonyms is formed and the mapping table is retained.

[0007] In one possible implementation, the formula for calculating unsupervised self-consistent contrastive learning optimization using InfoNCE loss is as follows: Where L is the contrastive learning loss function; , These represent vector representations of the same piece of evidence text encoded using two different data augmentation views. The vector representation of the second view of the j-th piece of evidence text within the batch; The cosine similarity function; This is a temperature coefficient used to adjust the smoothness of the similarity distribution; This indicates summing over all negative samples within the batch; It is an exponential function.

[0008] In one possible implementation, candidate rules are mined from the synonym set under indicator category constraints and weighted statistics, and a quality score for each candidate rule is calculated based on consistency checks, including: Weighted constraint association rule mining oriented towards the indicator system is performed on the aforementioned synonym set, and the credibility of the evaluation subject, time decay and text information content are introduced into the weighted support calculation. At the same time, an adaptive minimum support is set for different indicator categories, calculated using the following formula: in, Let be the number of transactions corresponding to metric category c, and a and b be the scaling and bias hyperparameters, respectively. While generating candidate multisets according to the Apriori principle and pruning them, the index category constraint pruning is applied, prioritizing expansion within the same category of indexes, allowing cross-category connections only through bridging evidence, and filtering out weak polarity or negative conflict terms. Calculate the weighted confidence and lift, obtain candidate rules through screening, and calculate the quality score of each candidate rule based on consistency verification.

[0009] In one possible implementation, weighted constraint association rule mining oriented towards the indicator system is performed on the synonym set, and the formula for calculating the weighted support is as follows: (The formula incorporates the credibility of the evaluation subject, time decay, and textual information content.) in, The weighted support of candidate set X; Total number of transactions; For the first One transaction; 1( ) is an indicator function, when X is contained in The value is 1 if the condition is met, otherwise it is 0. For the first Transaction weight of each transaction For matters The corresponding source credibility weight, For time decay, The weights are normalized to reflect the amount of information in the text.

[0010] In one possible implementation, the weighted confidence and lift are calculated, candidate rules are obtained through screening, and the quality score of each candidate rule is calculated based on consistency verification using the following formulas: in, Indicates the candidate rules, For the rule quality score, , , , These are the weighting coefficients. For coverage, For reliability, For correlation strength, For consistency, For weighted support function, For weighted confidence functions, For the weighted lift function, For the set of antecedent conditions of the rule, For the rule consequent action conclusion item set, The value represents the rule contradiction degree, used to characterize the co-occurrence of mutually exclusive evidence and the conflict of action direction, and its value ranges from [0,1].

[0011] In one possible implementation, the candidate rules are executable and specified, transforming them into executable inference rule models that include priority and mutual exclusion attributes, and generating rule files that can be loaded by the rule engine, including: The candidate rules are made executable and standardized. Based on the service area evaluation index and action dictionary, the rule text is unified into a condition-action-attribute structure. At the same time, the confidence and source coverage are converted into rule priority and mutually exclusive attributes to form an inference rule model. Based on the rule model, Drools rule files are automatically generated, defining a unique rule ID, indicator category, salience priority, mutual exclusion group, and agenda group attributes for each rule, and mapping the mutual exclusion relationship obtained from the consistency check to a mutual exclusion group.

