Searching method and system for vehicle-mounted system and storage medium

Abstract

The embodiment of the invention provides a search method and system for a vehicle-mounted system.The search method comprises the following steps that a search instruction is responded, environment characteristics and user characteristics are obtained, and the search instruction comprises initial text characteristics; the environment characteristics are used for representing the real-time state of the vehicle; the user features are used for representing user historical behaviors; performing semantic generalization processing on the initial text features to generate a text feature set; fusing the text feature set, the environment feature and the user feature into a fusion vector; performing first recall based on the text feature set to generate a first candidate set; performing second recall based on the fusion vector to generate a second candidate set; according to a preset fusion weight strategy, the first candidate set and the second candidate set are fused and sorted, and a search set is generated.

Description

Technical Field

[0001] This application relates to the technical field of vehicle-mounted systems, and particularly relates to a search method, system, and storage medium for vehicle-mounted systems. Background Art

[0002] With the development of intelligent connected vehicles, vehicle-mounted systems have increasingly become an important interaction entry for users to obtain information such as navigation, media, and services. As one of the core functions, vehicle-mounted search directly affects the information acquisition efficiency and experience of users during driving. Most existing vehicle-mounted systems adopt a traditional information retrieval architecture based on keyword matching. A typical example is the Elasticsearch (ES) system based on inverted index, which combines classic statistical models such as BM25 and TF-IDF to achieve search ranking.

[0003] However, during the user's driving process, the attention is limited and the interaction window is short. There are problems such as low recall rate and poor result relevance when the traditional information retrieval architecture is applied to vehicle-mounted systems. Summary of the Invention

[0004] Embodiments of this application provide a search method, system, and storage medium for vehicle-mounted systems, which can improve the recall rate and result relevance of the search system for vehicle-mounted systems.

[0005] In a first aspect, an embodiment of this application provides a search method for vehicle-mounted systems, including the following steps: Respond to a search instruction, and obtain an environmental feature and a user feature. The search instruction includes: an initial text feature; the environmental feature is used to characterize the real-time state of the vehicle; the user feature is used to characterize the user's historical behavior; Perform semantic generalization processing on the initial text feature to generate a text feature set; Fuse the text feature set, the environmental feature, and the user feature into a fusion vector; Perform a first recall based on the text feature set to generate a first candidate set; Perform a second recall based on the fusion vector to generate a second candidate set; According to a preset fusion weight strategy, fuse and sort the first candidate set and the second candidate set to generate a search set.

[0006] In some of these embodiments, performing semantic generalization processing on the initial text feature to generate a text feature set includes: Perform intention recognition on the initial text feature through a prompt word model to generate intention information, where the intention information includes a basic intention and a topic intention; Perform synonym expansion on the basic intention to generate a first synonym set; Perform synonym expansion on the subject intention to generate a second synonym set; Generate the text feature set based on the first synonym set and the second synonym set.

[0007] In some embodiments, the performing synonym expansion on the subject intention to generate a second synonym set includes: Generate a recall prompt word according to the user feature and the subject intention; Input the recall prompt word into a preset expansion model to generate a second synonym set.

[0008] In some embodiments, the performing synonym expansion on the subject intention to generate a second synonym set includes: Input the subject intention into a preset expansion model to generate a general synonym set; Calculate the relevance between the synonyms in the general synonym set and the user feature; Filter out the second synonym set from the general synonym set based on the relevance; the relevance of the synonyms in the second synonym set is greater than that of other synonyms in the general synonym set.

[0009] In some embodiments, the performing a first recall based on the text feature set to generate a first candidate set includes: Limit the recall scope for the first synonym set; Perform keyword matching on the second synonym set within the recall scope to generate the first candidate set.

[0010] In some embodiments, the fusing the text feature set, the environment feature, and the user feature into a fusion vector includes: Convert the text feature set into a text feature vector, convert the environment feature into an environment feature vector, and convert the user feature into a user feature vector; Perform weighted fusion on the text feature vector, the environment feature vector, and the user feature vector to generate a fusion vector.

[0011] In some embodiments, the performing a second recall based on the fusion vector to generate a second candidate set includes: Calculate the similarity between the fusion vector and each resource vector in a pre-constructed resource vector library; Filter out the target resource vectors in the resource vector library whose similarity is greater than a preset similarity; Generate the second candidate set based on the resources corresponding to the target resource vectors.

