Government affair hotline satisfaction degree prediction method and system based on multi-source feature vector fusion

Abstract

The application provides a government hotline satisfaction prediction method and system based on multi-source feature vector fusion, and relates to the technical field of government hotline intelligent analysis, which comprises the following steps: collecting government hotline business data and performing preprocessing to output standardized government hotline business data set; extracting direct business feature vector, constructing hotline field new word library based on hotline unlabeled corpus and performing field self-adaption pre-training on language model; performing improved emotion extraction on appeal text to output emotion feature vector; performing text feature quantization and syntax structure analysis on appeal text to output text feature vector, fusing direct business feature vector, emotion feature vector and text feature vector to output satisfaction prediction feature vector and input into gradient boosting decision tree model, and outputting work order satisfaction prediction result, risk level, recommended revisit mode and key influence factor explanation, so as to improve the accuracy, stability and explainability of emotion recognition and satisfaction prediction in the government hotline scene.

Description

Technical Field

[0001] This invention relates to the field of intelligent analysis technology for government hotlines, and in particular to a method and system for predicting government hotline satisfaction based on multi-source feature vector fusion. Background Technology

[0002] Government hotlines are an important service channel for the government to connect with the public. Hotline request records contain multi-source business data, including the request text, basic attributes, processing procedures, follow-up status, and follow-up results. Although existing hotline platforms can complete the work order flow and result follow-up, they still mainly rely on human experience to judge questions such as "which work orders are more likely to cause dissatisfaction," "which texts contain implicit negative emotions," and "how to provide risk warnings and follow-up suggestions in advance."

[0003] Existing satisfaction assessment schemes typically suffer from the following shortcomings: First, they lack domain adaptability to hotline texts, making it difficult for general sentiment classifiers to accurately identify euphemisms of dissatisfaction, compound complaints, rhetorical questions, and policy terminology in government contexts. Second, they fail to adequately utilize new hotline terms, compound phrases, and industry buzzwords, leading to decreased accuracy in text segmentation and sentiment trigger word recognition. Third, most schemes use only single text features or single business features, neglecting the synergistic effects between the request text, process indicators, processing results, and citizen attributes. Fourth, they lack a mechanism for linking predictive interpretation with follow-up strategies for actual business operations.

[0004] Therefore, it is necessary to provide a method and system for predicting government hotline satisfaction based on multi-source feature vector fusion to solve the above-mentioned technical problems. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a method and system for predicting government hotline satisfaction based on multi-source feature vector fusion. This method solves the problems in existing technologies, such as insufficient adaptation of emotion recognition to domain corpora, inadequate utilization of new hotline terms, weak feature fusion capability for satisfaction prediction, and poor interpretability of results.

[0006] The present invention provides a method for predicting satisfaction with government hotlines based on multi-source feature vector fusion, the method comprising: Collect government hotline business data and perform standardized preprocessing to output a standardized government hotline business dataset. Direct business feature vectors are extracted from the standardized government hotline business dataset. At the same time, based on the unlabeled hotline corpus in the standardized government hotline business dataset, a new hotline domain lexicon is constructed and the basic language model is pre-trained in a domain-adaptive manner. Based on the new hotline domain lexicon and the domain-adaptive pre-trained basic language model, an improved sentiment extraction method is performed on the petition text in the standardized government hotline business dataset, and the sentiment feature vector of the petition is output. Based on the new thesaurus in the hotline field, text feature quantification and syntactic structure analysis are performed on the appeal text to output the appeal text feature vector. Multi-source feature fusion is then performed on the direct business feature vector, the appeal sentiment feature vector, and the appeal text feature vector to output the satisfaction prediction feature vector. The satisfaction prediction feature vector is input into the pre-trained gradient boosting decision tree model, which outputs the satisfaction prediction result, risk level, recommended follow-up method, and explanation of key influencing factors of satisfaction prediction for the government hotline work order.

[0007] Preferably, the step of collecting government hotline business data and performing standardized preprocessing to output a standardized government hotline business dataset specifically includes: Collect the business data of the government hotline; The collected government hotline business data is standardized and its integrity is verified. Data with missing values ​​exceeding the preset missing value threshold and logical conflicts are removed, and the preliminary government hotline business dataset is output. Data anonymization processing is performed on the initially screened government hotline business dataset, and privacy fields are encrypted and anonymized to output the anonymized government hotline business dataset. The de-identified government hotline business dataset is subjected to deduplication and unified encoding processing to remove duplicate work order data. All fields are subjected to unified format encoding and numerical normalization to output a standardized government hotline business dataset. Perform text preprocessing on the request text field in the standardized government hotline business dataset, including sentence segmentation, word segmentation, character normalization, typo correction and stop word filtering, process any request text into an ordered word sequence, and output the text-standardized government hotline business dataset. The standardized government hotline business dataset is mapped and categorized for storage. It is divided into subsets of basic information of requests, subsets of call record information, subsets of follow-up record information, subsets of process record information, subsets of return visit record information, and subsets of basic citizen information. The standardized government hotline business dataset is then output.

[0008] Preferably, the step of constructing a new thesaurus for the hotline domain based on the unlabeled hotline corpus in the standardized government hotline business dataset specifically includes: Extract the unlabeled corpus of hotline data from the standardized government hotline business dataset, perform co-occurrence statistics on candidate words w in the unlabeled corpus, and calculate the co-occurrence of two adjacent candidate words. mutual information That is, two adjacent candidate words The probability of co-occurrence in the unlabeled corpus of the hotline, and based on two adjacent candidate words. mutual information The average mutual information PMI(w) between the candidate word w and the domain sentiment trigger word is calculated using the following formula: In the formula, They are respectively The probability of appearing in the unlabeled corpus of the hotline; Calculate the left information entropy of the candidate word w. The corresponding calculation formula is as follows: In the formula, linL(w) is the set of all left-side adjacent characters of candidate word w; P(l|w) is the conditional probability that the left-side adjacent character l appears to the left of candidate word w; Calculate the right information entropy of the candidate word w. The corresponding calculation formula is as follows: In the formula, rinR(w) is the set of all right-side adjacent characters of candidate word w; P(r|w) is the conditional probability that the right-side adjacent character r appears to the right of candidate word w. The inverse document frequency (IDF) (w) of the candidate word w is calculated using the following formula: In the formula, N is the total number of documents in the unannotated corpus of the hotline; df(w) is the number of documents in the unannotated corpus of the hotline that contain the candidate word w; Calculate the comprehensive score of the candidate word w. The corresponding calculation formula is as follows: In the formula, Left information entropy of candidate word w With right information entropy The minimum value in; The overall score Candidate words w that exceed a preset scoring threshold are added as new words in the hotline field new word library.

[0009] Preferably, the improved sentiment extraction of requests from the standardized government hotline business dataset, based on the new lexicon of the hotline domain and the domain-adaptive pre-trained basic language model, and outputting a sentiment feature vector, specifically includes: Extract words from the request text in the standardized government hotline business dataset. Corresponding word vectors ; Based on the new lexicon in the hotline field, calculate words Corresponding domain dictionary feature vector The corresponding calculation formula is as follows: In the formula, For words Word frequencies in the unlabeled corpus of the hotline; Words contained in the unannotated corpus of the hotline The number of documents; beta is the preset adjustment coefficient; For words Average mutual information with domain-specific emotion trigger words; Extract words from the stated appeal text The corresponding punctuation, tone, adverbs, negation words, and intensity words are used to output words. tone auxiliary feature vector ; Extract words from the stated appeal text The corresponding word position information and syntactic auxiliary features are used to output the word. Positional Syntactic Feature Vector ; The word vector The domain dictionary feature vector The tone auxiliary feature vector The positional syntax feature vector To combine and construct words Corresponding emotional appeal extraction input vector .

[0010] Preferably, the emotional appeal extraction input vector is based on a domain-adaptive pre-trained base language model. Perform context encoding and output words. Corresponding emotional context encoding vector The corresponding calculation formula is as follows: In the formula, BiGRU() is a bidirectional gated loop unit; For words The corresponding emotional context encoding vector; Calculation words Corresponding attention scoring vector The corresponding calculation formula is as follows: In the formula, tanh() is the hyperbolic tangent activation function; is the learnable attention scoring weight matrix; b is the attention scoring bias term; Calculation words Corresponding attention weights The corresponding calculation formula is as follows: In the formula, v is a learnable attention weight vector; For words The corresponding attention scoring vector; exp() is an exponential function with the natural constant e as the base; The formula for calculating the single-view global sentiment representation z is as follows: ; The single-view global sentiment representations corresponding to the title, body, and call summary of the appeal text are weighted and fused to output a multi-view fused global sentiment representation. The corresponding calculation formula is as follows: In the formula, These are single-perspective global sentiment representations corresponding to the title, body, and call summary of the appeal text, respectively. The fusion weight coefficients are learnable, and .

[0011] Preferably, the multi-view fusion global sentiment representation is performed through a fully connected layer. Perform feature mapping and output the sentiment fully connected feature vector; The sentiment fully connected feature vector is normalized, and the sentiment category probability vector C corresponding to the appeal text is calculated using the following formula: In the formula, The output layer weight matrix is ​​a learnable matrix. The output layer bias term; softmax() is the normalized exponential function; Based on the sentiment category probability vector C, the positive sentiment probability is split. Neutral emotional probability With the probability of negative emotions And calculate the emotional intensity score corresponding to the appeal text. The corresponding calculation formula is as follows: ; The emotion category probability vector C and the emotion intensity score are used as the basis for the expression. The data is integrated to construct the emotional feature vector corresponding to the stated appeal text.

