A vehicle cabin user intention classification method and device, and a vehicle

Abstract

The embodiment of the application relates to the technical field of natural language processing, and discloses a vehicle cabin user intention classification method and device and a vehicle, which comprises the following steps: receiving a voice instruction input by a user through a voice acquisition module of the vehicle cabin, and converting the voice instruction into original query text; performing semantic purification processing on the original query text, eliminating redundant words with a semantic contribution degree lower than a preset threshold, and obtaining purified text; performing part-of-speech weighting semantic enhancement processing on the purified text, and generating a text vector; performing semantic generalization and expansion based on a pre-constructed intention category word group set, and generating a benchmark vector of each intention category; performing two-dimensional similarity calculation on the text vector and the benchmark vector of each intention category respectively, and obtaining a comprehensive matching score of each intention category through weighted fusion; and determining an intention category with the highest comprehensive matching score as a final classification result. The embodiment can solve the technical problem of low accuracy of intention classification of colloquial query text in the prior art.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, specifically to a method, apparatus, and vehicle for classifying user intent in a vehicle cabin. Background Technology

[0002] With the rapid development of smart cockpit technology, in-vehicle voice interaction systems have become an important means of human-machine interaction. User intent classification in the vehicle cockpit is a core component of the voice interaction system, and its accuracy directly determines the quality of subsequent command execution and information feedback.

[0003] Currently, existing user intent classification methods mainly suffer from the following shortcomings:

[0004] First, the ability to process colloquial expressions is insufficient. In in-vehicle scenarios, users' voice input often contains a large number of interjections, filler words, and meaningless conjunctions (such as "that," "um," "just," etc.). Existing methods usually use a general stop word list for simple filtering, lacking a refined semantic purification mechanism for colloquial scenarios. This results in noisy words interfering with subsequent semantic understanding and reducing classification accuracy.

[0005] Second, the semantic representation capability is limited. Existing methods mostly use bag-of-words models or simple average pooling of word vectors to generate text representations, which fail to fully consider the differences in the contribution of different parts of speech to the expression of intent. Core intent words (such as nouns and verbs) and modifying words (such as adjectives and adverbs) are treated equally, resulting in the dilution of key semantic information.

[0006] Third, the generalization coverage of intent categories is insufficient. Existing methods rely on manually maintained fixed keyword lists for intent matching. When faced with diverse spoken expressions from users, it is difficult to cover language variations such as synonym substitution and near-synonym generalization, resulting in low recall rates. Furthermore, adding new intent categories requires a large amount of manual annotation and configuration work, leading to poor scalability.

[0007] Fourth, single similarity metrics have limitations. Relying solely on cosine similarity or a single distance metric for intent matching makes it difficult to balance directional similarity and absolute distance differences when faced with intent category word sets of varying sizes, which can easily lead to misclassification. Summary of the Invention

[0008] In view of the shortcomings of the prior art, the purpose of this application is to provide a method, device and vehicle for classifying user intent in vehicle cabin, which aims to solve the technical problems of low intent classification accuracy, weak semantic representation ability, insufficient generalization coverage of intent categories and easy misclassification by a single similarity measurement method in the context of in-vehicle voice interaction.

[0009] In a first aspect, embodiments of this application provide a method for classifying user intent in a vehicle cockpit, applied to a voice interaction system in a vehicle cockpit, comprising:

[0010] Receive voice commands input by the user through the voice acquisition module in the vehicle cabin, and convert the voice commands into raw query text;

[0011] The original query text is semantically cleaned by removing redundant words whose semantic contribution is lower than a preset threshold, resulting in cleaned text.

[0012] The purified text is then subjected to part-of-speech weighted semantic enhancement processing to generate a text vector;

[0013] Semantic generalization expansion is performed based on a pre-built set of intent category phrases to generate a baseline vector for each intent category;

[0014] The text vectors are compared with the baseline vectors of each intent category using a two-dimensional similarity calculation. The results are then weighted and fused to obtain a comprehensive matching score for each intent category. The intent category with the highest comprehensive matching score is determined as the final classification result.

[0015] In the above technical solution, through the coordinated efforts of semantic purification, part-of-speech weighted semantic enhancement, phrase semantic generalization expansion, and two-dimensional similarity calculation, a complete intent classification link is formed, which extends from noise removal, semantic enhancement, coverage expansion to accurate matching, effectively improving the accuracy and robustness of user intent classification in in-vehicle conversational scenarios.

[0016] In one embodiment, the semantic cleanup of the original query text input by the user includes: jointly filtering based on a pre-built stop word dictionary, semantic redundancy determination results, and word frequency weight values, and removing words that simultaneously meet the requirements of stop word hit, semantic redundancy determination as weak redundancy, and word frequency weight values ​​below the average threshold.

[0017] In the above technical solution, the triple joint screening mechanism of stop word dictionary, semantic redundancy judgment and word frequency weight value avoids the problem of accidental deletion or omission caused by single-dimensional filtering. While effectively removing spoken noise words, it retains core words that contribute semantics and improves the semantic purity of input text.

[0018] In one embodiment, the part-of-speech weighted semantic enhancement processing of the purified text includes: segmenting and tagging the purified text with words; setting differentiated weights for each word according to its part-of-speech category, wherein the weights of nouns and verbs are higher than those of adjectives and adverbs; mapping each word to a word vector, weighting and fusing them according to their corresponding part-of-speech weights, and normalizing them to generate the text vector.

[0019] In the above technical solution, by using a part-of-speech differentiated weighting strategy, nouns and verbs that carry the core intent semantics obtain higher representation weights in the text vector, effectively suppressing the dilution effect of modifying words on intent expression, and improving the text vector's ability to distinguish and express key intent information.