[0012] In some of these embodiments, the environmental features include: vehicle state parameters, interaction timing parameters; the first candidate set contains multiple candidate resources, each candidate resource corresponding to a first score, and the second candidate set contains multiple candidate resources, each candidate resource corresponding to a second score; Fusing and sorting the first candidate set and the second candidate set according to a preset fusion weight strategy to generate a search set includes: Normalize the first scores of the candidate resources in the first candidate set and the second scores of the candidate resources in the second candidate set; Obtain the vehicle state parameters and interaction timing parameters, and confirm the current interaction reliability; When the interaction reliability is lower than a preset threshold, reduce the weights of the candidate resources in the first candidate set, increase the weights of the candidate resources in the second candidate set, and generate corresponding fusion scores based on the weights of the candidate resources in the first candidate set and the normalized first scores, and generate corresponding fusion scores based on the weights of the candidate resources in the second candidate set and the normalized second scores; or, When the interaction reliability is higher than or equal to the preset threshold, increase the weights of the candidate resources in the first candidate set, reduce the weights of the candidate resources in the second candidate set, and generate corresponding fusion scores based on the weights of the candidate resources in the first candidate set and the normalized first scores, and generate corresponding fusion scores based on the weights of the candidate resources in the second candidate set and the normalized second scores; Sort the candidate resources in the first candidate set and the second candidate set uniformly based on the fusion scores to generate the search set.

[0013] In a second aspect, this embodiment also provides a search system for an in-vehicle system, including: A response module, configured to respond to a search instruction, obtain environmental features and user features, where the search instruction includes: initial text features; the environmental features are used to characterize the real-time state of the vehicle; the user features are used to characterize the historical behavior of the user; A semantic processing module, configured to perform semantic generalization processing on the initial text features to generate a text feature set; A vector conversion module, configured to fuse the text feature set, the environmental features, and the user features into a fusion vector; A first recall module, configured to perform a first recall based on the text feature set to generate a first candidate set; A second recall module, configured to perform a second recall based on the fusion vector to generate a second candidate set; A fusion module, configured to fuse and sort the first candidate set and the second candidate set according to a preset fusion weight strategy, and generate a search set.

[0014] In a third aspect, the present embodiment further provides a computer-readable storage medium, on which a computer program is stored. The computer program is loaded by a processor to execute the steps in the search method for a vehicle-mounted system according to any one of the above embodiments. Advantageous effects

[0015] The search method for a vehicle-mounted system disclosed in this application proposes a dual-channel recall method. By performing semantic generalization processing on a search instruction, combining user historical features and vehicle current environment features, keyword recall and fusion vector recall are respectively performed, and a weight regulation mechanism based on interactive reliability scoring is introduced to perform weighted fusion sorting on the recall results, so as to output search results with high relevance, high robustness and high responsiveness in different driving scenarios. Description of the drawings

[0016] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application. For those skilled in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0017] Figure 1 is a flowchart of the search method for a vehicle-mounted system provided by some embodiments of the present application Figure 1 ; Figure 2 is a flowchart of the search method for a vehicle-mounted system provided by some embodiments of the present application Figure 2 ; Figure 3 is a flowchart of the search method for a vehicle-mounted system provided by some embodiments of the present application Figure 3 . Detailed implementation manners

[0018] The following will clearly and completely describe the technical solutions in the embodiments of the present application with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, rather than all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative efforts belong to the scope of protection of the present application.

[0019] In the description of the present application, it should be understood that the orientation or positional relationship indicated by the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc. is based on the orientation or positional relationship shown in the drawings, and is only for the convenience of describing the present application and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as a limitation to the present application. In addition, the terms "first" and "second" are only used for descriptive purposes and cannot be construed as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, the features defined with "first" and "second" may explicitly or implicitly include one or more of the said features. In the description of the present application, "a plurality of" means two or more, unless otherwise specifically defined.

[0020] "A and / or B" includes the following three combinations: only A, only B, and the combination of A and B.

[0021] The use of "suitable for" or "configured to" in the present application means open and inclusive language, which does not exclude a device that is suitable for or configured to perform additional tasks or steps. Additionally, the use of "based on" means open and inclusive because a process, step, calculation, or other action "based on" one or more of the said conditions or values can in practice be based on additional conditions or values beyond those stated.