[0012] The preferred joint training and iterative optimization process of the improved appeal sentiment extraction algorithm is as follows: Multiple teacher models are trained based on manually labeled sentiment samples. The unlabeled hotline data is then input into the multiple teacher models, and the predicted probability distribution of the unlabeled hotline samples is output. Based on the predicted probability distribution, the prediction confidence of the unlabeled samples of the hotline is calculated using the following formula: In the formula, The prediction confidence level for the i-th unlabeled hotline sample; Predicted probability distribution for the i-th unlabeled hotline sample Information entropy; K is the total number of sentiment categories; Select unlabeled hotline samples whose prediction confidence is greater than or equal to a preset confidence threshold and whose prediction results from multiple teacher models are consistent. Generate pseudo-labels for the selected unlabeled hotline samples and output a training sample set with pseudo-labels. The formula for calculating the pseudo-labels is as follows: In the formula, is the pseudo-label for the i-th unlabeled hotline sample; argmax() is the function that takes the category corresponding to the maximum probability; tau is the preset confidence threshold; This is an indicator that the prediction results of multiple teacher models are consistent; it takes a value of 1 when they are consistent and a value of 0 when they are inconsistent. The manually labeled sentiment samples are merged with the set of training samples with pseudo-labels to construct a joint training sample set. A weighted focus loss function is constructed based on the joint training sample set, and the training loss calculation result of the improved appeal sentiment extraction algorithm is output. The learnable parameters of the improved appeal sentiment extraction algorithm are optimized through the backpropagation algorithm to complete the joint training and iterative optimization of the improved appeal sentiment extraction algorithm.

[0013] Preferably, the step of performing text feature quantification and syntactic structure analysis on the appeal text based on the new thesaurus in the hotline domain, outputting an appeal text feature vector, and performing multi-source feature fusion on the direct business feature vector, the appeal sentiment feature vector, and the appeal text feature vector to output a satisfaction prediction feature vector, specifically includes: Based on the new thesaurus in the hotline field, text feature quantification is performed on the appeal text to extract a statistical text feature vector; wherein, the statistical text feature vector includes the basic word count, part-of-speech count, average sentence length, stop word ratio, and linking word frequency; Based on the new lexicon in the hotline field, syntactic structure analysis is performed on the appeal text to extract the language structural text feature vector of the appeal text; wherein, the language structural text feature vector includes minimum edit distance, dependency syntax ratio, inter-sentence similarity, and narrative integrity index; The statistical text feature vector is concatenated with the language structural text feature vector to output the appeal text feature vector; Obtain the direct business feature vector and the appeal sentiment feature vector, and perform feature alignment and standardization processing with the appeal text feature vector to output the aligned direct business feature vector, appeal sentiment feature vector and appeal text feature vector. The aligned direct business feature vector, appeal sentiment feature vector, and appeal text feature vector are concatenated and fused to construct the satisfaction prediction feature vector. The corresponding calculation formula is as follows: In the formula, Let be the feature vector for predicting the satisfaction level of the j-th government hotline work order; Let be the direct business feature vector of the j-th government hotline work order; The emotional intensity score of the request text of the j-th government hotline work order; Let be the probability vector of the sentiment category of the request text of the j-th government hotline work order; Let be the feature vector of the request text of the j-th government hotline work order.

[0014] Preferably, the step of inputting the satisfaction prediction feature vector into the pre-trained gradient boosting decision tree model and outputting the satisfaction prediction result, risk level, recommended follow-up method, and explanation of key influencing factors of satisfaction prediction for the government hotline work order specifically includes: The satisfaction prediction feature vector is input into the pre-trained gradient boosting decision tree model. Multiple weak learners of the gradient boosting decision tree model are used to predict and calculate the satisfaction prediction feature vector, and the prediction result of each weak learner is output. The weighted summation of the prediction results from each weak learner yields the ensemble output of the gradient boosting decision tree model, calculated using the following formula: In the formula, To improve the gradient boosting decision tree for predicting the satisfaction level of the i-th government hotline work order, we need to define the feature vector. The integrated output value; M is the total number of weak learners; Let m be the learning rate of the m-th weak learner; Let m be the feature vector of the satisfaction prediction by the m-th weak learner. The predicted output value; The satisfaction probability of the government hotline work order is calculated by mapping the integrated output value through an activation function, and the corresponding calculation formula is as follows: In the formula, Let be the satisfaction probability of the i-th government hotline work order; sigmoid() is the sigmoid activation function; A preset probability threshold is obtained, including a first probability threshold and a second probability threshold. Based on the preset probability threshold, the government hotline work order is classified into risk levels: when the satisfaction probability is less than the first probability threshold, the government hotline work order is classified into a high-risk unsatisfactory level; when the satisfaction probability is greater than or equal to the first probability threshold and less than the second probability threshold, the government hotline work order is classified into a medium-risk pending verification level; when the satisfaction probability is greater than or equal to the second probability threshold, the government hotline work order is classified into a low-risk satisfaction level. Based on the risk level of the government hotline work order, a corresponding preset follow-up strategy rule is matched and a recommended follow-up method is generated: a manual follow-up method is recommended for government hotline work orders with a high-risk dissatisfaction level; a manual review method or a semi-automatic follow-up method is recommended for government hotline work orders with a medium-risk pending verification level; and an automatic follow-up method is recommended for government hotline work orders with a low-risk satisfaction level. Based on the feature contribution value calculation logic of the gradient boosting decision tree model, the contribution of each dimension of the feature in the satisfaction prediction feature vector is sorted, and the core influencing feature set that has the greatest impact on the satisfaction prediction result is selected. Combined with the negative emotional segments and business process abnormal indicators in the appeal text, the explanation of the key influencing factors of the satisfaction prediction is generated. Based on the explanation of the key influencing factors for the satisfaction prediction, the recommended follow-up method is optimized and adjusted, and key reminders for the follow-up are added.

[0015] A government hotline satisfaction prediction system based on multi-source feature vector fusion, the system comprising: The data acquisition and preprocessing module is used to collect government hotline business data and perform standardized preprocessing to output a standardized government hotline business dataset. The new lexicon construction and model pre-training module is used to extract direct business feature vectors from the standardized government hotline business dataset, and at the same time, based on the unlabeled hotline corpus in the standardized government hotline business dataset, construct a new lexicon for the hotline domain and perform domain-adaptive pre-training on the basic language model. The emotional appeal extraction module is used to perform improved emotional appeal extraction on the appeal text in the standardized government hotline business dataset based on the new lexicon of the hotline domain and the basic language model after domain adaptive pre-training, and output the emotional appeal feature vector. The appeal text output and feature fusion module is used to perform text feature quantification and syntactic structure analysis on the appeal text based on the new thesaurus in the hotline domain, output the appeal text feature vector, and perform multi-source feature fusion on the direct business feature vector, the appeal sentiment feature vector and the appeal text feature vector to output the satisfaction prediction feature vector. The satisfaction prediction result output module is used to input the satisfaction prediction feature vector into the pre-trained gradient boosting decision tree model and output the satisfaction prediction result, risk level, recommended follow-up method, and explanation of key influencing factors of satisfaction prediction for the government hotline work order.

[0016] Compared with related technologies, the government hotline satisfaction prediction method and system based on multi-source feature vector fusion provided by this invention has the following beneficial effects: This invention collects government hotline business data and performs standardized preprocessing to output a standardized government hotline business dataset. It extracts direct business feature vectors from this dataset and, based on unlabeled hotline corpora within the dataset, constructs a new hotline-specific lexicon and performs domain-adaptive pre-training on the basic language model. Based on the new hotline-specific lexicon and the domain-adaptive pre-trained basic language model, it performs improved sentiment extraction on the request texts in the standardized government hotline business dataset, outputting sentiment feature vectors. Finally, based on the new hotline-specific lexicon, it performs text feature quantification and syntactic structure analysis on the request texts, outputting request text feature vectors, and further analyzes the direct business features. This method integrates business feature vectors, sentiment feature vectors, and textual feature vectors to output a satisfaction prediction feature vector. This vector is then input into a pre-trained gradient boosting decision tree model to output the satisfaction prediction result, risk level, recommended follow-up method, and explanation of key influencing factors for government hotline work orders. This approach enables both domain adaptation to government hotline business corpora and the implementation of new word enhancement, sentiment extraction, text feature quantification, and classification prediction for government hotline satisfaction prediction. It improves the accuracy, stability, and interpretability of sentiment recognition and satisfaction prediction in government hotline scenarios, making it suitable for hotline follow-up warnings, service quality monitoring, and follow-up strategy linkage.