[0020] In one embodiment, the semantic generalization expansion based on the pre-constructed intent category word set includes: using a pre-trained word vector model to perform semantic clustering on the words in each intent category word set, and classifying words with word vector similarity greater than or equal to a preset clustering threshold into the corresponding word set; performing cross-duplicate processing on the expanded word sets, and establishing a word set association mapping table to achieve cross-category reuse.

[0021] In the above technical solution, semantic clustering based on word vector model realizes the automatic generalization and expansion of word groups. It can cover the diverse spoken expression variations of users without the need for manual maintenance of the thesaurus, which significantly improves the recall rate of intent categories. At the same time, through cross-duplicate removal and association mapping table mechanism, word group redundancy and cross-category conflicts are avoided, and the configuration cost of adding new intent categories is reduced.

[0022] In one embodiment, the step of performing two-dimensional similarity calculation between the text vector and the baseline vector of each intent category includes: calculating the cosine similarity and Manhattan distance between the text vector and the baseline vector of each intent category respectively; adaptively adjusting the weight coefficients of the cosine similarity and the Manhattan distance according to the number of generalized phrases contained in each intent category; and fusing the weighted cosine similarity and Manhattan distance to obtain the comprehensive matching score.

[0023] In the above technical solution, the dual-dimensional fusion of cosine similarity and Manhattan distance takes into account the two complementary dimensions of vector directional similarity and absolute distance difference. At the same time, the weight coefficients are adaptively adjusted according to the number of generalized phrases in each intent category, so that categories with larger phrase sizes focus more on directional similarity and categories with smaller phrase sizes focus more on absolute distance, effectively reducing the misclassification rate between categories of different sizes.

[0024] In one embodiment, before determining the final classification result, the method further includes: if the original query text hits a preset prohibited word of a certain intent category, then the highest score qualification of that category is blocked, and the candidate category with the second highest score is extracted as the final classification result.

[0025] In the above technical solution, by introducing a prohibited word verification mechanism, a safety net is added to the intent classification results, effectively preventing misclassification between categories that are semantically similar but have mutually exclusive business logic, and improving the classification reliability and business security of the system in boundary scenarios.

[0026] Secondly, embodiments of this application provide a vehicle cabin user intent classification device, comprising:

[0027] The voice interaction interface module is used to receive voice commands input by the user through the voice acquisition module in the vehicle cabin and convert the voice commands into raw query text.

[0028] The semantic purification module is used to perform semantic purification processing on the original query text, remove redundant words whose semantic contribution is lower than a preset threshold, and obtain the purified text.

[0029] The semantic enhancement module is used to perform part-of-speech weighted semantic enhancement processing on the purified text to generate text vectors;

[0030] The phrase generalization extension module is used to perform semantic generalization extension based on a pre-built set of intent category phrases, generating a baseline vector for each intent category;

[0031] The dual-dimensional vector matching module is used to calculate the dual-dimensional similarity between the text vector and the baseline vector of each intent category, and then weighted and fused to obtain the comprehensive matching score of each intent category. The intent category with the highest comprehensive matching score is determined as the final classification result.

[0032] In the above technical solution, through the coordinated cooperation of various functional modules, a complete intent classification process is realized, which extends from semantic cleansing, semantic enhancement, and phrase generalization to two-dimensional vector matching. The modular design facilitates independent optimization and flexible expansion of each module, thereby improving the maintainability of the system and the efficiency of engineering deployment.

[0033] In one embodiment, the semantic enhancement module is specifically used to: segment and tag the purified text with words, set differentiated weights for each word according to the part-of-speech category, wherein the weights of nouns and verbs are higher than the weights of adjectives and adverbs; map each word into a word vector, perform weighted fusion and normalization according to the corresponding part-of-speech weights, and generate the text vector.

[0034] In the above technical solution, the semantic enhancement module generates text vectors through a part-of-speech differentiated weighted fusion strategy, which gives core intent words higher representation weights and effectively improves the ability of text vectors to express key semantic information and their distinguishability.

[0035] In one embodiment, the dual-dimensional vector matching module is further configured to: if the original query text matches a preset prohibited word of a certain intent category, then block the highest score qualification of that category and extract the candidate category with the second highest score as the final classification result.

[0036] In the above technical solution, the dual-dimensional vector matching module integrates the prohibited word verification function, which adds security constraints at the business rule level on the basis of vector matching scoring. This effectively prevents misclassification between semantically similar but business-exclusive categories, and improves the business reliability of the classification results.

[0037] Thirdly, embodiments of this application provide a vehicle, including an in-vehicle voice interaction system and a processor. The processor is configured to execute the above-described vehicle cockpit user intent classification method to classify the intent of a query text input by a user through the in-vehicle voice interaction system, and to execute corresponding vehicle control commands or information feedback operations based on the classification results.

[0038] In the above technical solution, by applying the above vehicle cockpit user intent classification method to the vehicle's in-vehicle voice interaction system, the vehicle can accurately understand the user's true intent expressed in spoken language, and execute precise vehicle control or information feedback accordingly, thereby improving the intelligence level and user experience of in-vehicle voice interaction.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] 1. This application achieves precise removal of noisy words in colloquial text through a joint screening mechanism of stop word dictionary, semantic redundancy judgment and word frequency weight, which effectively improves the semantic purity of input text;

[0041] 2. This application strengthens the representation weight of core intent words in text vectors and improves the distinguishability of semantic representation by adopting a part-of-speech differentiated weighted fusion strategy.

[0042] 3. This application automatically covers synonym and near-synonymous expression variants through a semantic clustering generalization extension mechanism based on word vector models, which significantly improves the recall rate and system scalability of intent categories;

[0043] 4. This application adopts a two-dimensional adaptive weighted fusion of cosine similarity and Manhattan distance, which takes into account both directional similarity and absolute distance difference, and effectively reduces the misclassification rate. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of this application or the background art, the accompanying drawings used in the embodiments of this application will be described below.