[0022] In the present application, the term "exemplary" is used to mean "serving as an example, illustration, or explanation". Any embodiment described as "exemplary" in the present application is not necessarily to be construed as more preferred or advantageous than other embodiments. The following description is given to enable any person skilled in the art to make and use the present application. In the following description, details are set forth for purposes of explanation. It should be understood that those of ordinary skill in the art can recognize that the present application can be implemented without the use of these specific details. In other instances, well-known structures and processes are not elaborated in detail to avoid obscuring the description of the present application with unnecessary details. Therefore, the present application is not intended to be limited to the embodiments shown, but is to be accorded the widest scope consistent with the principles and features disclosed herein.

[0023] As described in the background art, there are technical bottlenecks in the application of the search architecture of traditional vehicle-mounted systems. Specifically, during the driving process, the user's attention is limited, so the search instructions issued are often inaccurate and have limited information. When the user's expression is not standardized, the instruction is colloquial, or there is no obvious match between the search term and the target content, the search architecture of the traditional vehicle-mounted system is prone to problems such as incomplete recall and irrelevant results. The above deficiencies not only seriously weaken the search experience but also distract the driver's attention and increase the driving risk.

[0024] Referring to Figure 1 , an embodiment of the present application provides a search method for a vehicle-mounted system, including: S100. Respond to the search instruction, obtain the environmental feature and the user feature. The search instruction includes: the initial text feature; the environmental feature is used to characterize the real-time state of the vehicle; the user feature is used to characterize the user's historical behavior; S200. Perform semantic generalization processing on the initial text feature to generate a text feature set; S300. Fuse the text feature set, the environmental feature, and the user feature into a fusion vector; S400. Perform the first recall based on the text feature set to generate a first candidate set; S500. Perform the second recall based on the fusion vector to generate a second candidate set; S600. According to the preset fusion weight strategy, fuse and sort the first candidate set and the second candidate set to generate a search set.

[0025] In step S100, respond to the search instruction, obtain the environmental feature and the user feature. The search instruction includes: the initial text feature; the environmental feature is used to characterize the real-time state of the vehicle; the user feature is used to characterize the user's historical behavior.

[0026] In the embodiment of the present application, the search instruction can be input in the form of text input or voice input. The search system of the vehicle-mounted system can uniformly parse and process different forms of input to obtain the initial text feature included in the search instruction.

[0027] After receiving and responding to a search instruction, the search system of the in-vehicle system can obtain environmental features and user features. Environmental features include, but are not limited to: vehicle state parameters, interaction timing parameters, time parameters, etc. Vehicle state parameters include, but are not limited to: current vehicle position information, current vehicle speed, current acceleration, etc. Time parameters include current time information. Interaction timing parameters include: peak interaction period and low interaction period; the peak interaction period is used to represent the stage when the user's attention is relatively concentrated and in an idle and interactive state. The low interaction period is used to represent the stage when the user's attention is limited and the driving task has a higher priority. For example, the peak interaction period can be: the initial stage of getting in the car (the first 0 - 30 seconds after the vehicle is just unlocked and started), when the vehicle is stationary (the vehicle is stationary for ≥ 10 seconds), etc.; the low interaction period can be: during driving. User features include, but are not limited to: user historical behavior records, user interest tags, user portrait information, historical geographical behavior characteristics, etc.

[0028] In step S200, semantic generalization processing is performed on the initial text features to generate a text feature set.

[0029] Refer to Figure 2 , in the embodiment of the present application, the text feature set is generated through the following steps. Specifically: S201. Use a prompt word model to perform intent recognition on the initial text features to generate intent information, where the intent information includes a basic intent and a topic intent; S202. Perform synonym expansion on the basic intent to generate a first synonym set; S203. Perform synonym expansion on the topic intent to generate a second synonym set; S204. Generate a text feature set based on the first synonym set and the second synonym set.

[0030] Specifically, the basic intent represents the type of operation the user intends to perform, usually a verb expression. For example, when the user inputs "I want to listen to the quiet of Zhou XX", the basic intent is "listen". After expanding the synonyms of "listen", the first synonym set includes, but is not limited to, ["play", "listen", "play some", "play it", "play a song"]. The thematic intent represents the main content the user is interested in, usually a noun or a phrase structure. For example, when the user inputs "I want to listen to the quiet of Zhou XX", the thematic intent is "the quiet of Zhou XX", and the second synonym set includes, but is not limited to, ["the quiet of Zhou XX", "quiet", "lyric", "quiet songs", "Zhou XX", "Chinese pop music"]. The generated text feature set can be a feature expression in a phrase structure generated by cross-combining the first synonym set and the second synonym set, such as ["play quiet", "play Zhou XX", "play lyric songs", "play some quiet", "listen to Zhou XX", "play some Chinese pop music"], or it can retain the structures of the first synonym set and the second synonym set respectively, or there can be text feature sets in both of the above forms at the same time.