[0012] In one possible implementation, the acquired factual data of the service area to be evaluated undergoes semantic alignment and structural transformation to generate standardized facts, which are then input into a rule engine loaded with the rule file. Efficient pattern matching is performed based on the Rete network, and rule actions are triggered according to priority and conflict resolution strategies, including: After semantic alignment and structuring, the acquired factual data of the service area to be evaluated is converted into standardized evidential facts and indicator facts. Standardized facts are input into the rule engine, and the Drools rule files automatically generated by the rule model are loaded through the KIE container and compiled into KieBase; Evidence facts and indicator facts are inserted into Working Memory in batches. When new or updated facts are added, update is called, so that the Rete network only performs incremental propagation to the affected nodes, thus achieving fast pattern matching. When facts flow through the Rete network and satisfy the conditions of the rule-based terminal nodes, a candidate activation set is generated; The candidate activation set is sorted from high to low salience, and mutual exclusion rules are implemented by combining mutual exclusion groups. The maximum number of times the cumulative score of the same indicator is applied is truncated. When mutually exclusive evidence pairs or contradictory action directions are detected, retain the rule with the higher quality score and suppress the other rule. If necessary, perform secondary filtering on activations of low-priority or cross-category connections; The then part uniformly calls ScoreService to add or deduct points, and records the trigger rule ID, matching evidence, action score, source weight, time decay factor and rule quality score in the interpretation log, forming an auditable interpretation chain.

[0013] Secondly, embodiments of this application provide a service level evaluation system for highway service areas based on knowledge reasoning, including: The data acquisition and preprocessing module is used to acquire multi-source evaluation data and preprocess the multi-source evaluation data to form a structured evaluation event transaction set; The synonym set construction module is used to encode the textual evidence in the evaluation event transaction set into vectors using BERT, perform domain adaptation using unsupervised self-consistent contrastive learning, and form synonym sets by combining cosine similarity and adaptive threshold clustering. The rule mining and calculation module is used to mine candidate rules from the synonym set based on indicator category constraints and weighted statistics, and calculate the quality score of each candidate rule based on consistency verification. The rule modeling module is used to perform executable specification on the candidate rules, convert the candidate rules into executable reasoning rule models containing priority and mutual exclusion attributes, and generate rule files that can be loaded by the rule engine. The rule engine execution module is used to perform semantic alignment and structural transformation on the acquired fact data to be processed in the service area to be evaluated to generate standardized facts, which are then input into the rule engine loaded with the rule file. Based on the Rete network, efficient pattern matching is completed, and rule actions are triggered according to priority and conflict resolution strategies. The results aggregation and output module is used to aggregate the scores or deductions of multiple rules at the indicator level and overall level, and output the final evaluation level and interpretable link.

[0014] Compared with the prior art, the beneficial effects of this application are as follows: This application achieves domain-adaptive semantic encoding and synonym clustering of textual evidence through unsupervised self-consistent contrastive learning. It automatically mines candidate rules by combining indicator category constraints and weighted statistics, significantly reducing the workload of manual rule formulation and lowering human resource costs. Simultaneously, the rule mining process incorporates weighted factors such as the credibility of the evaluation subject, time decay, and textual information content, combined with consistency checks to calculate rule quality scores, effectively improving the stability and reliability of the rules and reducing the generation of contradictory and invalid rules.

[0015] This application leverages BERT encoding and synonym clustering to achieve a unified mapping of different expression methods, avoiding evaluation errors caused by semantic understanding biases. By setting an adaptive minimum support, it takes into account the data distribution characteristics of different indicator categories, ensuring effective coverage of long-tail key rules. During the rule execution phase, the efficient pattern matching and priority and conflict resolution strategies based on the Rete network avoid repeated or contradictory triggering, improving the consistency of cross-service area subject evaluations and the accuracy of evaluation results.

[0016] This application can reduce the human cost of rule construction, improve rule stability and cross-entity consistency, reduce contradictions or repeated triggering, thereby improving the interpretability and auditability of evaluation results. Attached Figure Description

[0017] Figure 1A flowchart illustrating a knowledge-based reasoning-based service level evaluation method for highway service areas, provided as an embodiment of this application; Figure 2 A flowchart of rule matching and judgment provided for embodiments of this application; Figure 3 This is a schematic diagram illustrating the structure of the Drools inference engine principle provided in an embodiment of this application. Figure 4 This is a schematic diagram of a highway service area service level evaluation system based on knowledge reasoning, provided as an embodiment of this application. Detailed Implementation

[0018] The present solution will now be described in conjunction with the accompanying drawings and specific embodiments.