[0045] Figure 1 A flowchart illustrating the vehicle cockpit user intent classification method provided in this application embodiment;

[0046] Figure 2 A flowchart illustrating the word2Vec clustering-driven phrase generalization expansion and reuse strategy provided in this application embodiment;

[0047] Figure 3 A logic block diagram of the two-dimensional vector matching algorithm combining cosine similarity and Manhattan distance provided in the embodiments of this application;

[0048] Figure 4 This is a bar chart comparing the performance of the embodiments of this application with existing technologies in terms of classification accuracy and recall.

[0049] Figure 5 This is a schematic diagram of the structure of the vehicle cockpit user intent classification device provided in the embodiments of this application.

[0050] Explanation of reference numerals in the attached figures:

[0051] 100 - Voice interaction interface module, 200 - Semantic purification module, 300 - Semantic enhancement module, 400 - Phrase generalization extension module, 500 - Two-dimensional vector matching module. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0053] Currently, existing methods for classifying user intent in vehicle cabins primarily employ a general stop word list for simple filtering, followed by text representation generation using a bag-of-words model or equal-weighted word vector average pooling. Intent matching is then performed using a manually maintained fixed keyword list combined with a single cosine similarity metric. However, these methods have significant shortcomings in in-vehicle conversational scenarios: the general stop word list lacks refined purification capabilities for conversational scenarios; the equal-weighted fusion method fails to reflect the differences in contribution of different parts of speech to intent expression; the fixed keyword list struggles to cover diverse variations in user conversational expressions; and the single similarity metric is prone to misclassification when faced with intent category phrase sets of varying sizes. To address these issues, this application provides a method, apparatus, and vehicle for classifying user intent in vehicle cabins based on semantic purification, phrase generalization, and enhanced vector matching.

[0054] The vehicle cockpit user intent classification method provided in this application can be applied to scenarios requiring accurate understanding of users' spoken intent, such as in-vehicle voice interaction systems, intelligent customer service, smart speakers, search and recommendation systems, and dialogue systems. The following detailed explanation uses an in-vehicle voice interaction scenario as an example.

[0055] In one embodiment, reference is made to Figure 1 This application provides a method for classifying user intent in a vehicle cockpit, applied to a voice interaction system in a vehicle cockpit, comprising the following steps:

[0056] Step S100: Receive the voice command input by the user through the voice acquisition module in the vehicle cabin, and convert the voice command into the original query text.

[0057] Specifically, the in-vehicle voice interaction system receives the voice signal input by the user through the microphone and converts it into the original query text through the speech recognition engine. Since the user's voice input in the in-vehicle scenario often contains a large number of interjections, filler words, and meaningless conjunctions, such as "that," "um," "just," and "then," these redundant words can interfere with subsequent semantic understanding, so semantic cleanup processing is required.

[0058] Step S200: Perform semantic purification on the original query text, remove redundant words with semantic contribution below a preset threshold, and obtain the purified text.

[0059] In this embodiment, semantic purification processing is performed by the semantic purification module 200. The semantic purification module 200 performs joint screening based on a pre-built stop word dictionary, semantic redundancy determination results, and word frequency weight values, and removes words that simultaneously meet the requirements of stop word hit, weak semantic redundancy determination, and word frequency weight values ​​below the average threshold.

[0060] Specifically, the joint screening mechanism of the semantic sanitization module 200 includes the following three dimensions:

[0061] The first dimension involves pre-constructing a stop word dictionary specifically for spoken language, which includes commonly used redundant words in spoken language such as interjections, filler words, and meaningless conjunctions. This stop word dictionary differs from general stop word lists and is specifically customized for in-vehicle spoken language scenarios, including high-frequency redundant spoken expressions such as "um," "ah," "that," "that is to say," and "then what."

[0062] The second dimension involves configuring a semantic redundancy judgment algorithm. Combining the semantic relevance of words within their context, the algorithm categorizes words in the text into three levels: strongly redundant words, weakly related core words, and strongly core words. Strongly redundant words refer to words that do not carry effective semantics in any context, such as purely modal particles. Weakly related core words refer to words that may carry some semantics in a specific context but with low contribution. Strongly core words refer to keywords that carry the user's core intent, such as "navigation," "air conditioning," and "music."

[0063] The third dimension introduces the TF-IDF (Term Frequency - Inverse Document Frequency) algorithm to calculate the term frequency weight value of each word in the text, and compares the term frequency weight value with the global mean threshold.

[0064] The semantic cleansing module 200 only removes words that simultaneously meet the following three conditions: 1) they are found in the stop word dictionary; 2) the semantic redundancy determination result is weak redundancy; and 3) the TF-IDF value is lower than the mean threshold. All other core words that contribute semantically are retained to prevent over-cleaning from causing a loss of user query intent.

[0065] For example, if a user enters "Um, please turn on the air conditioner for me", after processing by the semantic cleansing module 200, "um", "for me" and "for me" are removed, and the core semantic words such as "help me", "turn on", and "air conditioner" are retained, resulting in the cleaned text "help me turn on the air conditioner".

[0066] This step employs a triple-screening mechanism combining a stop word dictionary, semantic redundancy assessment, and word frequency weighting to avoid the erroneous or missed deletions caused by single-dimensional filtering. Compared to existing technologies that rely solely on stop word lists for simple filtering, this embodiment effectively removes colloquial noise words while retaining core words that contribute semantically. It avoids both over-purification resulting in the loss of core semantics and incomplete purification leaving behind colloquial redundancy words, significantly improving the semantic purity of the input text and laying a high-quality data foundation for subsequent semantic enhancement and vector matching.

[0067] Step S300: Perform part-of-speech weighted semantic enhancement processing on the purified text to generate text vectors.

[0068] Specifically, the semantic enhancement module 300 receives the purified text output by the semantic purification module 200 and performs part-of-speech weighted semantic enhancement processing on it.