[0031] It is worth mentioning that in the embodiments of the present application, the synonyms included in the first synonym set and the second synonym set have different granularities. For example, in the second synonym set, the fine-grained expression (closer to the user's specific expression and retaining the complete directivity) is "the quiet of Zhou XX"; the medium-grained expression (abstracting some semantic keywords) is "quiet", "lyric", "quiet songs", "Zhou XX"; the coarse-grained expression (generalizing to the semantic category or content style) is "Chinese pop music".

[0032] Furthermore, when generating the second synonym set, the second synonym set can also be jointly generated based on the user characteristics and the thematic intent. In some embodiments, the second synonym set is generated through the following steps: Generate a recall prompt word according to the user characteristics and the thematic intent; Input the recall prompt word into a preset expansion model to generate the second synonym set.

[0033] Specifically, when the initial text feature is "I want to listen to light music", the thematic intent is "light music", and among the user characteristics, the user interest tags are: jazz, piano, serenade, and the user's historical behavior records are: "Bill Evans", "café jazz", and the user portrait information is: 30 - 40 years old, then a recall prompt word [User preference: jazz, piano, serenade; Historical behavior: played Bill Evans and café jazz many times; Current search intent: light music; Please expand the keywords related to "light music" based on the above information] can be generated based on the above thematic intent and user characteristics, and the above recall prompt word is input into a preset expansion model to generate the second synonym set. The expansion model can be a lightweight generation model that supports input of prompt words.

[0034] In some other embodiments, a second synonym set is generated through the following steps: Input the theme intention into a preset expansion model to generate a general synonym set; Calculate the relevance between the synonyms in the general synonym set and the user characteristics; Based on the relevance, screen out the second synonym set from the general synonym set; the relevance of the synonyms in the second synonym set is greater than that of other synonyms in the general synonym set.

[0035] Specifically, through a preset expansion model (such as BERT, GPT, dictionary system), general synonym expansion is performed on "light music" to generate a general synonym set: ["relaxing music", "pure music", "background music", "piano music", "jazz", "serenade", "Lo-fi", "sleep melody"]; among the user characteristics, the user interest tags are: jazz, piano, serenade, and the user's historical behavior records are: "Bill Evans", "nighttime Lo-fi", and the user profile information is: 30 - 40 years old. For each general synonym, calculate its semantic relevance with the user interest tags and the user's historical behavior records (vector cosine similarity or rule-based scoring can be used). The second synonym set screened out is ["piano music", "jazz", "serenade", "Lo-fi"].

[0036] In the above two embodiments of generating the second synonym set, user characteristics are introduced into the synonym expansion link of the theme intention, which can improve the accuracy of the first candidate set during the first recall. At the same time, when the user inputs a search instruction, it often has ambiguity and incompleteness (such as "play something light"), or there are non-standard spoken expressions. Through the fusion of generalization and personalization, this solution can still infer and cover the true intention even when the user input is not standardized. For example: "Listen to light music" can be automatically associated with "jazz serenade", "piano relaxation music", "Lo-fi background", significantly improving the robustness of the recall and the consistency of the user experience.

[0037] In step S400, a first recall is performed based on the text feature set to generate a first candidate set. In the embodiments of the present application, the first candidate set is specifically generated through the following steps: Limit the recall range for the first synonym set; Perform keyword matching on the second synonym set within the recall range to generate a first candidate set.

[0038] Specifically, when the above-mentioned publicly generated text feature set is mentioned, it can be stated that the text feature set can respectively retain the structures of the first synonym set and the second synonym set. In this embodiment, the text feature set that retains the structures of the first synonym set and the second synonym set is utilized. First, the first synonym set generated by the basic intention recognition is used as the behavior limiting condition to screen the index fields or semantic tags in the resource library, thereby limiting the recall scope. Example: When the basic intention is "listen", and the first synonym set is ["play", "listen", "play some", "play it"], the system will only retrieve media resources (such as songs, albums, radio stations) when recalling resources, rather than navigation, question-and-answer, or service resources. Subsequently, within the above-mentioned limited recall scope, keyword matching is performed on the synonyms in the second synonym set to screen out candidate resources with high relevance to form the first candidate set. Example: The second synonym set is ["Zhou XX", "quiet", "lyric", "serenade", "jazz"], and based on the exact match or fuzzy match (Fuzzy) of the inverted index of ES (Elasticsearch), and using BM25, TF-IDF, or a customized scoring mechanism to score and rank the candidate resources, finally, the candidate resources with scores higher than the preset threshold will be included in the first candidate set for subsequent fusion ranking.