[0019] Figure 1 A flowchart illustrating a knowledge-based reasoning-based service level evaluation method for highway service areas, provided in this application embodiment, is shown below. Figure 1 This embodiment of a knowledge-based reasoning-based method for evaluating the service level of highway service areas includes: S101, acquire multi-source evaluation data and preprocess the multi-source evaluation data to form a structured evaluation event transaction set.

[0020] In this embodiment, a combination of webpage data collection and incremental database synchronization is used to acquire multi-source evaluation data. This multi-source evaluation data includes historical rating records, indicator data, and text data. The text data includes questionnaires, platform comments, and internal inspection records. The credibility weights of multi-source evaluation data from different sources or channels are labeled. ∈[0,1], for example, internal inspection is greater than questionnaires, which are greater than platform comments. Duplicate descriptions from the same reviewer within a set time window are deduplicated and merged to reduce statistical bias caused by fake reviews and duplicate descriptions. Negative words in the review text, such as "not," "none," and "not yet," and degree words, such as "very," "relatively," and "slightly," are identified. Based on the identification results, corresponding polarity reversal coefficients and intensity coefficients are generated for subsequent weighting and action direction judgment. The star rating is... } Mapped to in The score is a linear mapping of the star rating; the emotional polarity is... Mapped to in The score is the result of the emotional polarity mapping; and according to... Normalized to [0,1], where The final score is the combined score after merging. , To integrate weighting coefficients, the star rating and sentiment score are used to represent their contribution to the overall score. Simultaneously, evidence labels corresponding to the indicator system are generated, such as hygiene, dining, safety, and parking, outputting a structured evaluation event transaction set to provide input for subsequent indicator constraint mining.

[0021] S102 encodes the textual evidence in the evaluation event transaction set into vectors using BERT, performs domain adaptation using unsupervised self-consistent contrastive learning, and combines cosine similarity with adaptive threshold clustering to form a synonym set.

[0022] In this embodiment, each piece of text evidence... Input BERT encoder Obtain vector ,pass Word fragment masking generates two views of the same text. , and respectively encoded to obtain , As positive sample pairs, different evidence texts In particular, using texts from different indicator categories as negative samples allows the model to bring synonymous expressions closer together and distance heteronyms from each other within the domain context. The loss is optimized using unsupervised self-consistent comparison learning, and the optimization formula is as follows: Where L is the contrastive learning loss function; , These represent vector representations of the same piece of evidence text encoded using two different data augmentation views. The vector representation of the second view of the j-th piece of evidence text within the batch; The cosine similarity function; This is a temperature coefficient used to adjust the smoothness of the similarity distribution; This indicates summing over all negative samples within the batch; It is an exponential function. The negative sample term in the denominator is used to construct the discriminant boundary for contrastive learning, separating the representations of different evidence or different indicator categories in the vector space. After training, the encoder is fixed. The output vector is then used for subsequent clustering and rule mining. Based on this, semantic vectors are clustered, and the merging threshold is adaptively determined through the intra-cluster similarity distribution. The calculation formula is as follows: in, , The first The mean and standard deviation of pairwise similarity within each cluster. This is the threshold adjustment coefficient. After merging, a set of synonyms is formed. And retain the mapping table Evidence text → Synonym ID, use the synonym ID to replace the original text to construct a transaction.

[0023] S103 involves mining candidate rules from the synonym set under the constraints of indicator category and weighted statistics, and calculating the quality score of each candidate rule based on consistency verification.

[0024] In this embodiment, weighted constraint association rule mining oriented towards the indicator system is performed on the synonym itemset. The credibility of the evaluation subject, time decay, and textual information content are introduced into the weighted support calculation. The calculation formula is as follows: in, The weighted support of candidate set X; Total number of transactions; For the first One transaction; 1( ) is an indicator function, when X is contained in The value is 1 if the condition is met, otherwise it is 0. For the first Transaction weight of each transaction For matters The corresponding source credibility weight is defined as a prior hyperparameter based on domain knowledge. For example, internal inspections or official monitoring have a weight of 0.9–1.0; structured questionnaires have a weight of 0.7–0.8; and platform comments or unstructured text have a weight of 0.5–0.6. The decay is time-dependent, where Δt is the time interval from the current point, and γ is the decay coefficient. Normalized weights for text information content, Let x be the length of the text. This represents the maximum text length within the same dataset or time window.