[0017] Compared to methods that only use a general sentiment classifier, this invention improves the recognition ability of policy terms, industry terms, euphemisms, rhetorical questions, and complex emotional expressions through domain-adaptive pre-training and hotline neologism enhancement. Compared to the original weighted voting transfer learning method, this invention introduces multi-granularity attention fusion and confidence pseudo-label retraining, which can improve the accuracy and generalization ability of hotline domain sentiment extraction without significantly increasing the cost of manual annotation. By retaining sentiment probability and sentiment intensity, rather than just outputting binary labels, this invention can provide more granular information for the satisfaction prediction model, improving prediction accuracy and ranking ability. By unifying the modeling of business features, sentiment features, and text features, and using gradient-boosted decision trees to output satisfaction probabilities, this invention can better adapt to the modeling needs of mixed and heterogeneous features in hotline business. In addition, the output results of this invention can be directly linked to follow-up strategies and early warning handling, which has strong business implementation value. Attached Figure Description

[0018] Figure 1 A flowchart illustrating the government hotline satisfaction prediction method based on multi-source feature vector fusion provided in this embodiment of the invention; Figure 2 A system block diagram of a government hotline satisfaction prediction system based on multi-source feature vector fusion provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] like Figure 1 The diagram shown is a flowchart of a government hotline satisfaction prediction method based on multi-source feature vector fusion provided in an embodiment of the present invention. Figure 1 The execution entity of the method shown can be a software and / or hardware device. The execution entity of this application can include, but is not limited to, at least one of the following: user equipment, network equipment, etc. User equipment can include, but is not limited to, computers, smartphones, personal digital assistants (PDAs), and the aforementioned electronic devices. Network equipment can include, but is not limited to, a single network server, a server group consisting of multiple network servers, or a cloud based on cloud computing consisting of a large number of computers or network servers. Cloud computing is a type of distributed computing, consisting of a super virtual computer composed of a group of loosely coupled computers. This embodiment does not limit this. Steps S1 to S4 are detailed as follows: S1 collects government hotline business data and performs standardized preprocessing to output a standardized government hotline business dataset. S2, extract direct business feature vectors from the standardized government hotline business dataset, and at the same time, construct a new hotline domain lexicon based on the unlabeled hotline corpus in the standardized government hotline business dataset and perform domain-adaptive pre-training on the basic language model. S3, based on the new lexicon of the hotline domain and the basic language model after domain adaptive pre-training, perform improved emotional extraction of the request text in the standardized government hotline business dataset, and output the emotional feature vector of the request. S4. Based on the new thesaurus in the hotline domain, perform text feature quantification and syntactic structure analysis on the appeal text, output the appeal text feature vector, and perform multi-source feature fusion on the direct business feature vector, the appeal sentiment feature vector and the appeal text feature vector to output the satisfaction prediction feature vector. S5, input the satisfaction prediction feature vector into the pre-trained gradient boosting decision tree model, and output the satisfaction prediction result, risk level, recommended follow-up method, and explanation of key influencing factors of satisfaction prediction for the government hotline work order.

[0021] In practical applications, the system collects multi-source business data, including basic information on requests, call records, follow-up records, process records, return visit records, and basic citizen information, for all dimensions of government hotline business scenarios. Simultaneously, it performs standardized preprocessing operations, covering field cleaning, data anonymization, deduplication, unified encoding, and format standardization. At the same time, it performs text standardization processing on the request text, such as sentence segmentation, word segmentation, character normalization, typo correction, and stop word filtering. Finally, it outputs a compliant, standardized, and non-redundant government hotline business dataset.

[0022] This study extracts directly quantifiable business features from standardized government hotline business datasets, covering core business dimensions such as request privacy, urgency, completion rate, process complexity, agent proficiency, and response efficiency, forming direct business feature vectors. Based on the massive unlabeled hotline corpus within the standardized government hotline business dataset, a new lexicon for the hotline domain is constructed using a nested balanced binary tree new word discovery algorithm. Simultaneously, domain-adaptive pre-training of the basic language model is performed, enabling it to learn policy expressions, habitual expressions, and implicit sentiment patterns in government hotline scenarios, thus addressing the pain point of insufficient domain adaptability in general models.

[0023] Based on the domain-adapted basic language model and the new lexicon for the hotline domain, an improved emotional extraction method is used to extract the sentiment of requests from the standardized government hotline business dataset. Through a multi-granularity attention fusion mechanism, it accurately identifies complex expressions such as veiled dissatisfaction, rhetorical questions and denials, and compound emotions in the government context. Finally, it outputs a sentiment feature vector of requests that includes sentiment category, sentiment probability, and sentiment intensity.

[0024] Based on a new thesaurus for the hotline domain, feature quantification and syntactic structure analysis of appeal texts were performed, extracting two main categories of text features: statistical and linguistic structural features, forming appeal text feature vectors. Then, the direct business feature vectors, appeal sentiment feature vectors, and appeal text feature vectors were standardized, aligned, and fused to construct a unified satisfaction prediction feature vector. This achieves unified modeling of multi-dimensional heterogeneous features, addressing the problem of insufficient prediction capability of single features.

[0025] The standardized satisfaction prediction feature vector is input into the pre-trained gradient boosting decision tree model to calculate and output the satisfaction probability. The government hotline work orders are classified into high, medium and low risk levels, and corresponding differentiated follow-up strategies of manual follow-up, semi-automatic review and automatic follow-up are matched. At the same time, based on the feature contribution of the gradient boosting decision tree model, key influencing factors are traced and interpretable prediction results are output, realizing a deep linkage between technical prediction and government hotline business handling.

[0026] In the specific implementation process, the collection of government hotline business data and the execution of standardized preprocessing to output a standardized government hotline business dataset specifically include: Collect the business data of the government hotline; The collected government hotline business data is standardized and its integrity is verified. Data with missing values ​​exceeding the preset missing value threshold and logical conflicts are removed, and the preliminary government hotline business dataset is output. Data anonymization processing is performed on the initially screened government hotline business dataset, and privacy fields are encrypted and anonymized to output the anonymized government hotline business dataset. The de-identified government hotline business dataset is subjected to deduplication and unified encoding processing to remove duplicate work order data. All fields are subjected to unified format encoding and numerical normalization to output a standardized government hotline business dataset. Perform text preprocessing on the request text field in the standardized government hotline business dataset, including sentence segmentation, word segmentation, character normalization, typo correction and stop word filtering, process any request text into an ordered word sequence, and output the text-standardized government hotline business dataset. The standardized government hotline business dataset is mapped and categorized for storage. It is divided into subsets of basic information of requests, subsets of call record information, subsets of follow-up record information, subsets of process record information, subsets of return visit record information, and subsets of basic citizen information. The standardized government hotline business dataset is then output.

[0027] Collect comprehensive business data from government hotlines, including basic information on requests, call records, follow-up records, process records, return visit records, and basic citizen information, thereby achieving complete data aggregation across the entire business chain.

[0028] We conduct field standardization and integrity verification on the collected government hotline business data from all dimensions. By setting a threshold for the percentage of missing data, we remove invalid samples with excessive missing values ​​or conflicting business logic, thus completing the initial data quality management and outputting the effective business dataset after initial screening.

[0029] Encryption and anonymization are performed on citizen privacy fields and sensitive business fields in the valid business dataset after initial screening. The government data security management regulations are strictly followed to avoid compliance risks in data use and output an anonymized compliant business dataset.

[0030] The de-identified compliant business dataset is subjected to deduplication and unified encoding processing to remove redundant data from duplicate work orders. At the same time, all fields are subjected to unified format encoding and numerical normalization processing to solve the problems of inconsistent formats and numerical units of multi-source heterogeneous data, and output a standardized business dataset.

[0031] Text preprocessing is performed on the request text field in the format-standardized dataset, which sequentially completes sentence segmentation, word segmentation, character normalization, typo correction and stop word filtering, transforming unstructured natural language text into structured ordered word sequences, providing standardized text input for subsequent sentiment extraction and text feature quantification, and outputting a standardized business dataset.

[0032] The standardized dataset is then mapped and categorized for storage, divided into six subsets of business information according to business dimensions, and the final standardized government hotline business dataset is output.

[0033] The construction of a new lexicon for the hotline domain based on the unlabeled hotline corpus in the standardized government hotline business dataset specifically includes: Extract the unlabeled corpus of hotline data from the standardized government hotline business dataset, perform co-occurrence statistics on candidate words w in the unlabeled corpus, and calculate the co-occurrence of two adjacent candidate words. mutual information That is, two adjacent candidate words The probability of co-occurrence in the unlabeled corpus of the hotline, and based on two adjacent candidate words. mutual information The average mutual information PMI(w) between the candidate word w and the domain sentiment trigger word is calculated using the following formula: In the formula, They are respectively The probability of appearing in the unlabeled corpus of the hotline; Calculate the left information entropy of the candidate word w. The corresponding calculation formula is as follows: In the formula, linL(w) is the set of all left-side adjacent characters of candidate word w; P(l|w) is the conditional probability that the left-side adjacent character l appears to the left of candidate word w; Calculate the right information entropy of the candidate word w. The corresponding calculation formula is as follows: In the formula, rinR(w) is the set of all right-side adjacent characters of candidate word w; P(r|w) is the conditional probability that the right-side adjacent character r appears to the right of candidate word w. The inverse document frequency (IDF) (w) of the candidate word w is calculated using the following formula: In the formula, N is the total number of documents in the unannotated corpus of the hotline; df(w) is the number of documents in the unannotated corpus of the hotline that contain the candidate word w; Calculate the comprehensive score of the candidate word w. The corresponding calculation formula is as follows: In the formula, Left information entropy of candidate word w With right information entropy The minimum value in; The overall score Candidate words w that exceed a preset scoring threshold are added as new words in the hotline field new word library.

[0034] A complete, unlabeled corpus of hotline language was extracted from a standardized government hotline business dataset. This corpus comprehensively covers colloquial expressions, policy terminology, business handling phrases, localized expressions, and emotion-related compound phrases in government service scenarios. Based on this, a co-occurrence statistical analysis of candidate words was conducted. By statistically analyzing the co-occurrence patterns of candidate words in the corpus, the mutual information of adjacent candidate words was calculated to measure the semantic cohesion between words. Simultaneously, based on the statistical results, the average mutual information between candidate words and domain-specific emotion trigger words was calculated to quantify the correlation strength between candidate words and the emotional expression of government service requests, and to initially screen out candidate phrases that are highly correlated with the emotional semantics of government service scenarios.