[0039] In step S300, the text feature set, environmental features, and user features are fused into a fusion vector; in step S500, a second recall is performed based on the fusion vector to generate a second candidate set. Specifically, in the embodiment of this application, the fusion vector is formed through the following steps: Convert the text feature set into a text feature vector, convert the environmental features into an environmental feature vector, and convert the user features into a user feature vector; Perform weighted fusion on the text feature vector, environmental feature vector, and user feature vector to generate a fusion vector.

[0040] Specifically, the text feature set can generate a corresponding text feature vector V-text through a preset text encoding model (such as lightweight semantic models like BERT, SBERT, SimCSE, etc.). The environmental features (including vehicle state parameters, interaction timing parameters, time parameters, etc.) can be encoded as numerical features and then input into an embedding model or use a multi-dimensional normalization method to generate an environmental feature vector V-env. The user features (including user historical behavior records, user interest tags, user portrait information, historical geographical behavior characteristics, etc.) can be converted into a vector representation with a fixed dimension, and the conversion methods include but are not limited to methods such as label vector mapping, behavior sequence embedding, and portrait feature encoding. The user feature vector is represented by V-user.

[0041] When performing weighted fusion on the text feature vector, environmental feature vector, and user feature vector to generate a fusion vector, the dimensions of the three vectors may be different. The text feature vector V-text may be 128-dimensional, the environmental feature vector V-env may be 64-dimensional, and the user feature vector V-user may be 64-dimensional. First, the three vectors need to be mapped to the same dimension through a linear transformation, and then weighted summation is performed using weights α, β, and γ to obtain the fusion vector V-fusion, where V-fusion = α * V-text + β * V-env + γ * V-user, and α, β, and γ can be preset fixed values (such as 0.6, 0.15, 0.25).

[0042] In the search system of traditional in-vehicle systems, relying solely on text vectors is likely to miss the true intentions of users. When environmental features are added during the second recall, it can cover ambiguous expressions or non-standard sentences. At the same time, different users enter the same text, but the search system can recall different resources according to their preferences. For example, for "play light music", some people like "piano music", while others like "Lo-fi serenade".

[0043] In the embodiments of the present application, the second candidate set is generated through the following steps: Calculate the similarity between the fusion vector and each resource vector in the pre-constructed resource vector library; Screen out the target resource vectors in the resource vector library whose similarity is greater than the preset similarity; Generate a second candidate set based on the resources corresponding to the target resource vectors.

[0044] In the pre-constructed resource vector library, various target resources in the resource library (such as songs, albums, news, radio programs, etc.) are uniformly encoded, and each resource content (including title, tags, description fields, etc.) is converted into a resource semantic vector with a fixed dimension. When calculating the similarity between the fusion vector and each resource vector in the pre-constructed resource vector library, cosine similarity can be used for calculation. Compared with the first recall, the second recall can capture deeper and implicit semantic associations through the vector expressions of the text features, environmental features, and user features after synonym expansion, and can effectively complement the retrieval results in cases where the user input is incomplete, keywords are missing, or the expression is ambiguous.

[0045] Refer to Figure 3 , in step S600, according to the preset fusion weight strategy, the first candidate set and the second candidate set are fused and sorted to generate a search collection.

[0046] Specifically, the first candidate set contains multiple candidate resources, and each candidate resource corresponds to a first score. The second candidate set contains multiple candidate resources, and each candidate resource corresponds to a second score. Embodiments of the present application generate a search set based on the following steps: S601. Normalize the first scores of the candidate resources in the first candidate set and the second scores of the candidate resources in the second candidate set; S602. Obtain vehicle state parameters and interaction timing parameters, and confirm the current interaction reliability; S603. When the interaction reliability is lower than a preset threshold, reduce the weights of the candidate resources in the first candidate set, increase the weights of the candidate resources in the second candidate set, and generate corresponding fusion scores based on the weights of the candidate resources in the first candidate set and the normalized first scores, and generate corresponding fusion scores based on the weights of the candidate resources in the second candidate set and the normalized second scores; or, S604. When the interaction reliability is higher than or equal to the preset threshold, increase the weights of the candidate resources in the first candidate set, reduce the weights of the candidate resources in the second candidate set, and generate corresponding fusion scores based on the weights of the candidate resources in the first candidate set and the normalized first scores, and generate corresponding fusion scores based on the weights of the candidate resources in the second candidate set and the normalized second scores; S605. Sort the candidate resources in the first candidate set and the second candidate set uniformly based on the fusion scores to generate a search set.