[0025] At the same time, an adaptive minimum support is set for different indicator categories, calculated using the following formula: in, Let be the number of transactions corresponding to indicator category c, and let a and b be the scaling and bias hyperparameters, respectively. These are used to suppress noise in small sample categories and cover long-tail but critical rules in large sample categories, such as low-frequency, high-risk security events. a and b can be determined through anchor point calibration: given ( )and( Solve the system of equations. , Received, among which , These are the minimum and maximum reference values ​​for the sample size. , The upper and lower bounds of the minimum support reached at the sample extrema; or performing a grid search on (a,b) on the validation set to minimize the conflict rate and repetition rate of the rule set.

[0026] While generating candidate multisets according to the Apriori principle and pruning them, the indicator category constraint pruning is applied, prioritizing expansion within the same category of indicators, allowing cross-category connections only through bridging evidence, and filtering out weak polarity or negative conflict items. Bridging evidence refers to composite evidence or shared facility evidence that can simultaneously characterize two indicator categories. For example, "overflowing trash cans" can bridge hygiene and environmental management, and "missing warning signs" can bridge safety and management regulations.

[0027] Finally, the weighted confidence score and lift are calculated using the following formula: Association rules are obtained through filtering, and rule quality scores are calculated based on rule consistency verification. The co-occurrence rate of mutually exclusive evidence is used to penalize unreasonable co-occurrences (such as "warning signs have been set up" and "warning signs are missing"), while the contradiction degree of action direction is used to penalize opposite actions under similar conditions. The calculation formulas are as follows: in, Indicates the candidate rules, For the rule quality score, , , , These are the weighting coefficients, and they satisfy... , For coverage, For reliability, For correlation strength, For consistency, For weighted support function, For weighted confidence functions, For the weighted lift function, For the set of antecedent conditions of the rule, For the rule consequent action conclusion item set, The rule contradiction degree is used to characterize the co-occurrence of mutually exclusive evidence and the conflict of action direction. The value range is [0,1], which can be obtained by weighting the co-occurrence rate of mutually exclusive evidence and the contradiction degree of action direction.

[0028] S104, executes the candidate rules into executable rules, converts the candidate rules into executable reasoning rule models containing priority and mutual exclusion attributes, and generates rule files that the rule engine can load.

[0029] In this embodiment, candidate rules are made executable and standardized. Based on the service area evaluation index and action dictionary, the rule text is unified into a condition-action-attribute structure. The condition part supports single conditions, multiple conditions, and AND and OR logical connections, and extracts limitation information such as negation, range, and threshold through rule templates. The action part extracts constraints such as addition and deduction values, applicable indicators, upper limit of score, and upper limit of frequency. At the same time, confidence and source coverage are converted into rule priority and mutual exclusion and agenda attributes to form a rule model that can be directly consumed by the inference engine.

[0030] For language phenomena such as "and / or not satisfied" or "not set", the evidence items and threshold conditions are standardized into executable predicates. For example, if the rule condition is "selling alcoholic beverages and not setting warning signs", the transformed condition is evidence('selling alcoholic beverages') and not evidence('warning signs have been set'), the action conclusion is to deduct 10 points, and the statistical attributes of this rule are mapped to priority and mutual exclusion constraints.