[0035] By calculating the left and right information entropy of candidate words, the stability of candidate word boundaries can be accurately determined. The left information entropy measures the distribution richness of the characters adjacent to the left of the candidate word, and the right information entropy measures the distribution richness of the characters adjacent to the right of the candidate word. The higher the values ​​of the two indicators, the more stable the left and right word boundaries of the candidate word are, and the greater its possibility of forming an independent word. This solves the core problem of the segmentation error of out-of-vocabulary words in the field of government hotlines by general word segmentation models.

[0036] The inverse document frequency of candidate words is calculated to measure their domain specificity in the hotline corpus. This allows for the selection of distinctive words that frequently appear in government hotline scenarios and possess specific meanings specific to those scenarios, while eliminating common words used in general scenarios to ensure the domain adaptability of the new thesaurus.

[0037] By combining three core indicators—mutual information, minimum left and right information entropy, and inverse document frequency—a comprehensive score is calculated for candidate words. A preset scoring threshold is then used to complete the final selection of candidate words. Candidate words that meet the comprehensive score criteria are identified as new hotline terms and included in the new hotline-related thesaurus. The final new hotline-related thesaurus covers core categories such as phrases related to government affairs, policy noun phrases, emotional reinforcement phrases, and negation / contrast phrases, effectively improving the accuracy of sentiment recognition and text analysis in government contexts.

[0038] The above approach can address the issue of insufficient accuracy in recognizing specialized terminology, new policy terms, business compound phrases, and emotion-triggered words specific to the government hotline domain by general natural language processing models. It can also construct a specialized domain-specific new lexicon adapted to government demand scenarios based on standardized government hotline business data, providing prior domain knowledge support for subsequent sentiment extraction and text feature quantification.

[0039] The improved sentiment extraction of requests from the standardized government hotline business dataset is performed on the new lexicon and the domain-adaptive pre-trained basic language model based on the hotline domain, and the sentiment feature vector of the requests is output. Specifically, this includes: Extract words from the request text in the standardized government hotline business dataset. Corresponding word vectors ; Based on the new lexicon in the hotline field, calculate words Corresponding domain dictionary feature vector The corresponding calculation formula is as follows: In the formula, For words Word frequencies in the unlabeled corpus of the hotline; Words contained in the unannotated corpus of the hotline The number of documents; beta is the preset adjustment coefficient; For words Average mutual information with domain-specific emotion trigger words; Extract words from the stated appeal text The corresponding punctuation, tone, adverbs, negation words, and intensity words are used to output words. tone auxiliary feature vector ; Extract words from the stated appeal text The corresponding word position information and syntactic auxiliary features are used to output the word. Positional Syntactic Feature Vector ; The word vector The domain dictionary feature vector The tone auxiliary feature vector The positional syntax feature vector To combine and construct words Corresponding emotional appeal extraction input vector .

[0040] From the complaint text in the standardized government hotline business dataset, the basic word vector corresponding to each word segment is extracted. This vector is generated by a basic language model that is pre-trained with domain-adaptive technology, which fully preserves the basic semantic information of words in the context of the government hotline. This achieves the initial conversion of unstructured text into structured numerical vectors, providing underlying semantic support for sentiment extraction.

[0041] Based on a pre-built domain-specific lexicon for hotline communication, a domain-specific lexicon feature vector is calculated for each word. This feature vector integrates word frequency and domain distribution features in the hotline corpus, as well as the strength of association with domain-specific emotion trigger words. By balancing the weights of each dimension through a preset adjustment coefficient, the feature weights of new words in the government hotline domain, policy terms, negative emotion trigger words, and emotion-enhancing phrases are particularly strengthened. This addresses the problem of insufficient attention paid to domain-specific emotional vocabulary by general models and fully leverages the enhanced role of the domain-specific lexicon in emotion recognition.

[0042] To address the complex implicit emotional expressions commonly found in government requests, such as euphemisms of dissatisfaction, rhetorical questions, layered complaints, and negations, this study extracts features from each word, including punctuation, tone, adverbs, negation words, and intensity words, forming a tone-assisted feature vector. This vector accurately captures key emotional modifiers in the text, such as tone reinforcement, semantic reversal, and emotional progression, compensating for the shortcomings of general models in recognizing implicit emotions in government contexts.

[0043] By extracting the word position information and syntactic auxiliary features corresponding to each word, a positional syntactic feature vector is formed. This fully preserves the word order information, syntactic structure and grammatical role information of words in the sentence, providing structural support for the model to capture long-distance semantic dependencies and inter-sentence sentiment logic.

[0044] By concatenating the aforementioned basic word vectors, domain dictionary feature vectors, tone-assisted feature vectors, and positional syntax feature vectors, a standardized input vector for extracting the sentiment of each word in the appeal text is constructed. This input vector simultaneously integrates basic semantics, domain-specific information, sentiment modification information, and syntactic structure information, comprehensively covering the sentiment expression dimensions of government appeal texts.

[0045] The above approach addresses the shortcomings of general sentiment classifiers, such as insufficient ability to capture features specific to government affairs contexts and low weights for novel domain words and sentiment trigger words. It provides standardized input vectors rich in domain semantics and sentiment information for subsequent context encoding and multi-granularity attention fusion, which helps improve the accuracy of sentiment recognition for government affairs appeals.

[0046] Based on the domain-adaptive pre-trained base language model, the input vector for extracting the stated sentiment is obtained. Perform context encoding and output words. Corresponding emotional context encoding vector The corresponding calculation formula is as follows: In the formula, BiGRU() is a bidirectional gated loop unit; For words The corresponding emotional context encoding vector; Calculation words Corresponding attention scoring vector The corresponding calculation formula is as follows: In the formula, tanh() is the hyperbolic tangent activation function; is the learnable attention scoring weight matrix; b is the attention scoring bias term; Calculation words Corresponding attention weights The corresponding calculation formula is as follows: In the formula, v is a learnable attention weight vector; For words The corresponding attention scoring vector; exp() is an exponential function with the natural constant e as the base; The formula for calculating the single-view global sentiment representation z is as follows: ; The single-view global sentiment representations corresponding to the title, body, and call summary of the appeal text are weighted and fused to output a multi-view fused global sentiment representation. The corresponding calculation formula is as follows: In the formula, These are single-perspective global sentiment representations corresponding to the title, body, and call summary of the appeal text, respectively. The fusion weight coefficients are learnable, and .

[0047] Based on a pre-trained adaptive language model for the government hotline domain, a bidirectional gated recurrent unit is used to perform contextual encoding on the previously constructed emotional appeal extraction input vector. This encoding process can simultaneously capture the long-distance semantic dependencies of the text in both the forward and backward directions, deeply mining the contextual semantic relationships and emotional logic between words in long government hotline texts and complex appeal sentences. This addresses the pain point of general encoding models' insufficient understanding of the semantics of government-specific contexts, and ultimately outputs the emotional appeal contextual encoding vector corresponding to each word, fully preserving the contextual semantics and emotional association information of words in specific appeal scenarios.

[0048] First, based on the context encoding vector and the features of the hotline domain dictionary, the attention scoring vector corresponding to each word is calculated. A non-linear activation function is then used to perform feature transformation, achieving deep fusion of semantic features and domain-specific features. Next, based on the attention scoring vector, the attention weight corresponding to each word is calculated, completing the differentiated weight allocation for different words in the text. This process explicitly strengthens the weights of core emotion-related words such as new words in the hotline domain, policy entities, negative emotion trigger words, and negation / contrast phrases, while weakening the interference of meaningless function words on emotion judgment, accurately locating the core emotional expression segments in the appeal text.

[0049] Based on the context encoding vector of each word and the corresponding attention weight, the aggregation of word-level features to text-level features is completed by weighted summation, generating a single-view global sentiment representation. This integrates discrete word-level semantic encodings into a unified text global sentiment semantic vector, fully preserving the overall sentiment tendency information of the appeal text.

[0050] For the three core text modules of a government hotline work order—the request title, the request body, and the call summary—the aforementioned encoding and feature aggregation processes are completed separately to obtain single-view global sentiment representations for each module. Then, through learnable fusion weight coefficients, the multi-view sentiment representations are weighted and fused to output the final multi-view fused global sentiment representation. This process fully leverages the complementarity of sentiment information from different text modules, comprehensively capturing the full-dimensional sentiment expression of the work order, and further improving the completeness and accuracy of sentiment semantic representation.

[0051] The above approach provides a key technical module that addresses the challenge of recognizing complex emotional expressions such as euphemisms, rhetorical questions, and repeated reminders in government contexts. Through deep contextual semantic encoding and multi-dimensional attention mechanism optimization, it achieves accurate representation of the emotional semantics of the appeal text, providing a high-quality global semantic foundation for subsequent emotion classification and intensity calculation.

[0052] The multi-view fusion global sentiment representation is achieved through a fully connected layer. Perform feature mapping and output the sentiment fully connected feature vector; The sentiment fully connected feature vector is normalized, and the sentiment category probability vector C corresponding to the appeal text is calculated using the following formula: In the formula, The output layer weight matrix is ​​a learnable matrix. The output layer bias term; softmax() is the normalized exponential function; Based on the sentiment category probability vector C, the positive sentiment probability is split. Neutral emotional probability With the probability of negative emotions And calculate the emotional intensity score corresponding to the appeal text. The corresponding calculation formula is as follows: ; The emotion category probability vector C and the emotion intensity score are used as the basis for the expression. The data is integrated to construct the emotional feature vector corresponding to the stated appeal text.