[0047] Specifically, the candidate resources in the first candidate set and the candidate resources in the second candidate set are recalled through different recall paths. By normalizing the first scores of the candidate resources in the first candidate set and the second scores of the candidate resources in the second candidate set respectively, the differences in the scoring scales in different recall paths can be eliminated, ensuring the effectiveness and comparability of subsequent weighted calculations. The normalization method can be min-max normalization, Z-score standardization, Softmax transformation, etc.

[0048] Furthermore, by obtaining vehicle state parameters and interaction timing parameters, they are used as interaction environment evaluation indicators. Vehicle state parameters include but are not limited to the current vehicle position information, current vehicle speed, current acceleration, etc.; interaction timing parameters include: interaction peak period and interaction low peak period; the interaction peak period is used to represent the stage when the user's attention is relatively concentrated and in an idle or interactive state. The interaction low peak period is used to represent the stage when the user's attention is limited and the driving task has a higher priority. For example, the interaction peak period can be: the initial stage of getting in the car (the first 0 - 30 seconds after the vehicle is just unlocked and started), when the vehicle is stationary (the vehicle is stationary for ≥ 10 seconds), etc.; the interaction low peak period can be: during driving.

[0049] Specifically, in the embodiments of the present application, the interactive reliability score is calculated according to a preset rule. Regarding the vehicle state parameters, the obtained current vehicle speed is as follows: stationary (0 km / h) corresponds to 1 point, low speed (<30 km / h) corresponds to 0.7 points, and high speed (≥60 km / h) corresponds to 0.2 points; the obtained current acceleration is as follows: close to 0 (indicating that the vehicle is running smoothly) corresponds to 1 point, and frequent speed changes (indicating that the vehicle may be accelerating or braking suddenly) corresponds to 0.4 points. Through linear weighting, the weight of the current vehicle speed is 0.6, and the weight of the current acceleration is 0.4. The score R-vehicle for evaluating the interactive reliability of the vehicle state parameters can be determined by R-vehicle = 0.6 × f(speed) + 0.4 × f(acceleration), where f(speed) represents the score of the current vehicle speed and f(acceleration) represents the score of the current acceleration. Regarding the interaction timing parameters, the peak interaction period corresponds to 1 point, and the off-peak interaction period corresponds to 0.5 points. The comprehensive score R-interact of the current interactive reliability can be determined by R-interact = 0.5*R-vehicle + 0.5*R-context, where R-context represents the score of the interaction timing parameters. In some embodiments, the preset threshold is 0.6. When the interactive reliability score is lower than the preset threshold, the system believes that the initial text features may be incomplete or interfered by noise. Therefore, the dependence on the first candidate set obtained by the first recall should be weakened, and the proportion of the second candidate set obtained by the second recall should be increased. At this time, the weight of the first candidate set is lower than that of the second candidate set. When the interactive reliability score is higher than or equal to the preset threshold, the system gives priority to trusting the first candidate set obtained by the first recall, enhances the influence of the first candidate set, and appropriately weakens the inference recall of the second candidate set. At this time, the weight of the first candidate set is higher than that of the second candidate set.

[0050] The first candidate set is recalled through keyword matching, with fast response and high accuracy, but it requires high clarity of user expression. For example, the first candidate set has a high hit rate for expressions like "listen to Zhou XX", but a low hit rate for expressions like "I want something comfortable". The second candidate set is recalled through fused vectors, which can adapt to fuzzy expressions and has strong semantic reasoning ability, but there may be insufficient accuracy or an overly wide recall range. In an actual in-vehicle scenario, the quality of the search results fluctuates due to factors such as driving state, voice interference, and time. In the embodiments of the present application, by adjusting the weights of the first candidate set and the second candidate set during fusion, when the user's expression is clear and the system has stable interaction (such as in a parking state), the weight of the first candidate set can be enhanced to improve the accuracy and intuitive hit rate of the recall results. When the user's voice input is incomplete, the expression is fuzzy, or there is misrecognition (such as on a highway), the weight of the second candidate set is increased, enabling the system to better "understand" the user's potential intention, thereby reducing the situation of no results or incorrect results. At the same time, in the embodiments of the present application, when the interaction reliability is relatively high, lightweight keyword recall is preferentially invoked, which can reduce the average response latency and improve the interaction fluency and system response efficiency.