[0031] Drools rule files are automatically generated based on the rule model, defining a unique rule ID, metric category, salience priority, mutual exclusion group, and agenda group attributes for each rule. The salience is determined by the rule quality score. Automatically generated, its calculation formula is as follows: in For the rules, The quality score for the rule, This is a Sigmoid function. It maps the mutual exclusion relationships obtained from consistency checks to mutual exclusion groups to avoid conflicting rules being triggered simultaneously. The `when` part binds standardized Facts such as `EvidenceFact` and `IndicatorFact` and references the synonym set mapping results to complete evidence matching. The `then` part calls the unified `ScoreService` to deduct or add points and record explanation information. It iterates through the matched rule set `R_match` in descending order of `salience`. If the activation-group to which rule `r` belongs has already been triggered, it skips it. If triggering causes the score of the same indicator to exceed the upper limit `cap`, it truncates it. If there is a conflicting evidence pair with a triggered rule (given by the mutual exclusion table `MutexEvidence`), it retains the one with higher `Q(r)`. Otherwise, it executes `r` and records the explanation. For example, a rule can be represented as "rule 'Safety Management - Restaurant Settings' salience50 when evidence('Selling Alcoholic Beverages') and not evidence('Warning Signs Set') thenscore(10,'Safety Management') end".

[0032] S105, after semantic alignment and structural transformation of the acquired fact data to be processed in the service area to be evaluated to generate standardized facts, the data is input into the rule engine loaded with the rule file. Efficient pattern matching is performed based on the Rete network, and rule actions are triggered according to priority and conflict resolution strategies.

[0033] In this embodiment, inference execution employs fast pattern matching based on the Rete algorithm, combined with rule priority and conflict resolution strategies to achieve stable output. During runtime, service area facts are input by staff or system interfaces (inserted into working memory), and the inference engine completes fact propagation and node matching in the Rete network. When multiple rules are satisfied simultaneously, high-quality rules are triggered first based on salience, and the final trigger set is selected by combining mutual exclusion groups, action limit constraints, and contradiction detection results to avoid duplicate deductions or mutual contradictions. After the rules are triggered, the scores and deduction results are aggregated according to indicator dimensions and an explanation link is generated, ultimately outputting a comprehensive service level evaluation result.

[0034] See Figure 2 and Figure 3In this embodiment, the specific implementation steps based on the Rete algorithm include: abstracting the input data into standardized Facts, including EvidenceFact and IndicatorFact. EvidenceFact contains fields such as synonym ID / evidence label, category of indicator, polarity, intensity coefficient, source subject weight, timestamp, and original text. IndicatorFact contains structured monitoring values ​​such as number of parking spaces, congestion level, and equipment online rate, as well as fields required for threshold comparison.

[0035] The inspection text / user evaluation is processed by sentence denoising, negation, degree word processing, and sentiment star rating mapping. The text is then vectorized and mapped to a synonym set to obtain a standardized evidence set. Numerical indicators undergo unit conversion, missing word completion, and interval / threshold discretization to form an IndicatorFact set. The generated DRL rule file is loaded into a KIE container and compiled into KieBase. During compilation, the "when" part of the rule is decomposed into Alpha nodes for single-condition filtering and Beta nodes for multi-condition connection, and terminal nodes are constructed to support incremental matching. Runtime metadata is generated for the rule's salience, agenda group, and mutual exclusion group constraints. EvidenceFact and IndicatorFact are batch-inserted into Working Memory. When a new / updated fact occurs, the update function is called, ensuring the Rete network only performs incremental propagation to affected nodes, avoiding full recalculation and achieving fast pattern matching. When a fact flows through the Rete network and satisfies the rule's terminal node conditions, a candidate activation set is generated. Candidate activations carry the necessary information for interpretation, including the matched evidence ID, indicator category, action score, source coverage, and rule quality score. Sort by salience from highest to lowest, and implement mutual exclusion rules using activation-groups; apply a maximum limit to the number of times the cumulative score for the same indicator is applied; retain the rule quality score when mutually exclusive evidence pairs or contradictory action directions are detected. The higher priority rule suppresses the other rule; if necessary, the AgendaFilter can be used to further filter the activation of low-priority or cross-category connections. Then, the ScoreService is uniformly invoked to add or deduct points, and the triggering rule ID, matching evidence (synonym ID and original text fragment), action score, source weight, time decay factor, and rule quality score are recorded in the explanation log. This forms an auditable explanation chain.