[0053] By performing high-dimensional feature mapping on the multi-view fusion global sentiment representation through a fully connected layer, the global sentiment semantic vector, which integrates multi-dimensional information such as the appeal title, body text, and call summary, is transformed into a sentiment fully connected feature vector adapted to the sentiment classification task. This completes the deep purification and dimensional adaptation of sentiment semantic features, further strengthens the effective information in the features that is strongly correlated with sentiment classification, and weakens the interference of redundant noise on subsequent sentiment judgment.

[0054] Normalization is performed on the fully connected feature vector of sentiment, and the mapping from the feature space to the probability space is completed through the normalization exponential function to calculate and generate the sentiment category probability vector corresponding to the appeal text. This vector fully quantifies the probability of the appeal text belonging to three preset sentiment categories: positive, neutral, and negative. It realizes a multi-class quantitative representation of the sentiment tendency of the appeal text, and can accurately capture complex compound emotional tendencies such as subtle dissatisfaction, superimposed complaints, and neutral appeals in the context of government affairs. It makes up for the deficiency of general classifiers in recognizing the accuracy of sentiment expressions specific to government affairs.

[0055] Based on the probability vector of sentiment categories, the probability of positive sentiment and the probability of negative sentiment are extracted and used to calculate a continuous sentiment intensity score corresponding to the appeal text. This score enables a refined quantification of the sentiment polarity and intensity of the appeal text. Compared with traditional discrete sentiment tags, it can more accurately reflect different levels of emotional expression in government appeals, such as mild complaints, strong grievances, and implicit dissatisfaction, providing more discriminative and fine-grained sentiment features to support satisfaction prediction.

[0056] Finally, the probability vector of sentiment categories and the sentiment intensity score are integrated to construct a standardized sentiment feature vector. This vector covers two core dimensions: sentiment category distribution and sentiment intensity. It can be directly used as input for the multi-source feature fusion process, working together with business features and text features to construct the satisfaction prediction feature vector, thereby improving the accuracy and stability of government hotline satisfaction prediction.

[0057] The above method can realize the transformation of emotional semantic representation into quantifiable emotional features, providing fine-grained and highly discriminative emotional dimension input for subsequent government hotline satisfaction prediction models. This solves the problems of traditional solutions that only output binary emotional labels, have insufficient information granularity, and cannot adapt to the complex emotion recognition in government contexts.

[0058] The joint training and iterative optimization process of the improved appeal sentiment extraction algorithm is as follows: Multiple teacher models are trained based on manually labeled sentiment samples. The unlabeled hotline data is then input into the multiple teacher models, and the predicted probability distribution of the unlabeled hotline samples is output. Based on the predicted probability distribution, the prediction confidence of the unlabeled samples of the hotline is calculated using the following formula: In the formula, The prediction confidence level for the i-th unlabeled hotline sample; Predicted probability distribution for the i-th unlabeled hotline sample Information entropy; K is the total number of sentiment categories; Select unlabeled hotline samples whose prediction confidence is greater than or equal to a preset confidence threshold and whose prediction results from multiple teacher models are consistent. Generate pseudo-labels for the selected unlabeled hotline samples and output a training sample set with pseudo-labels. The formula for calculating the pseudo-labels is as follows: In the formula, is the pseudo-label for the i-th unlabeled hotline sample; argmax() is the function that takes the category corresponding to the maximum probability; tau is the preset confidence threshold; This is an indicator that the prediction results of multiple teacher models are consistent; it takes a value of 1 when they are consistent and a value of 0 when they are inconsistent. The manually labeled sentiment samples are merged with the set of training samples with pseudo-labels to construct a joint training sample set. A weighted focus loss function is constructed based on the joint training sample set, and the training loss calculation result of the improved appeal sentiment extraction algorithm is output. The learnable parameters of the improved appeal sentiment extraction algorithm are optimized through the backpropagation algorithm to complete the joint training and iterative optimization of the improved appeal sentiment extraction algorithm.

[0059] Based on manually labeled sentiment samples from government hotline scenarios, supervised training of multiple heterogeneous teacher models was completed, forming a multi-model integrated domain sentiment prior knowledge system. Then, the massive amount of unlabeled corpus from government hotlines was input into the trained teacher models to obtain the predicted probability distribution of each unlabeled sample. This provides a multi-perspective and highly reliable prediction basis for subsequent pseudo-label generation, fully exploring the value of the extremely high proportion of unlabeled corpus in government hotline scenarios and overcoming the industry bottleneck of poor model fitting caused by insufficient labeled samples.

[0060] Based on the predicted probability distribution output by the multi-teacher model, the prediction confidence of each unlabeled sample is calculated. This metric quantifies the certainty of the model's prediction results through the information entropy of the predicted probability distribution; the lower the information entropy, the more stable the model's judgment of the sample's sentiment category and the higher the confidence. Subsequently, sample screening is completed through a pre-set confidence threshold, retaining only unlabeled samples with qualified confidence and completely consistent prediction results from multiple teacher models. Pseudo-labels are generated for these samples, and a training sample set with pseudo-labels is constructed. This dual screening mechanism strictly ensures the labeling quality of pseudo-labels, avoiding noise interference from low-quality pseudo-labels to the model training.

[0061] Manually labeled gold-labeled samples are merged with screened pseudo-labeled samples to construct a semi-supervised joint training sample set. Addressing the issues of low proportion of negative sentiment samples and uneven category distribution in government hotline corpora, a weighted focus loss function is constructed. This function can differentiate the loss weights of easily classified and difficult-to-classify samples, as well as samples of different sentiment categories. Simultaneously, it matches pseudo-labeled samples with weights positively correlated with confidence levels, weakening the dominant role of easily classified samples in the training process and strengthening the optimization contribution of difficult and rare category samples. Finally, the backpropagation algorithm iteratively optimizes all learnable parameters of the model, completing the joint training and iterative optimization of the model. This achieves deep adaptation of the model to the government hotline domain, significantly improving the accuracy, robustness, and generalization ability of sentiment extraction.

[0062] The above methods can solve the problems of scarce manually labeled samples, unbalanced distribution of sentiment categories, and insufficient generalization ability of general models in the context of government hotlines. By jointly optimizing the semi-supervised learning of confidence pseudo-labels and weighted focus loss, the accuracy and generalization performance of the model in the field of government hotlines can be greatly improved without significantly increasing the cost of manual labeling.

[0063] Based on the new thesaurus in the hotline domain, the process involves performing text feature quantification and syntactic structure analysis on the appeal text to output an appeal text feature vector. Furthermore, it involves multi-source feature fusion of the direct business feature vector, the appeal sentiment feature vector, and the appeal text feature vector to output a satisfaction prediction feature vector. Specifically, this includes: Based on the new thesaurus in the hotline field, text feature quantification is performed on the appeal text to extract a statistical text feature vector; wherein, the statistical text feature vector includes the basic word count, part-of-speech count, average sentence length, stop word ratio, and linking word frequency; Based on the new lexicon in the hotline field, syntactic structure analysis is performed on the appeal text to extract the language structural text feature vector of the appeal text; wherein, the language structural text feature vector includes minimum edit distance, dependency syntax ratio, inter-sentence similarity, and narrative integrity index; The statistical text feature vector is concatenated with the language structural text feature vector to output the appeal text feature vector; Obtain the direct business feature vector and the appeal sentiment feature vector, and perform feature alignment and standardization processing with the appeal text feature vector to output the aligned direct business feature vector, appeal sentiment feature vector and appeal text feature vector. The aligned direct business feature vector, appeal sentiment feature vector, and appeal text feature vector are concatenated and fused to construct the satisfaction prediction feature vector. The corresponding calculation formula is as follows: In the formula, Let be the feature vector for predicting the satisfaction level of the j-th government hotline work order; Let be the direct business feature vector of the j-th government hotline work order; The emotional intensity score of the request text of the j-th government hotline work order; Let be the probability vector of the sentiment category of the request text of the j-th government hotline work order; Let be the feature vector of the request text of the j-th government hotline work order.

[0064] Based on a pre-built domain-specific new thesaurus, the text of the appeal is subjected to full-dimensional text feature quantification, and statistical text feature vectors are extracted. These features characterize the language expression pattern of the appeal from a surface statistical dimension, covering core indicators such as basic word count, part-of-speech count, average sentence length, stop word ratio, and linking word frequency. They can accurately reflect the content length, expression complexity, narrative logic, and language habits of the appeal. At the same time, the text segmentation accuracy is optimized by relying on the domain-specific new thesaurus, avoiding statistical feature bias caused by errors in domain-specific new word segmentation, and providing a text surface dimension discrimination basis for satisfaction prediction.

[0065] Based on a new thesaurus for hotline-related issues, deep syntactic structure analysis is conducted on the texts of public appeals to extract structural textual feature vectors. These features, from the perspectives of deep syntax and semantic structure, uncover textual structural features highly correlated with public satisfaction. They encompass minimum edit distance, dependency syntax ratios, sentence similarity, and narrative integrity indicators, effectively identifying the repetition, focus, logical completeness, and standardization of expression in appeals. This accurately captures latent features indicating high dissatisfaction risks in government appeals, such as repeated follow-ups, vague requests, and logical inconsistencies, thus compensating for the limitations of statistical textual features in capturing deep semantic structures.