[0069] In this embodiment, the semantic enhancement module 300 specifically performs the following operations:

[0070] First, the purified text is segmented and labeled with parts of speech to identify the part-of-speech categories of each word, including nouns, verbs, adjectives, adverbs, etc.

[0071] Then, differentiated weights were assigned to each word based on its part of speech. Nouns and verbs, which carry the core semantic intent, were assigned a high weight range of 0.6–0.8, while adjectives and adverbs, which only serve a modifying function, were assigned a medium-low weight range of 0.2–0.4. This weighting reflects the different contributions of different parts of speech to the expression of intent: nouns typically carry the object of intent, such as "air conditioner," "navigation," and "music"; verbs typically carry the action of intent, such as "turn on," "turn off," and "adjust," together constituting the core semantics of the user's intent; while adjectives and adverbs primarily serve a modifying and limiting function, contributing relatively little to intent classification.

[0072] Finally, the pre-trained Word2Vec (Word to Vector, word-to-vector) model, which is a model that converts words into vector representations, is called to map each word to a word vector of a fixed dimension. All word vectors are weighted and summed according to the corresponding part-of-speech weights and then normalized to generate a sentence-level text vector that highlights the core semantics, so as to weaken the interference of spoken modifier words on vector representation.

[0073] For example, in the purified text "help me turn on the air conditioner", "turn on" is a verb, which is assigned a weight of 0.8; "air conditioner" is a noun, which is assigned a weight of 0.7; "help me" is a pronoun plus verb combination, which is assigned a weight of 0.5. After each word is mapped to a word vector by the Word2Vec model, weighted fusion and normalization are performed according to the above weights to generate the final text vector.

[0074] This step adopts a differentiated part-of-speech weighting strategy, so that nouns and verbs carrying core intention semantics obtain higher representation weights in the text vector, which effectively suppresses the dilution effect of modifier words on intention expression. Compared with the method of equal-weight accumulation of words adopted in the prior art, this embodiment significantly improves the discrimination and expression ability of the text vector for key intention information, so that the generated text vector can more accurately reflect the real intention of the user.

[0075] Step S400: Perform semantic generalization and expansion based on the pre-constructed intention category phrase set to generate a reference vector for each intention category.

[0076] Specifically, the phrase generalization and expansion module 400 is responsible for automatically generalizing and expanding the basic phrase set of each intention category.

[0077] In this embodiment, the phrase generalization and expansion module 400 performs the following operations:

[0078] First, each basic phrase set is pre-constructed according to different business intention categories. For example, the basic phrase set of the "air conditioner control" category may include "turn on the air conditioner", "turn off the air conditioner", "adjust temperature", "cooling", "heating", etc.; the basic phrase set of the "navigation control" category may include "navigate to", "go there", "how to get there", "route planning", etc.

[0079] Then, the Word2Vec word vector model is trained using domain spoken corpora to perform semantic clustering calculation on the words in each basic phrase set, traverse the global vocabulary library, and automatically assign synonyms whose word vector similarity is greater than or equal to the preset clustering threshold to the corresponding phrase set, so as to complete the automatic generalization and expansion without manually sorting out synonyms one by one. For example, the preset clustering threshold is 0.85.

[0080] Next, the expanded word groups are cross-matched, and the word vector similarity is calculated pairwise. If the similarity is greater than or equal to a preset deduplication threshold, it is determined to be a synonymous duplicate, and a single word group is retained while duplicate entries are removed, thus simplifying the size of the generalized word group library and avoiding data redundancy. For example, the preset deduplication threshold is 0.9.

[0081] Finally, a word set association mapping table is established, marking common business word sets such as commuting, querying, and settings as reusable resources. When a new intent category is added, the common word set is directly called from the mapping table, and only the word set specific to the current category needs to be added to complete the configuration, enabling rapid reuse across scenarios and categories.

[0082] After generalization expansion is completed, the phrase generalization expansion module 400 performs weighted fusion of all word vectors in the expanded phrase set for each intent category to generate a baseline vector for each intent category.

[0083] This step achieves automated generalization and expansion of word groups through semantic clustering based on the Word2Vec model. It covers diverse spoken expression variations of users without requiring manual maintenance of thesaurus, significantly improving the recall rate of intent categories. At the same time, through cross-duplicate removal and association mapping mechanisms, it avoids word group redundancy and cross-category conflicts, greatly reducing the configuration cost of adding new intent categories and the workload of subsequent corpus maintenance, giving the system good scalability.

[0084] Step S500: Perform two-dimensional similarity calculations between the text vector and the baseline vector of each intent category, and then perform weighted fusion to obtain the comprehensive matching score of each intent category. The intent category with the highest comprehensive matching score is determined as the final classification result.

[0085] Specifically, the dual-dimensional vector matching module 500 receives the text vector output by the semantic enhancement module 300 and the baseline vectors of each intent category output by the phrase generalization extension module 400, and performs dual-dimensional similarity calculation.

[0086] In this embodiment, the two-dimensional vector matching module 500 performs the following operations:

[0087] First, we calculate the cosine similarity and Manhattan distance between the text vector and the baseline vector for each intent category. The cosine similarity measures the degree of semantic alignment between the vectors, with a value range of [-1, 1]. A larger value indicates a more consistent semantic alignment. The Manhattan distance measures the absolute distance between the vectors in the feature space. A smaller distance indicates a higher degree of semantic fit.

[0088] Then, the weight coefficients of cosine similarity and Manhattan distance are adaptively adjusted based on the number of generalized phrases contained in each intent category. Specifically, when a certain intent category contains a large number of generalized phrases, the weight of cosine similarity is increased, making the matching more focused on directional similarity; when a certain intent category contains a small number of generalized phrases, the weight of Manhattan distance is increased, making the matching more focused on absolute distance.