[0036] S106 Finally, the scores or deductions of multiple rules are aggregated at the indicator level and the overall level to output the final evaluation level and interpretable link.

[0037] In this embodiment, the cumulative scores in ScoreService are aggregated according to the indicator dimensions to obtain the scores of each indicator and the total score; the service level is output according to the preset level mapping, and an explanation link is carried for review and tracing.

[0038] Corresponding to the knowledge-reasoning-based highway service area service level evaluation method provided in the above embodiments, this application also provides an embodiment of a knowledge-reasoning-based highway service area service level evaluation system.

[0039] See Figure 4 This application provides a knowledge-based reasoning-based highway service area service level evaluation system 20, comprising: The data acquisition and preprocessing module 201 is used to acquire multi-source evaluation data and preprocess the multi-source evaluation data to form a structured evaluation event transaction set.

[0040] The synonym itemset construction module 202 is used to encode the textual evidence in the evaluation event transaction set into vectors using BERT, perform domain adaptation using unsupervised self-consistent contrastive learning, and form synonym itemsets by combining cosine similarity and adaptive threshold clustering.

[0041] The rule mining and calculation module 203 is used to mine candidate rules from the synonym set based on the indicator category constraints and weighted statistics, and calculate the quality score of each candidate rule based on consistency verification.

[0042] The rule modeling module 204 is used to perform executable specification on candidate rules, converting candidate rules into executable inference rule models that include priority and mutual exclusion attributes, and generating rule files that can be loaded by the rule engine.

[0043] The rule engine execution module 205 is used to perform semantic alignment and structural transformation on the acquired fact data to be processed in the service area to be evaluated to generate standardized facts, and then input them into the rule engine loaded with the rule file. Based on the Rete network, it completes efficient pattern matching and triggers rule actions according to priority and conflict resolution strategies.

[0044] The result aggregation and output module 206 is used to aggregate the scores or deductions of multiple rules at the indicator level and the overall level, and output the final evaluation level and interpretable link.

[0045] In this embodiment, "multiple" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent the existence of A alone, the simultaneous existence of A and B, or the existence of B alone. A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0046] It should be noted that, in this document, 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 a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0047] The above description is merely a specific embodiment of this application. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the protection scope of this application. The protection scope of this application should be determined by the protection scope of the claims.

Claims

1. A method for evaluating the service level of highway service areas based on knowledge reasoning, characterized in that, include: Acquire multi-source evaluation data and preprocess the multi-source evaluation data to form a structured evaluation event transaction set; The textual evidence in the evaluation event transaction set is encoded into vectors using BERT, and unsupervised self-consistent contrastive learning is used for domain adaptation. Cosine similarity and adaptive threshold clustering are combined to form a synonym set. Candidate rules are mined from the synonym set using indicator category constraints and weighted statistics, and the quality score of each candidate rule is calculated based on consistency verification. The candidate rules are made executable and standardized, which converts them into executable reasoning rule models containing priority and mutual exclusion attributes, and generates rule files that can be loaded by the rule engine. After semantic alignment and structural transformation of the acquired fact data of the service area to be evaluated to generate standardized facts, the data is input into the rule engine loaded with the rule file. Efficient pattern matching is performed based on the Rete network, and rule actions are triggered according to priority and conflict resolution strategies. Finally, the scores or deductions of multiple rules are aggregated at the indicator level and the overall level to output the final evaluation level and interpretable link.