[0066] By concatenating statistical text feature vectors with linguistic structural text feature vectors to form a complete appeal text feature vector, the appeal text achieves full-dimensional feature representation from surface statistical patterns to deep syntactic structures, comprehensively covering text dimensional information related to public satisfaction.

[0067] The system acquires direct business feature vectors and emotional appeal feature vectors, and performs feature dimension alignment and numerical standardization processing with the appeal text feature vectors. This solves the problems of inconsistent dimensions and mismatched dimensions among the three types of heterogeneous features, ensuring the fusion of multi-source features and the standardization of model input.

[0068] The aligned three types of features are concatenated and fused to form a standardized feature vector for predicting satisfaction. This vector integrates three core dimensions: the business attributes of the entire work order process, the deep emotional tendency of the appeal text, and the text expression and structural features. This achieves unified modeling of the heterogeneous features of government hotline work orders and fully explores the synergistic relationship between business processes, emotional expression, text features, and public satisfaction.

[0069] The above approach can address the issues of insufficient single-feature modeling capabilities and inadequate collaborative utilization of multi-source heterogeneous features in existing solutions, providing standardized and highly discriminative input features for subsequent gradient-boosting decision tree prediction models, thereby improving the accuracy of satisfaction prediction.

[0070] The process of inputting the satisfaction prediction feature vector into a pre-trained gradient boosting decision tree model and outputting the satisfaction prediction result, risk level, recommended follow-up method, and explanation of key influencing factors for satisfaction prediction for government hotline work orders specifically includes: The satisfaction prediction feature vector is input into the pre-trained gradient boosting decision tree model. Multiple weak learners of the gradient boosting decision tree model are used to predict and calculate the satisfaction prediction feature vector, and the prediction result of each weak learner is output. The weighted summation of the prediction results from each weak learner yields the ensemble output of the gradient boosting decision tree model, calculated using the following formula: In the formula, To improve the gradient boosting decision tree for predicting the satisfaction level of the i-th government hotline work order, we need to define the feature vector. The integrated output value; M is the total number of weak learners; Let m be the learning rate of the m-th weak learner; Let m be the feature vector of the satisfaction prediction by the m-th weak learner. The predicted output value; The satisfaction probability of the government hotline work order is calculated by mapping the integrated output value through an activation function, and the corresponding calculation formula is as follows: In the formula, Let be the satisfaction probability of the i-th government hotline work order; sigmoid() is the sigmoid activation function; A preset probability threshold is obtained, including a first probability threshold and a second probability threshold. Based on the preset probability threshold, the government hotline work order is classified into risk levels: when the satisfaction probability is less than the first probability threshold, the government hotline work order is classified into a high-risk unsatisfactory level; when the satisfaction probability is greater than or equal to the first probability threshold and less than the second probability threshold, the government hotline work order is classified into a medium-risk pending verification level; when the satisfaction probability is greater than or equal to the second probability threshold, the government hotline work order is classified into a low-risk satisfaction level. Based on the risk level of the government hotline work order, a corresponding preset follow-up strategy rule is matched and a recommended follow-up method is generated: a manual follow-up method is recommended for government hotline work orders with a high-risk dissatisfaction level; a manual review method or a semi-automatic follow-up method is recommended for government hotline work orders with a medium-risk pending verification level; and an automatic follow-up method is recommended for government hotline work orders with a low-risk satisfaction level. Based on the feature contribution value calculation logic of the gradient boosting decision tree model, the contribution of each dimension of the feature in the satisfaction prediction feature vector is sorted, and the core influencing feature set that has the greatest impact on the satisfaction prediction result is selected. Combined with the negative emotional segments and business process abnormal indicators in the appeal text, the explanation of the key influencing factors of the satisfaction prediction is generated. Based on the explanation of the key influencing factors for the satisfaction prediction, the recommended follow-up method is optimized and adjusted, and key reminders for the follow-up are added.

[0071] This invention preferentially selects gradient boosting decision trees as the core model for satisfaction prediction. The core reason is that this model can perfectly adapt to the modeling needs of mixed and heterogeneous features in the context of government hotlines. Compared with deep learning black-box models, it has stronger adaptability to the mixture of discrete business features, continuous sentiment features, and structured text features, better resistance to overfitting and generalization ability, and has a natural advantage in interpretability, fully complying with the compliance requirements of government data governance. The model completes supervised training and cross-validation optimization based on the labeled follow-up results of historical work orders from government hotlines. It consists of multiple serially iterative classification and regression tree weak learners. Each weak learner optimizes the fit based on the residual of the previous learning round, which can accurately capture the nonlinear correlation between multi-source features and public satisfaction.

[0072] In the prediction execution phase, the constructed satisfaction prediction feature vector, which integrates the full-process business characteristics of work orders, deep emotional characteristics of demands, and textual statistical and structural characteristics, is first input into the pre-trained gradient boosting decision tree model. Each weak learner in the model performs independent prediction calculations on the input feature vector and outputs the single-model prediction result. During this process, different weak learners can capture the potential correlation between different dimensions of business processes, emotional tendencies, and textual expressions and satisfaction, realizing in-depth mining and utilization of multi-source fusion features. Finally, based on the preset learning rate weights, the prediction results of all weak learners are weighted and summed to obtain the final integrated output value of the model. The learning rate can balance the model's fitting ability and generalization ability, avoid model overfitting, and ensure the stability of the prediction results in different batches of government hotline data.

[0073] After obtaining the model ensemble output, a sigmoid activation function is used to perform a nonlinear mapping, converting the continuous ensemble output into a standardized satisfaction probability within the range of 0 to 1. This probability value accurately quantifies the likelihood of the public giving a satisfactory evaluation for the corresponding government hotline work order; a higher value indicates a higher probability of public satisfaction, and vice versa. Compared to existing solutions that only output binary classification results of satisfaction / dissatisfaction, this invention outputs a continuous satisfaction probability, achieving a refined quantification of the risk of dissatisfaction with work orders, rather than a simple binary division. This provides a quantitative basis for the subsequent accurate classification of risk levels and perfectly adapts to the complex risk gradient distribution characteristics of government hotline work orders.

[0074] A preset dual-threshold probability system is established, comprising a first probability threshold and a second probability threshold. Both thresholds are determined based on the distribution characteristics of historical work order data and the allocation of government hotline follow-up resources, through cross-validation and business adaptation optimization, balancing the accuracy of risk identification with the feasibility of business handling. Based on this dual-threshold system, work orders are classified into three risk levels: when the satisfaction probability is below the first probability threshold, it is classified as a high-risk dissatisfaction level; these work orders have a very high probability of receiving dissatisfaction feedback and are the core focus of follow-up handling and risk prevention. When the satisfaction probability is between the two thresholds, it is classified as a medium-risk pending verification level; these work orders have potential dissatisfaction risks, and the handling of requests and public sentiment are uncertain, requiring focused verification. When the satisfaction probability is above the second probability threshold, it is classified as a low-risk satisfaction level; these work orders have a very high probability of public satisfaction and a very low risk of dissatisfaction.

[0075] This three-tiered classification mechanism enables automated, standardized, and precise identification of dissatisfaction risks, solving the industry pain points of high subjectivity, low efficiency, and high false negative rate caused by the reliance on human experience in the existing model. At the same time, it provides a precise basis for the formulation of differentiated follow-up strategies and achieves the optimal allocation of government hotline follow-up resources.

[0076] Based on the risk level of government hotline work orders, corresponding preset follow-up strategy rules are matched and recommended follow-up methods are generated. For high-risk, unsatisfactory work orders, manual follow-up is prioritized. These work orders often have core issues such as unresolved requests, strong negative emotions, and abnormal business processes, requiring in-depth human intervention to communicate, verify requests, and promote problem-solving, thus avoiding unsatisfactory evaluations at the source. For medium-risk, pending verification work orders, manual review or semi-automatic follow-up is recommended. A lightweight verification process confirms the handling of requests and the public's true attitude, identifying potential risks while controlling labor costs. For low-risk, satisfactory work orders, automatic follow-up is recommended. Satisfaction confirmation is completed through standardized intelligent voice and SMS, significantly reducing follow-up labor costs and improving the efficiency of follow-up for all work orders. This achieves direct linkage between prediction results and follow-up strategies, solving the core problem of the disconnect between technical prediction and business processing in existing solutions, and has strong business implementation value.

[0077] To meet the core requirements of algorithm interpretability and compliance in government scenarios, this design simultaneously outputs interpretability analysis and key influencing factors for prediction results. Based on the feature contribution value calculation logic of a gradient boosting decision tree model, the design quantifies and ranks the contribution of all-dimensional input features, calculates the positive or negative impact of each feature on the current prediction result, and selects a set of core influencing features, including key indicators such as request completion rate, request process complexity, request response efficiency, emotional intensity, and text complexity. Simultaneously, it combines negative emotional fragments in the request text and abnormal business process indicators to generate natural language explanations of key influencing factors that can directly guide business operations, clearly identifying the core reasons leading to high dissatisfaction risk in work orders. This design makes prediction results traceable and interpretable, meets the compliance requirements of government data governance, and provides clear direction for responsible departments to optimize service processes and address public demands in a targeted manner.

[0078] Finally, based on the explanation of key influencing factors, the initially matched follow-up strategy was optimized and adjusted a second time, and targeted follow-up key reminders were added. This ensured that the recommended follow-up strategy not only had a tiered handling framework, but also had implementable guidelines, further improving the accuracy and effectiveness of follow-up handling.