[0051] In some embodiments, before sorting the candidate resources in the first candidate set and the second candidate set based on the fusion score to generate a search collection in step S605, a similarity matching judgment is also made on the candidate resources in the first candidate set and the second candidate set. When it is determined that there is the same candidate resource, the candidate resource is regarded as a duplicate resource, and only one result is retained, and the scores of the candidate resource in the two recall paths are fused. The scores of the fused duplicate candidate resources in the two recall paths can be calculated by weighted summation, and a redundant matching weighting term can be introduced to strengthen its correlation expression.

[0052] For example, the resource "Zhou XX · Quiet" is hit in the first candidate set, with a fusion score of 0.81; the same resource is also hit in the second candidate set, with a fusion score of 0.74. Based on the weights of the fusion scores, the fusion scores of the candidate resource can be fused, and the fusion score of the candidate resource is recalculated. For example, in the fusion weight of the current fusion score, the weight of the candidate resource in the first candidate set is 0.6, the weight of the candidate resource in the first candidate set is 0.4, and the redundant matching weighting term = 0.03. The recalculated fusion score score-fused of the candidate resource is score_fused = 0.6×0.81 + 0.4×0.74 + 0.03 ≈ 0.786.

[0053] In summary, the search method for in-vehicle systems disclosed in this application proposes a dual-channel recall method. By performing semantic generalization on the search instruction, combining the user's historical features and the vehicle's current environmental features, keyword recall and fusion vector recall are respectively performed, and a weight regulation mechanism based on interactive reliability scoring is introduced to perform weighted fusion sorting on the recall results, so as to output search results with high relevance, high robustness, and high responsiveness in different driving scenarios.

[0054] An embodiment of this application also discloses a search system for in-vehicle systems, including: a response module, a semantic processing module, a vector conversion module, a first recall module, a second recall module, and a fusion module.

[0055] The response module is used to respond to the search instruction and obtain the environmental features and user features. The search instruction includes: initial text features; the environmental features are used to characterize the real-time state of the vehicle; the user features are used to characterize the user's historical behavior. The semantic processing module is used to perform semantic generalization on the initial text features to generate a text feature set. The vector conversion module is used to fuse the text feature set, environmental features, and user features into a fusion vector. The first recall module is used to perform the first recall based on the text feature set to generate a first candidate set. The second recall module is used to perform the second recall based on the fusion vector to generate a second candidate set. The fusion module is used to fuse and sort the first candidate set and the second candidate set according to a preset fusion weight strategy to generate a search set.

[0056] An embodiment of this application also discloses a computer-readable storage medium, on which a computer program is stored. The computer program is loaded by a processor to execute the steps in the search method for in-vehicle systems described in any one of the above. In the embodiment of this application, the storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM), etc.

[0057] In the above embodiments, the descriptions of the respective embodiments have their own emphases. For parts not detailed in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0058] The above has introduced in detail a washing machine and its control method, control system, and computer-readable storage medium provided by the embodiments of this application. Specific examples are used in this article to elaborate on the principle and implementation manner of this application. The descriptions of the above embodiments are only used to help understand the method and its core idea of this application; at the same time, for those skilled in the art, according to the idea of this application, there will be changes in the specific implementation manner and application scope. In summary, the content of this specification should not be construed as a limitation to this application.

Claims

1. A search method for an in-vehicle system, characterized in that, Including: In response to a search instruction, obtain environmental features and user features. The search instruction includes: initial text features; the environmental features are used to characterize the real-time state of the vehicle; the user features are used to characterize the user's historical behavior. Perform semantic generalization processing on the initial text features to generate a text feature set. Fuse the text feature set, the environmental features, and the user features into a fusion vector. Perform a first recall based on the text feature set to generate a first candidate set. Perform a second recall based on the fusion vector to generate a second candidate set. According to a preset fusion weight strategy, fuse and sort the first candidate set and the second candidate set to generate a search set.