[0089] Finally, the weighted cosine similarity and Manhattan distance are fused to obtain a comprehensive matching score for each intent category. The intent category with the highest comprehensive matching score is determined as the final classification result.

[0090] This step integrates cosine similarity and Manhattan distance in a dual-dimensional manner, taking into account both vector directional similarity and absolute distance difference—two complementary dimensions. This overcomes the limitation of existing technologies that rely solely on cosine similarity and cannot quantify the closeness between semantically similar pairs. Furthermore, it adaptively adjusts the weighting coefficients based on the number of generalized phrases for each intent category, effectively reducing misclassification rates between categories of different sizes. This achieves a balance between matching accuracy and inference efficiency without introducing complex models or additional computational processes.

[0091] As an optional implementation of this application, the specific implementation of semantic cleanup processing is further defined.

[0092] The semantic cleanup module 200 pre-builds a speech-specific stop word dictionary. This dictionary differs from general NLP stop word lists and is specifically customized and optimized for in-vehicle speech scenarios. The types of words included in the stop word dictionary include, but are not limited to: pure interjections (such as "um", "ah", "oh", "uh"), filler words (such as "that", "this", "it is"), meaningless conjunctions (such as "then what", "next", "so what"), and polite redundant expressions (such as "trouble", "please ask", "sorry"), etc.

[0093] The semantic redundancy determination algorithm is based on the semantic relevance of words in context. Specifically, for each word in the text, its semantic relevance score with its preceding and following words is calculated, and the word is classified into strongly redundant words, weakly related core words, or strongly related core words based on the score. Strongly redundant words are words that do not carry effective semantics in any context; weakly related core words are words that may carry some semantics in a specific context but have a low overall contribution; and strongly related core words are keywords that carry the user's core intent.

[0094] The term frequency weights are calculated using the TF-IDF algorithm. For each word in the current query text, its term frequency (TF) in the current text is multiplied by its inverse document frequency (IDF) in the entire corpus to obtain the TF-IDF value of that word. The average of the TF-IDF values ​​of all words in the current text is then used as the mean threshold.

[0095] The joint filtering logic of the semantic cleanup module 200 is as follows: a word is removed only when it meets all three of the following conditions: 1) the word matches the stop word dictionary; 2) the semantic redundancy determination algorithm determines the word as weakly redundant; and 3) the word's TF-IDF value is lower than the mean threshold. All three conditions are indispensable to ensure that core words with semantic contributions are not mistakenly deleted.

[0096] Compared to single-dimensional filtering methods, the triple-screening mechanism offers higher accuracy and robustness. The stop word dictionary provides basic redundancy recognition capabilities, the semantic redundancy judgment algorithm provides context-aware semantic contribution evaluation capabilities, and the TF-IDF value provides a statistically based word frequency weight reference. The three work synergistically to effectively remove noisy, colloquial words while preserving core semantically contributing words to the greatest extent possible, avoiding the problems of over-filtering leading to lost intent or incomplete filtering leaving redundant words.

[0097] As an optional implementation of this application, the specific implementation of the part-of-speech weighted semantic enhancement processing is further defined.

[0098] When performing word segmentation and part-of-speech tagging on the purified text, the semantic enhancement module 300 can use word segmentation tools based on hidden Markov models or conditional random fields to perform Chinese word segmentation, and use part-of-speech tagging sets to tag each word segmentation result to identify noun, verb, adjective, adverb and other part-of-speech categories.

[0099] The specific settings for part-of-speech weights are as follows: nouns and verbs are set to a high weight range of 0.6 to 0.8, adjectives and adverbs are set to a medium-low weight range of 0.2 to 0.4, and other parts of speech (such as pronouns, prepositions, conjunctions, etc.) are set to a medium weight range of 0.4 to 0.6.

[0100] The formula for weighted fusion is:

[0101] in, The final generated text vector, Let i be the part-of-speech weight of the i-th word. Let be the Word2Vec word vector of the i-th word, and Normalize is the L2 normalization operation.

[0102] This application significantly increases the contribution of core intent words (nouns and verbs) to the final text vector by clearly defining part-of-speech weight intervals and using a Word2Vec word vector weighted fusion formula. Experiments show that, compared to the equal-weight fusion method, the text vectors generated by part-of-speech weighted fusion have significantly improved discriminative power in intent classification tasks, effectively mitigating the interference of colloquial modifiers on vector representation.

[0103] As an optional implementation of this application, to further define the specific implementation of phrase generalization expansion, refer to... Figure 2 The phrase generalization and reuse strategy of the phrase generalization extension module 400 includes the following specific steps:

[0104] Step S410, Construction of Basic Phrases. Basic phrase sets are pre-constructed according to different business intent categories. For example:

[0105] "Air Conditioning Control" category: Turn on air conditioning, turn off air conditioning, adjust temperature, cool, heat, adjust air volume;

[0106] "Navigation Control" category: Navigation to, where to go, how to get there, route planning, destination;

[0107] "Music Controls" category: Play music, skip tracks, pause, next track, volume adjustment;

[0108] "Window Control" category: Open window, close window, raise window, lower window.

[0109] Step S420, Word2Vec semantic clustering expansion. A Word2Vec word vector model is trained using a domain-specific spoken language corpus. Word vectors are calculated for each word in each basic word group set. Then, the global vocabulary is traversed, and the similarity between the global vocabulary and the word vectors in each basic word group set is calculated. Synonyms with a similarity greater than or equal to 0.85 are automatically assigned to their corresponding word group sets. For example, the category "air conditioning control" might be automatically expanded to include synonyms such as "cold air," "warm air," "temperature adjustment," "temperature adjustment," and "automatic air conditioning."