2. The method for evaluating the service level of highway service areas based on knowledge reasoning according to claim 1, characterized in that, The process of acquiring multi-source evaluation data and preprocessing the multi-source evaluation data to form a structured evaluation event transaction set includes: Multi-source evaluation data is acquired by combining web page data collection with incremental database synchronization. The credibility weights of the multi-source evaluation data from different sources are labeled; Duplicate descriptions from the same evaluation subject within a set time window are deduplicated and merged. Identify negative words and degree words in the evaluation text, and generate corresponding polarity reversal coefficients and intensity coefficients based on the identification results; By mapping the star rating and the polarity value obtained from sentiment analysis to a unified score range, and generating evidence labels corresponding to the indicator system, a structured evaluation event transaction set is formed.

3. The method for evaluating the service level of highway service areas based on knowledge reasoning according to claim 1, characterized in that, The textual evidence in the evaluation event transaction set is encoded into vectors using BERT, and unsupervised self-consistent contrastive learning is used for domain adaptation. Cosine similarity and adaptive threshold clustering are then combined to form synonym sets, including: Each piece of textual evidence in the evaluation event transaction set is input into BERT for encoding to obtain a vector; pass Word fragment masking generates two views of the same text. , and respectively encoded to obtain , ; The , As positive sample pairs, different evidence texts are used as negative samples. Combining cosine similarity and temperature coefficient, InfoNCE loss is used for unsupervised self-consistent contrastive learning optimization. After optimization, the merging threshold is adaptively determined using intra-cluster similarity, and the calculation formula is as follows: ； in, The merging threshold, , The first The mean and standard deviation of pairwise similarity within each cluster. This is the threshold adjustment coefficient; After merging, a set of synonyms is formed and the mapping table is retained.

4. The method for evaluating the service level of highway service areas based on knowledge reasoning according to claim 3, characterized in that, The formula for optimizing unsupervised self-consistent contrastive learning using InfoNCE loss is as follows: Where L is the contrastive learning loss function; , These represent vector representations of the same piece of evidence text encoded using two different data augmentation views. The vector representation of the second view of the j-th piece of evidence text within the batch; The cosine similarity function; This is a temperature coefficient used to adjust the smoothness of the similarity distribution; This indicates summing over all negative samples within the batch; It is an exponential function.

5. The method for evaluating the service level of highway service areas based on knowledge reasoning according to claim 1, characterized in that, Candidate rules are mined from the synonym set using indicator category constraints and weighted statistics, and a quality score for each candidate rule is calculated based on consistency checks, including: Weighted constraint association rule mining oriented towards the indicator system is performed on the aforementioned synonym set, and the credibility of the evaluation subject, time decay and text information content are introduced into the weighted support calculation. At the same time, an adaptive minimum support is set for different indicator categories, calculated using the following formula: in, is the number of transactions corresponding to metric category c, and a and b are the scaling and bias hyperparameters, respectively. While generating candidate multisets according to the Apriori principle and pruning them, the index category constraint pruning is applied, prioritizing expansion within the same category of indexes, allowing cross-category connections only through bridging evidence, and filtering out weak polarity or negative conflict terms. Calculate the weighted confidence and lift, obtain candidate rules through screening, and calculate the quality score of each candidate rule based on consistency verification.

6. The method for evaluating the service level of highway service areas based on knowledge reasoning according to claim 5, characterized in that, Weighted constraint association rule mining based on the indicator system is performed on the aforementioned synonym set. The formula for weighted support calculation incorporates the credibility of the evaluation subject, time decay, and textual information content as follows: in, The weighted support of candidate set X; Total number of transactions; For the first One transaction; 1( ) is an indicator function, when X is contained in The value is 1 if the condition is met, otherwise it is 0. For the first Transaction weight of each transaction For matters The corresponding source credibility weight, For time decay, The weights are normalized to reflect the amount of information in the text.