2. The search method for an in-vehicle system according to claim 1, wherein The performing semantic generalization processing on the initial text features to generate a text feature set includes: Use a prompt model to perform intent recognition on the initial text features to generate intent information, where the intent information includes a basic intent and a topic intent. Perform synonym expansion on the basic intent to generate a first synonym set. Perform synonym expansion on the topic intent to generate a second synonym set. Generate the text feature set based on the first synonym set and the second synonym set.

3. The search method for a vehicle-mounted system according to claim 2, wherein, The performing synonym expansion on the topic intent to generate a second synonym set includes: Generate a recall prompt word according to the user features and the topic intent. Input the recall prompt word into a preset expansion model to generate a second synonym set.

4. The search method for an in-vehicle system according to claim 2, wherein The performing synonym expansion on the topic intent to generate a second synonym set includes: Input the topic intent into a preset expansion model to generate a general synonym set. Calculate the relevance between the synonyms in the general synonym set and the user features. Filter out the second synonym set from the general synonym set based on the relevance; the relevance of the synonyms in the second synonym set is greater than the relevance of other synonyms in the general synonym set.

5. The search method for a vehicle-mounted system according to claim 2, wherein The performing a first recall based on the text feature set to generate a first candidate set includes: Limit the recall scope for the first synonym set. Perform keyword matching on the second synonym set within the recall scope to generate the first candidate set.

6. The search method for an in-vehicle system according to claim 1, wherein The fusing the text feature set, the environmental features, and the user features into a fusion vector includes: Convert the text feature set into a text feature vector, convert the environmental features into an environmental feature vector, and convert the user features into a user feature vector. Perform weighted fusion on the text feature vector, the environmental feature vector, and the user feature vector to generate a fusion vector.

7. The search method for an in-vehicle system according to claim 6, wherein The performing a second recall based on the fusion vector to generate a second candidate set includes: Calculate the similarity between the fusion vector and each resource vector in a pre-constructed resource vector library. Filter out target resource vectors in the resource vector library whose similarity is greater than a preset similarity. Generate the second candidate set based on the resources corresponding to the target resource vectors.

8. The search method for an in-vehicle system according to claim 1, wherein The environmental features include: vehicle state parameters, interaction timing parameters; the first candidate set contains multiple candidate resources, each candidate resource corresponding to a first score, and the second candidate set contains multiple candidate resources, each candidate resource corresponding to a second score; The fusing and sorting of the first candidate set and the second candidate set according to a preset fusing weight strategy to generate a search collection includes: Normalizing the first scores of the candidate resources in the first candidate set and the second scores of the candidate resources in the second candidate set; Obtaining the vehicle state parameters and interaction timing parameters, and confirming the current interaction reliability; When the interaction reliability is lower than a preset threshold, reducing the weights of the candidate resources in the first candidate set, increasing the weights of the candidate resources in the second candidate set, and generating corresponding fusing scores based on the weights of the candidate resources in the first candidate set and the normalized first scores, and generating corresponding fusing scores based on the weights of the candidate resources in the second candidate set and the normalized second scores; or, When the interaction reliability is higher than or equal to the preset threshold, increasing the weights of the candidate resources in the first candidate set, reducing the weights of the candidate resources in the second candidate set, and generating corresponding fusing scores based on the weights of the candidate resources in the first candidate set and the normalized first scores, and generating corresponding fusing scores based on the weights of the candidate resources in the second candidate set and the normalized second scores; Sorting the candidate resources in the first candidate set and the second candidate set uniformly based on the fusing scores to generate the search collection.

9. A search system for a vehicle-mounted system, characterized in that, It includes: A response module, configured to respond to a search instruction, obtain environmental features and user features, the search instruction including: initial text features; the environmental features are used to characterize the real-time state of the vehicle; the user features are used to characterize the historical behavior of the user; A semantic processing module, configured to perform semantic generalization processing on the initial text features to generate a text feature collection; A vector conversion module, configured to fuse the text feature collection, the environmental features and the user features into a fused vector; A first recall module, configured to perform a first recall based on the text feature collection to generate a first candidate set; A second recall module, configured to perform a second recall based on the fused vector to generate a second candidate set; A fusing module, configured to fuse and sort the first candidate set and the second candidate set according to a preset fusing weight strategy to generate a search collection.

10. A computer-readable storage medium, characterized in that, A computer program is stored thereon, and the computer program is loaded by a processor to execute the steps in the search method for an in-vehicle system according to any one of claims 1-8.

Images