[0110] Step S430: Cross-combination deduplication. Cross-combination matching is performed on the expanded multiple word groups, and the word vector similarity is calculated pairwise. If the word vector similarity between two word groups is greater than or equal to 0.9, they are determined to be synonymous duplicates. The single word group with a broader semantic coverage is retained, and duplicate entries are removed, thus simplifying the size of the generalized word group library.

[0111] Step S440: Establish a phrase set association mapping table. Mark common business phrase sets such as commuting, querying, and settings as reusable resources and include them in the mapping table for unified management. When a new intent category is added, the common phrase set is directly called from the mapping table, and only the phrases specific to the current category need to be added to complete the configuration.

[0112] Step S450, baseline vector generation. All word vectors within the expanded word set for each intent category are weighted and fused. The fusion weights can be set with reference to the part-of-speech weights to generate the baseline vector for each intent category.

[0113] This application employs an integrated phrase generalization scheme that combines Word2Vec clustering expansion, cross-combination deduplication, and cross-category reuse to achieve automated expansion and refined management of phrase sets. Compared to existing technologies that rely on manual maintenance of synonym lists, this embodiment significantly improves generalization efficiency and reduces the costs of subsequent category iterations and corpus maintenance. Simultaneously, the cross-deduplication mechanism effectively avoids redundant expansion of the generalized corpus, and the association mapping table mechanism enables rapid reuse across scenarios and categories, giving the system extremely high configuration efficiency when facing new intent categories.

[0114] As an optional implementation of this application, to further define the specific implementation of two-dimensional vector matching, refer to... Figure 3 The two-dimensional vector matching algorithm of the two-dimensional vector matching module 500 includes the following specific steps:

[0115] Step S510, cosine similarity calculation. Calculate the text vector. Baseline vectors for each intent category Cosine similarity between them:

[0116] .

[0117] The range of cosine similarity is [-1, 1]. The larger the value, the more consistent the two vectors are in the semantic direction. It is used to determine whether a text belongs to a certain intent category.

[0118] Step S520, Manhattan distance calculation. Calculate the text vector. Baseline vectors for each intent category Manhattan distance between them:

[0119] .

[0120] Manhattan distance measures the absolute distance between vectors in the feature space. The smaller the distance, the higher the semantic fit, and it is used to distinguish the similarity between vectors of the same category.

[0121] Step S530, Adaptive weight adjustment. Based on the number N of generalized phrases contained in each intent category, the weight coefficients of cosine similarity and Manhattan distance are adaptively adjusted. Specifically:

[0122] When N is large (there are many generalized phrases), increase the cosine similarity weight α and decrease the Manhattan distance weight β to make the matching focus more on directional similarity;

[0123] When N is small (few generalized phrases), decrease the cosine similarity weight α and increase the Manhattan distance weight β to make the matching focus more on absolute distance.

[0124] The weight adjustment formula is: α=N / (N+K), β=K / (N+K), where K is the adjustment constant.

[0125] Step S540, Weighted Fusion Score. The weighted cosine similarity and normalized Manhattan distance are fused to obtain a comprehensive matching score:

[0126]

[0127] in, This represents the normalized Manhattan distance. The comprehensive matching scores for each intent category are ranked, and the category with the highest score is selected as the initial classification result.

[0128] This implementation employs a dual-dimensional vector matching strategy instead of the traditional single cosine similarity matching mode. By combining cosine similarity to determine semantic direction consistency with Manhattan distance to quantify semantic closeness, it significantly improves the ability to distinguish between spoken query categories with similar expressions and finer-grained intents. An adaptive weight adjustment mechanism allows the system to dynamically optimize the matching strategy based on the word size of different categories, effectively reducing the misclassification rate between categories of different sizes. The entire matching process only adds distance calculations to the original vector calculations, without introducing complex models or additional computational processes. The inference time remains essentially unchanged, effectively improving classification accuracy while maintaining lightweight operation.

[0129] As an optional implementation of this application, the specific implementation of the prohibited word verification mechanism is further defined.

[0130] Before determining the final classification result, the two-dimensional vector matching module 500 also performs a prohibited word check operation. Specifically:

[0131] Pre-defined prohibited keyword libraries are set for each intent category. For example, prohibited words for the "air conditioning control" category could include "navigation," "route," and "destination"; prohibited words for the "navigation control" category could include "air conditioning," "temperature," and "cooling." The prohibited words are set based on mutually exclusive business logic, meaning that when a prohibited word of a certain category appears in the query text, it indicates that the user's intent is likely not to belong to that category.

[0132] If the query text matches a prohibited word in a certain intent category, the highest-scoring category is masked, and the candidate category with the second-highest score is extracted. Further, the Manhattan distance difference between the highest-scoring category and the second-highest-scoring category is calculated. If the difference is greater than or equal to 0.1, the second-highest category is directly determined as the final intent classification result; if the difference is less than 0.1, other auxiliary information can be combined for a comprehensive determination. Alternatively, in a simplified implementation, the second-highest candidate is directly taken as the final result after a prohibited word is matched.

[0133] For example, if a user inputs "navigate to the nearest gas station," and the overall matching score for the "air conditioning control" category is unexpectedly high due to semantic generalization, but the query text matches the prohibited word "navigation" within the "air conditioning control" category, then the highest-scoring category for "air conditioning control" is masked, and the second-highest-scoring category, "navigation control," is extracted. The calculated Manhattan distance difference between the two categories is 0.15 (≥0.1), so "navigation control" is directly determined as the final classification result.

[0134] This application introduces a prohibited word verification mechanism, adding a business rule-level security safeguard to the intent classification results. This mechanism works in conjunction with the two-dimensional vector matching logic to effectively prevent misclassification between semantically similar but logically mutually exclusive categories without increasing computational overhead, thus improving the system's classification reliability and business security in boundary scenarios.