[0079] The above approach can solve the problems of existing technologies, such as the disconnect between prediction results and business scenarios, insufficient interpretability of black-box predictions, and inability to directly support the optimization of follow-up strategies and the improvement of service quality. It can realize the full-process transformation from standardized multi-source fusion feature vectors to business outputs that can be implemented, while taking into account prediction accuracy, risk classification accuracy, result interpretability, and adaptability to government business.

[0080] like Figure 2The diagram shown is a system block diagram of a government hotline satisfaction prediction system based on multi-source feature vector fusion provided in an embodiment of the present invention. The system includes: The data acquisition and preprocessing module is used to collect government hotline business data and perform standardized preprocessing to output a standardized government hotline business dataset. The new lexicon construction and model pre-training module is used to extract direct business feature vectors from the standardized government hotline business dataset, and at the same time, based on the unlabeled hotline corpus in the standardized government hotline business dataset, construct a new lexicon for the hotline domain and perform domain-adaptive pre-training on the basic language model. The emotional appeal extraction module is used to perform improved emotional appeal extraction on the appeal text in the standardized government hotline business dataset based on the new lexicon of the hotline domain and the basic language model after domain adaptive pre-training, and output the emotional appeal feature vector. The appeal text output and feature fusion module is used to perform text feature quantification and syntactic structure analysis on the appeal text based on the new thesaurus in the hotline domain, output the appeal text feature vector, and perform multi-source feature fusion on the direct business feature vector, the appeal sentiment feature vector and the appeal text feature vector to output the satisfaction prediction feature vector. The satisfaction prediction result output module is used to input the satisfaction prediction feature vector into the pre-trained gradient boosting decision tree model and output the satisfaction prediction result, risk level, recommended follow-up method, and explanation of key influencing factors of satisfaction prediction for the government hotline work order.

[0081] Figure 2 The apparatus of the illustrated embodiment can be used to perform corresponding actions. Figure 1 The steps in the method embodiments shown are implemented in a similar manner and have similar technical effects, and will not be repeated here.

[0082] An electronic device includes a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor performs the steps of the government hotline satisfaction prediction method based on multi-source feature vector fusion as described above.

[0083] like Figure 3 The diagram shown is a hardware structure schematic of an electronic device according to an embodiment of the present invention. The electronic device 30 includes: a processor 31, a memory 32, and a computer program; wherein... The memory 32 is used to store the computer program, and the memory may also be flash memory. The computer program is, for example, an application program or functional module that implements the above method.

[0084] Processor 31 is configured to execute the computer program stored in the memory to implement the various steps performed by the device in the above method. For details, please refer to the relevant descriptions in the preceding method embodiments.

[0085] Alternatively, the memory 32 can be either standalone or integrated with the processor 31.

[0086] When the memory 32 is a device independent of the processor 31, the device may further include: Bus 33 is used to connect the memory 32 and the processor 31.

[0087] A readable storage medium storing a computer program, which, when executed by a processor, is used to implement the steps of the government hotline satisfaction prediction method based on multi-source feature vector fusion as described above.

[0088] The readable storage medium can be a computer storage medium or a communication medium. A communication medium includes any medium that facilitates the transfer of computer programs from one location to another. A computer storage medium can be any available medium accessible to a general-purpose or special-purpose computer. For example, a readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application-Specific Integrated Circuit (ASIC). Alternatively, the ASIC can be located in a user equipment. Of course, the processor and the readable storage medium can also exist as discrete components in a communication device. The readable storage medium can be a read-only memory (ROM), random access memory (RAM), CD-ROM, magnetic tape, floppy disk, and optical data storage device, etc.

[0089] The present invention also provides a program product including executable instructions stored in a readable storage medium. At least one processor of the device can read the executable instructions from the readable storage medium, and the at least one processor executes the executable instructions to cause the device to implement the methods provided in the various embodiments described above.

[0090] In the embodiments of the above-described device, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly manifested as execution by a hardware processor, or execution by a combination of hardware and software modules within the processor.

[0091] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical feature vectors; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting government hotline satisfaction based on multi-source feature vector fusion, characterized in that, The method includes: Collect government hotline business data and perform standardized preprocessing to output a standardized government hotline business dataset. Direct business feature vectors are extracted from the standardized government hotline business dataset. At the same time, based on the unlabeled hotline corpus in the standardized government hotline business dataset, a new hotline domain lexicon is constructed and the basic language model is pre-trained in a domain-adaptive manner. Based on the new hotline domain lexicon and the domain-adaptive pre-trained basic language model, an improved sentiment extraction method is performed on the petition text in the standardized government hotline business dataset, and the sentiment feature vector of the petition is output. Based on the new thesaurus in the hotline field, text feature quantification and syntactic structure analysis are performed on the appeal text to output the appeal text feature vector. Multi-source feature fusion is then performed on the direct business feature vector, the appeal sentiment feature vector, and the appeal text feature vector to output the satisfaction prediction feature vector. The satisfaction prediction feature vector is input into the pre-trained gradient boosting decision tree model, which outputs the satisfaction prediction result, risk level, recommended follow-up method, and explanation of key influencing factors of satisfaction prediction for the government hotline work order.

2. The government hotline satisfaction prediction method based on multi-source feature vector fusion according to claim 1, characterized in that, The process of collecting government hotline business data and performing standardized preprocessing to output a standardized government hotline business dataset specifically includes: Collect the business data of the government hotline; The collected government hotline business data is standardized and its integrity is verified. Data with missing values ​​exceeding the preset missing value threshold and logical conflicts are removed, and the preliminary government hotline business dataset is output. Data anonymization processing is performed on the initially screened government hotline business dataset, and privacy fields are encrypted and anonymized to output the anonymized government hotline business dataset. The de-identified government hotline business dataset is subjected to deduplication and unified encoding processing to remove duplicate work order data. All fields are subjected to unified format encoding and numerical normalization to output a standardized government hotline business dataset. Perform text preprocessing on the request text field in the standardized government hotline business dataset, including sentence segmentation, word segmentation, character normalization, typo correction and stop word filtering, process any request text into an ordered word sequence, and output the text-standardized government hotline business dataset. The standardized government hotline business dataset is mapped and categorized for storage. It is divided into subsets of basic information of requests, subsets of call record information, subsets of follow-up record information, subsets of process record information, subsets of return visit record information, and subsets of basic citizen information. The standardized government hotline business dataset is then output.

3. The government hotline satisfaction prediction method based on multi-source feature vector fusion according to claim 1, characterized in that, The construction of a new lexicon for the hotline domain based on the unlabeled hotline corpus in the standardized government hotline business dataset specifically includes: Extract the unlabeled corpus of hotline data from the standardized government hotline business dataset, perform co-occurrence statistics on candidate words w in the unlabeled corpus, and calculate the co-occurrence of adjacent candidate words. mutual information That is, two adjacent candidate words The probability of co-occurrence in the unlabeled corpus of the hotline, and based on two adjacent candidate words. mutual information The average mutual information PMI(w) between the candidate word w and the domain sentiment trigger word is calculated using the following formula: In the formula, They are respectively The probability of appearing in the unlabeled corpus of the hotline; Calculate the left information entropy of the candidate word w. The corresponding calculation formula is as follows: In the formula, linL(w) is the set of all left-side adjacent characters of candidate word w; P(l|w) is the conditional probability that the left-side adjacent character l appears to the left of candidate word w; Calculate the right information entropy of the candidate word w. The corresponding calculation formula is as follows: In the formula, rinR(w) is the set of all right-side adjacent characters of candidate word w; P(r|w) is the conditional probability that the right-side adjacent character r appears to the right of candidate word w. The inverse document frequency (IDF) (w) of the candidate word w is calculated using the following formula: In the formula, N is the total number of documents in the unannotated corpus of the hotline; df(w) is the number of documents in the unannotated corpus of the hotline that contain the candidate word w; Calculate the comprehensive score of the candidate word w. The corresponding calculation formula is as follows: In the formula, Left information entropy of candidate word w With right information entropy The minimum value in; The comprehensive score Candidate words w that exceed a preset scoring threshold are added as new words in the hotline field new word library.

4. The method for predicting government hotline satisfaction based on multi-source feature vector fusion according to claim 1, characterized in that, The improved sentiment extraction of requests from the standardized government hotline business dataset is performed on the new lexicon and the domain-adaptive pre-trained basic language model based on the hotline domain, and the sentiment feature vector of the requests is output. Specifically, this includes: Extract words from the request text in the standardized government hotline business dataset. Corresponding word vectors ; Based on the new lexicon in the hotline field, calculate words Corresponding domain dictionary feature vector The corresponding calculation formula is as follows: In the formula, For words Word frequencies in the unlabeled corpus of the hotline; Words contained in the unannotated corpus of the hotline The number of documents; beta is the preset adjustment coefficient; For words Average mutual information with domain-specific emotion trigger words; Extract words from the stated appeal text The corresponding punctuation, tone, adverbs, negation words, and intensity words are used to output words. tone auxiliary feature vector ; Extract words from the stated appeal text The corresponding word position information and syntactic auxiliary features are used to output the word. Positional Syntactic Feature Vector ; The word vector The domain dictionary feature vector The tone auxiliary feature vector The positional syntax feature vector To combine and construct words Corresponding emotional appeal extraction input vector .