7. The method for evaluating the service level of highway service areas based on knowledge reasoning according to claim 5, characterized in that, The weighted confidence and lift are calculated, candidate rules are obtained through screening, and the quality score of each candidate rule is calculated based on consistency verification. The formulas are as follows: in, Indicates the candidate rules, For the rule quality score, , , , These are the weighting coefficients. For coverage, For reliability, For correlation strength, For consistency, For weighted support function, For weighted confidence functions, For the weighted lift function, For the set of antecedent conditions of the rule, For the rule consequent action conclusion item set, The value represents the rule contradiction degree, used to characterize the co-occurrence of mutually exclusive evidence and the conflict of action direction, and its value ranges from [0,1].

8. The method for evaluating the service level of highway service areas based on knowledge reasoning according to claim 1, characterized in that, The candidate rules are made executable by specification, which transforms them into executable inference rule models containing priority and mutual exclusion attributes, and generates rule files that can be loaded by the rule engine, including: The candidate rules are made executable and standardized. Based on the service area evaluation index and action dictionary, the rule text is unified into a condition-action-attribute structure. At the same time, the confidence and source coverage are converted into rule priority and mutually exclusive attributes to form an inference rule model. Based on the rule model, Drools rule files are automatically generated, defining a unique rule ID, indicator category, salience priority, mutual exclusion group, and agenda group attributes for each rule, and mapping the mutual exclusion relationship obtained from the consistency check to a mutual exclusion group.

9. The method for evaluating the service level of highway service areas based on knowledge reasoning according to claim 1, characterized in that, After semantic alignment and structural transformation of the acquired factual data of the service area to be evaluated to generate standardized facts, the data is input into a rule engine loaded with the rule file. Efficient pattern matching is performed based on the Rete network, and rule actions are triggered according to priority and conflict resolution strategies, including: After semantic alignment and structuring, the acquired factual data of the service area to be evaluated is converted into standardized evidential facts and indicator facts. Standardized facts are input into the rule engine, and the Drools rule files automatically generated by the rule model are loaded through the KIE container and compiled into KieBase; Evidence facts and indicator facts are inserted into Working Memory in batches. When new or updated facts are added, update is called, so that the Rete network only performs incremental propagation to the affected nodes, thus achieving fast pattern matching. When facts flow through the Rete network and satisfy the conditions of the rule-based terminal nodes, a candidate activation set is generated; The candidate activation set is sorted from high to low salience, and mutual exclusion rules are implemented by combining mutual exclusion groups. The maximum number of times the cumulative score of the same indicator is applied is truncated. When mutually exclusive evidence pairs or contradictory action directions are detected, retain the rule with the higher quality score and suppress the other rule. If necessary, perform secondary filtering on activations of low-priority or cross-category connections; The then part uniformly calls ScoreService to add or deduct points, and records the trigger rule ID, matching evidence, action score, source weight, time decay factor and rule quality score in the interpretation log, forming an auditable interpretation chain.

10. A service level evaluation system for highway service areas based on knowledge reasoning, characterized in that, include: The data acquisition and preprocessing module is used to acquire multi-source evaluation data and preprocess the multi-source evaluation data to form a structured evaluation event transaction set; The synonym set construction module is used to encode the textual evidence in the evaluation event transaction set into vectors using BERT, perform domain adaptation using unsupervised self-consistent contrastive learning, and form synonym sets by combining cosine similarity and adaptive threshold clustering. The rule mining and calculation module is used to mine candidate rules from the synonym set based on indicator category constraints and weighted statistics, and calculate the quality score of each candidate rule based on consistency verification. The rule modeling module is used to perform executable specification on the candidate rules, convert the candidate rules into executable reasoning rule models containing priority and mutual exclusion attributes, and generate rule files that can be loaded by the rule engine. The rule engine execution module is used to perform semantic alignment and structural transformation on the acquired fact data to be processed in the service area to be evaluated to generate standardized facts, which are then input into the rule engine loaded with the rule file. Based on the Rete network, efficient pattern matching is completed, and rule actions are triggered according to priority and conflict resolution strategies. The results aggregation and output module is used to aggregate the scores or deductions of multiple rules at the indicator level and overall level, and output the final evaluation level and interpretable link.

Images