[0135] In one embodiment, reference is made to Figure 5 This application provides a vehicle cabin user intent classification device, including:

[0136] The voice interaction interface module 100 is used to receive voice commands input by the user through the voice acquisition module in the vehicle cockpit and convert the voice commands into raw query text.

[0137] The semantic purification module 200 is used to perform semantic purification on the original query text input by the user, removing redundant words whose semantic contribution is lower than a preset threshold, and obtaining the purified text. Specifically, the semantic purification module 200 performs joint screening based on a pre-built stop word dictionary, semantic redundancy judgment results, and word frequency weight values, and removes words that simultaneously meet the requirements of stop word hit, weak semantic redundancy judgment, and word frequency weight value lower than the average threshold.

[0138] The semantic enhancement module 300 is used to perform part-of-speech weighted semantic enhancement processing on the purified text to generate text vectors. Specifically, the semantic enhancement module 300 performs word segmentation and part-of-speech tagging on the purified text, sets differentiated weights for each word according to its part-of-speech category, with nouns and verbs having higher weights than adjectives and adverbs; maps each word to a word vector, performs weighted fusion and normalization according to the corresponding part-of-speech weights, and generates text vectors.

[0139] The phrase generalization extension module 400 is used to perform semantic generalization extension based on a pre-built set of intent category phrases, generating a baseline vector for each intent category. Specifically, the phrase generalization extension module 400 uses a pre-trained word vector model to perform semantic clustering on the words in each intent category phrase set, assigning words with word vector similarity greater than or equal to a preset clustering threshold to the corresponding phrase set; it performs cross-duplicate processing on the extended phrases and establishes a phrase set association mapping table to achieve cross-category reuse.

[0140] The two-dimensional vector matching module 500 calculates the two-dimensional similarity between the text vector and the baseline vector of each intent category, and then weights and fuses them to obtain a comprehensive matching score for each intent category. The intent category with the highest comprehensive matching score is determined as the final classification result. Specifically, the two-dimensional vector matching module 500 calculates the cosine similarity and Manhattan distance between the text vector and the baseline vector of each intent category; adaptively adjusts the weight coefficients of the cosine similarity and Manhattan distance according to the number of generalized phrases contained in each intent category; and fuses the weighted cosine similarity and Manhattan distance to obtain the comprehensive matching score.

[0141] The dual-dimensional vector matching module 500 is also used to: if the original query text hits a preset prohibited word of a certain intent category, then the highest score qualification of that category is blocked, and the candidate category with the second highest score is extracted as the final classification result.

[0142] This application achieves a complete intent classification process, from semantic cleansing, semantic enhancement, and phrase generalization to two-dimensional vector matching, through the coordinated operation of a voice interaction interface module 100, a semantic cleansing module 200, a semantic enhancement module 300, a phrase generalization extension module 400, and a two-dimensional vector matching module 500. The modular design facilitates independent optimization and flexible expansion of each module, improving system maintainability and engineering deployment efficiency.

[0143] In one embodiment, this application provides a vehicle including an in-vehicle voice interaction system and a processor. The processor is configured to execute the vehicle cockpit user intent classification method of any of the above embodiments to classify the intent of the query text input by the user through the in-vehicle voice interaction system, and execute corresponding vehicle control commands or information feedback operations based on the classification results.

[0144] Specifically, the in-vehicle voice interaction system includes a microphone array, a voice recognition engine, an intent classification processor, and an execution feedback unit. Users input voice commands through the microphone array, the voice recognition engine converts the voice signal into text, the intent classification processor performs intent classification on the text using the aforementioned vehicle cabin user intent classification method, and the execution feedback unit executes corresponding vehicle control commands (such as controlling the air conditioning, navigation, windows, music, etc.) or information feedback operations (such as voice broadcasting, screen display, etc.) based on the classification results.

[0145] The vehicle can be, but is not limited to, a pure electric vehicle (PEV / BEV), a hybrid electric vehicle (HEV), a range-extended electric vehicle (REEV), a plug-in hybrid electric vehicle (PHEV), a new energy vehicle, or a gasoline vehicle.

[0146] This embodiment applies the above-mentioned vehicle cockpit user intent classification method to the vehicle's in-vehicle voice interaction system, enabling the vehicle to accurately understand the user's true intent expressed in spoken language and execute precise vehicle control or information feedback accordingly. This effectively improves the intelligence level and user experience of in-vehicle voice interaction and reduces the risk of erroneous operation caused by misclassification of intent.

[0147] Reference Figure 4 The comparison diagram shown is intended to objectively verify the technical effects of the embodiments of this application. In this embodiment, a control experiment was set up on a self-constructed in-vehicle colloquial query corpus. The experimental configuration is as follows.

[0148] Test dataset: A total of 5,000 spoken language query texts were collected from real in-vehicle scenarios, covering 10 intent categories including air conditioning control, navigation control, music control, window control, and telephone control. The corpus generally contains interjections, filler words, and redundant spoken expressions such as omissions and inversions, which significantly interfere with intent recognition and can fully test the robustness of the method under spoken language conditions.

[0149] The baseline method is the existing vehicle cockpit user intent classification method. It first uses a general stop word list to filter the text, then generates text representations through bag-of-words model or equal-weighted word vector average pooling, and finally relies on a manually maintained fixed keyword list and combines single cosine similarity for intent matching.

[0150] The method of this application: namely, the complete technical solution provided in the embodiments of this application, which sequentially connects four processing steps: "semantic purification - part-of-speech weighted semantic enhancement - phrase generalization expansion - two-dimensional vector matching".

[0151] Experimental Results: This experiment used precision and recall as evaluation metrics. Precision measures the correctness of the classification results, while recall measures the completeness of the identification of each intent. Figure 4 As can be seen, in terms of accuracy, the method of this application is approximately 99%, while the first comparative method is approximately 76%, representing an improvement of approximately 23 percentage points over the baseline method; in terms of recall, the method of this application is approximately 94%, while the first comparative method is approximately 62%, representing an improvement of approximately 32 percentage points over the baseline method.