5. The government hotline satisfaction prediction method based on multi-source feature vector fusion according to claim 4, characterized in that, Based on the domain-adaptive pre-trained base language model, the input vector for extracting the stated sentiment is obtained. Perform context encoding and output words. Corresponding emotional context encoding vector The corresponding calculation formula is as follows: In the formula, BiGRU() is a bidirectional gated loop unit; For words The corresponding emotional context encoding vector; Calculation words Corresponding attention scoring vector The corresponding calculation formula is as follows: In the formula, tanh() is the hyperbolic tangent activation function; is the learnable attention scoring weight matrix; b is the attention scoring bias term; Calculation words Corresponding attention weights The corresponding calculation formula is as follows: In the formula, v is a learnable attention weight vector; For words The corresponding attention scoring vector; exp() is an exponential function with the natural constant e as the base; The formula for calculating the single-view global sentiment representation z is as follows: ； The single-view global sentiment representations corresponding to the title, body, and call summary of the appeal text are weighted and fused to output a multi-view fused global sentiment representation. The corresponding calculation formula is as follows: In the formula, These are single-perspective global sentiment representations corresponding to the title, body, and call summary of the appeal text, respectively. The fusion weight coefficients are learnable, and .

6. The government hotline satisfaction prediction method based on multi-source feature vector fusion according to claim 5, characterized in that, The multi-view fusion global sentiment representation is achieved through a fully connected layer. Perform feature mapping and output the sentiment fully connected feature vector; The sentiment fully connected feature vector is normalized, and the sentiment category probability vector C corresponding to the appeal text is calculated using the following formula: In the formula, The output layer weight matrix is ​​a learnable matrix. The output layer bias term; softmax() is the normalization exponential function; Based on the sentiment category probability vector C, the positive sentiment probability is split. Neutral emotional probability With the probability of negative emotions And calculate the emotional intensity score corresponding to the appeal text. The corresponding calculation formula is as follows: ； The emotion category probability vector C and the emotion intensity score are used as the basis for the expression. The data is integrated to construct the emotional feature vector corresponding to the stated appeal text.

7. The method for predicting government hotline satisfaction based on multi-source feature vector fusion according to claim 1, characterized in that, The joint training and iterative optimization process of the improved appeal sentiment extraction algorithm is as follows: Multiple teacher models are trained based on manually labeled sentiment samples. The unlabeled hotline data is then input into the multiple teacher models, and the predicted probability distribution of the unlabeled hotline samples is output. Based on the predicted probability distribution, the prediction confidence of the unlabeled samples of the hotline is calculated using the following formula: In the formula, The prediction confidence level for the i-th unlabeled hotline sample; Predicted probability distribution for the i-th unlabeled hotline sample Information entropy; K is the total number of sentiment categories; Select unlabeled hotline samples whose prediction confidence is greater than or equal to a preset confidence threshold and whose prediction results from multiple teacher models are consistent. Generate pseudo-labels for the selected unlabeled hotline samples and output a training sample set with pseudo-labels. The formula for calculating the pseudo-labels is as follows: In the formula, is the pseudo-label for the i-th unlabeled hotline sample; argmax() is the function that takes the category corresponding to the maximum probability; tau is the preset confidence threshold; This is an indicator that the prediction results of multiple teacher models are consistent; it takes a value of 1 when they are consistent and a value of 0 when they are inconsistent. The manually labeled sentiment samples are merged with the set of training samples with pseudo-labels to construct a joint training sample set. A weighted focus loss function is constructed based on the joint training sample set, and the training loss calculation result of the improved appeal sentiment extraction algorithm is output. The learnable parameters of the improved appeal sentiment extraction algorithm are optimized through the backpropagation algorithm to complete the joint training and iterative optimization of the improved appeal sentiment extraction algorithm.

8. The method for predicting government hotline satisfaction based on multi-source feature vector fusion according to claim 1, characterized in that, Based on the new thesaurus in the hotline domain, the process involves performing text feature quantification and syntactic structure analysis on the appeal text to output an appeal text feature vector. Furthermore, it involves multi-source feature fusion of the direct business feature vector, the appeal sentiment feature vector, and the appeal text feature vector to output a satisfaction prediction feature vector. Specifically, this includes: Based on the new thesaurus in the hotline field, text feature quantification is performed on the appeal text to extract a statistical text feature vector; wherein, the statistical text feature vector includes the basic word count, part-of-speech count, average sentence length, stop word ratio, and linking word frequency; Based on the new lexicon in the hotline field, syntactic structure analysis is performed on the appeal text to extract the language structural text feature vector of the appeal text; wherein, the language structural text feature vector includes minimum edit distance, dependency syntax ratio, inter-sentence similarity, and narrative integrity index; The statistical text feature vector is concatenated with the language structural text feature vector to output the appeal text feature vector; Obtain the direct business feature vector and the appeal sentiment feature vector, and perform feature alignment and standardization processing with the appeal text feature vector to output the aligned direct business feature vector, appeal sentiment feature vector and appeal text feature vector. The aligned direct business feature vector, appeal sentiment feature vector, and appeal text feature vector are concatenated and fused to construct the satisfaction prediction feature vector. The corresponding calculation formula is as follows: In the formula, Let be the feature vector for predicting the satisfaction level of the j-th government hotline work order; Let be the direct business feature vector of the j-th government hotline work order; The emotional intensity score of the request text of the j-th government hotline work order; Let be the probability vector of the sentiment category of the request text of the j-th government hotline work order; Let be the feature vector of the request text of the j-th government hotline work order.

9. The method for predicting government hotline satisfaction based on multi-source feature vector fusion according to claim 1, characterized in that, The process of inputting the satisfaction prediction feature vector into a pre-trained gradient boosting decision tree model and outputting the satisfaction prediction result, risk level, recommended follow-up method, and explanation of key influencing factors for satisfaction prediction for government hotline work orders specifically includes: The satisfaction prediction feature vector is input into the pre-trained gradient boosting decision tree model. Multiple weak learners of the gradient boosting decision tree model are used to predict and calculate the satisfaction prediction feature vector, and the prediction result of each weak learner is output. The weighted summation of the prediction results from each weak learner yields the ensemble output of the gradient boosting decision tree model, calculated using the following formula: In the formula, To improve the gradient boosting decision tree for predicting the satisfaction level of the i-th government hotline work order, we need to define the feature vector. The integrated output value; M is the total number of weak learners; Let m be the learning rate of the m-th weak learner; Let m be the feature vector of the satisfaction prediction by the m-th weak learner. The predicted output value; The satisfaction probability of the government hotline work order is calculated by mapping the integrated output value through an activation function, and the corresponding calculation formula is as follows: In the formula, Let be the satisfaction probability of the i-th government hotline work order; sigmoid() is the sigmoid activation function; A preset probability threshold is obtained, including a first probability threshold and a second probability threshold. Based on the preset probability threshold, the government hotline work order is classified into risk levels: when the satisfaction probability is less than the first probability threshold, the government hotline work order is classified into a high-risk unsatisfactory level; when the satisfaction probability is greater than or equal to the first probability threshold and less than the second probability threshold, the government hotline work order is classified into a medium-risk pending verification level; when the satisfaction probability is greater than or equal to the second probability threshold, the government hotline work order is classified into a low-risk satisfaction level. Based on the risk level of the government hotline work order, a corresponding preset follow-up strategy rule is matched and a recommended follow-up method is generated: a manual follow-up method is recommended for government hotline work orders with a high-risk dissatisfaction level; a manual review method or a semi-automatic follow-up method is recommended for government hotline work orders with a medium-risk pending verification level; and an automatic follow-up method is recommended for government hotline work orders with a low-risk satisfaction level. Based on the feature contribution value calculation logic of the gradient boosting decision tree model, the contribution of each dimension of the feature in the satisfaction prediction feature vector is sorted, and the core influencing feature set that has the greatest impact on the satisfaction prediction result is selected. Combined with the negative emotional segments and business process abnormal indicators in the appeal text, the explanation of the key influencing factors of the satisfaction prediction is generated. Based on the explanation of the key influencing factors for the satisfaction prediction, the recommended follow-up method is optimized and adjusted, and key reminders for the follow-up are added.

10. A government hotline satisfaction prediction system based on multi-source feature vector fusion, applied to the government hotline satisfaction prediction method based on multi-source feature vector fusion as described in any one of claims 1-9, characterized in that, The system includes: The data acquisition and preprocessing module is used to collect government hotline business data and perform standardized preprocessing to output a standardized government hotline business dataset. The new lexicon construction and model pre-training module is used to extract direct business feature vectors from the standardized government hotline business dataset, and at the same time, based on the unlabeled hotline corpus in the standardized government hotline business dataset, construct a new lexicon for the hotline domain and perform domain-adaptive pre-training on the basic language model. The emotional appeal extraction module is used to perform improved emotional appeal extraction on the appeal text in the standardized government hotline business dataset based on the new lexicon of the hotline domain and the basic language model after domain adaptive pre-training, and output the emotional appeal feature vector. The appeal text output and feature fusion module is used to perform text feature quantification and syntactic structure analysis on the appeal text based on the new thesaurus in the hotline domain, output the appeal text feature vector, and perform multi-source feature fusion on the direct business feature vector, the appeal sentiment feature vector and the appeal text feature vector to output the satisfaction prediction feature vector. The satisfaction prediction result output module is used to input the satisfaction prediction feature vector into the pre-trained gradient boosting decision tree model and output the satisfaction prediction result, risk level, recommended follow-up method, and explanation of key influencing factors of satisfaction prediction for the government hotline work order.

Images