[0152] Furthermore, the aforementioned gains can be attributed to each stage of this scheme: the significant improvement in accuracy mainly stems from the removal of colloquial noise words by "semantic purification" and the strengthening of the representation weights of core intent words such as nouns and verbs by "part-of-speech weighted semantic enhancement," making the text vectors closer to the true intent and thus suppressing misclassification caused by noise words and modifiers; the significant improvement in recall mainly stems from the automatic coverage of colloquial variant expressions such as synonyms and near-synonyms by "semantic clustering-based phrase generalization expansion," and the suppression of missed classifications between categories of different sizes by "cosine similarity + Manhattan distance" dual-dimensional matching. In summary, the experimental results show that, compared with traditional methods, the classification accuracy of this application is improved by approximately 23 percentage points and the recall is improved by approximately 32 percentage points, fully validating the effectiveness and synergistic gains of each stage of semantic purification, part-of-speech weighted semantic enhancement, phrase generalization expansion, and dual-dimensional vector matching.

[0153] The above embodiments are merely preferred embodiments provided to fully illustrate the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are all within the scope of protection of the present invention.

Claims

1. A method for classifying user intent in a vehicle cockpit, characterized in that, Voice interaction systems applied in vehicle cockpits include: Receive voice commands input by the user through the voice acquisition module in the vehicle cabin, and convert the voice commands into raw query text; The original query text is semantically cleaned by removing redundant words whose semantic contribution is lower than a preset threshold, resulting in cleaned text. The purified text is then subjected to part-of-speech weighted semantic enhancement processing to generate a text vector; Semantic generalization expansion is performed based on a pre-built set of intent category phrases to generate a baseline vector for each intent category; The text vectors are compared with the baseline vectors of each intent category using a two-dimensional similarity calculation. The results are then weighted and fused to obtain a comprehensive matching score for each intent category. The intent category with the highest comprehensive matching score is determined as the final classification result.

2. The vehicle cockpit user intent classification method according to claim 1, characterized in that, The semantic cleanup of the user-input original query text includes: joint filtering based on a pre-built stop word dictionary, semantic redundancy determination results, and word frequency weight values, and removing words that simultaneously meet the requirements of stop word hit, weak semantic redundancy determination, and word frequency weight values ​​below the average threshold.

3. The vehicle cockpit user intent classification method according to claim 1, characterized in that, The part-of-speech weighted semantic enhancement processing of the purified text includes: segmenting and tagging the purified text with words; setting differentiated weights for each word according to its part-of-speech category, wherein the weights of nouns and verbs are higher than those of adjectives and adverbs; mapping each word to a word vector, weighting and normalizing them according to their corresponding part-of-speech weights to generate the text vector.

4. The vehicle cockpit user intent classification method according to claim 1, characterized in that, The semantic generalization extension based on the pre-constructed intent category word set includes: using a pre-trained word vector model to perform semantic clustering of words in each intent category word set, and classifying words with word vector similarity greater than or equal to a preset clustering threshold into the corresponding word set; performing cross-duplicate processing on the extended word sets, and establishing a word set association mapping table to achieve cross-category reuse.

5. The vehicle cockpit user intent classification method according to claim 1, characterized in that, The step of performing two-dimensional similarity calculations between the text vector and the baseline vector of each intent category includes: calculating the cosine similarity and Manhattan distance between the text vector and the baseline vector of each intent category respectively; adaptively adjusting the weight coefficients of the cosine similarity and the Manhattan distance according to the number of generalized phrases contained in each intent category; and fusing the weighted cosine similarity and Manhattan distance to obtain the comprehensive matching score.

6. The vehicle cockpit user intent classification method according to claim 1, characterized in that, Before determining the final classification result, the process also includes: if the original query text hits a preset prohibited word of a certain intent category, then the highest score qualification of that category is blocked, and the candidate category with the second highest score is extracted as the final classification result.

7. A vehicle cabin user intent classification device, characterized in that, include: The voice interaction interface module is used to receive voice commands input by the user through the voice acquisition module in the vehicle cabin and convert the voice commands into raw query text. The semantic purification module is used to perform semantic purification processing on the original query text, remove redundant words whose semantic contribution is lower than a preset threshold, and obtain the purified text. The semantic enhancement module is used to perform part-of-speech weighted semantic enhancement processing on the purified text to generate text vectors; The phrase generalization extension module is used to perform semantic generalization extension based on a pre-built set of intent category phrases, generating a baseline vector for each intent category; The dual-dimensional vector matching module is used to calculate the dual-dimensional similarity between the text vector and the baseline vector of each intent category, and then weighted and fused to obtain the comprehensive matching score of each intent category. The intent category with the highest comprehensive matching score is determined as the final classification result.

8. The vehicle cabin user intent classification device according to claim 7, characterized in that, The semantic enhancement module is specifically used for: segmenting and tagging the purified text with words; setting differentiated weights for each word according to its part-of-speech category, wherein the weights of nouns and verbs are higher than those of adjectives and adverbs; mapping each word to a word vector; weighting and normalizing the vectors according to their corresponding part-of-speech weights; and generating the text vector.

9. The vehicle cockpit user intent classification device according to claim 7, characterized in that, The dual-dimensional vector matching module is also used to: if the original query text hits a preset prohibited word of a certain intent category, then block the highest score qualification of that category and extract the candidate category with the second highest score as the final classification result.

10. A vehicle, characterized in that, The system includes an in-vehicle voice interaction system and a processor, the processor being configured to perform the vehicle cockpit user intent classification method as described in any one of claims 1 to 6, to classify the intent of a query text input by a user through the in-vehicle voice interaction system, and to execute corresponding vehicle control commands or information feedback operations based on the classification results.

Images