Driver intention identification method, system, equipment and medium

Abstract

The invention relates to a driver intention identification method, system and device and a medium. The method comprises the steps of obtaining multiple recruitment communication records of a driver and active behaviors of the driver in a recruitment platform within a preset duration; wherein the recruitment communication record comprises single-party dialogue content between the driver and the recruitment platform and a corresponding time identifier; inputting each recruitment communication record into an abstract extraction model to obtain a dialogue abstract of each recruitment communication record; wherein the abstract extraction model is a natural language processing model; sorting the dialogue abstracts of each time according to respective time sequence, and combining the active behaviors and the sorted dialogue abstracts to form a synthetic text; and inputting the synthesized text into the large model for intention recognition, and recognizing the intention of the driver to join the recruitment platform. The abstract is extracted from multiple rounds of recruitment communication records of the driver, the synthetic text is generated in combination with the active behavior information, and the intention of the driver to join the recruitment platform is recognized more accurately.

Description

Technical Field

[0001] The present invention relates to the technical field of intention recognition, and in particular to a method, system, device and medium for recognizing driver's intention. Background Art

[0002] In the driver recruitment process, the recruitment platform needs to screen out drivers who meet the requirements, and how to improve the quality of driver leads and recruitment efficiency is a technical difficulty. The driver recruitment method usually relies on the inviter to invite the driver by phone and identify the driver's intention through conversation analysis. At present, the conversation analysis method is usually based on rule matching or shallow natural language processing methods, but due to the length limitation of text processing technology, the current conversation analysis method is often unable to process a large number of or long conversation records, resulting in the user's intention to rely on the analysis of a single round of conversation. Not only is the recruitment efficiency low, but it is also impossible to accurately and comprehensively identify the driver's potential intentions, resulting in poor quality of driver leads. Therefore, it is necessary to provide a method, system, device and medium for identifying driver intentions. Summary of the invention

[0003] In view of the above-mentioned shortcomings of the prior art, an object of the present invention is to provide a method, system, device and medium for identifying driver intention, which improves the technical problem of low accuracy and efficiency in the prior art when identifying driver intention.

[0004] To achieve the above-mentioned purpose and other related purposes, the present invention provides a method for identifying driver intentions, comprising: obtaining multiple recruitment communication records of a driver and the driver's active behavior on the recruitment platform within a preset time period; wherein the recruitment communication record includes the unilateral conversation content between the driver and the recruitment platform and its corresponding time stamp; inputting each recruitment communication record into a summary extraction model to obtain a conversation summary of each recruitment communication record; wherein the summary extraction model is a natural language processing model; sorting the conversation summaries of each time according to their respective chronological order, and merging the active behavior and the sorted conversation summaries to form a synthetic text; inputting the synthetic text into a large model for intention recognition to identify the driver's intention to join the recruitment platform.

[0005] In one embodiment of the present invention, the summary extraction model is a T5 model. For each recruitment communication record, the recruitment communication record is input into the summary extraction model to obtain a conversation summary of the recruitment communication record, including: preprocessing the recruitment communication record; inputting the preprocessed recruitment communication record into the encoder network of the T5 model, extracting features layer by layer based on multiple cascaded self-attention layers to obtain semantic features of the recruitment communication record; inputting the semantic features into the decoder network of the T5 model, gradually generating text information based on an autoregressive mechanism, and when a preset end condition is met, arranging the text information generated each time in sequence to generate a conversation summary.

[0006] In one embodiment of the present invention, the semantic feature is input into the decoder network of the T5 model, and text information is gradually generated based on the autoregressive mechanism. When the preset end condition is met, the text information generated each time is arranged in sequence to generate a conversation summary, including: extracting the first semantic feature vector of the semantic feature; inputting the extracted semantic feature vector into the decoder network, decoding it based on the self-attention mechanism, and obtaining the text information corresponding to the extracted semantic feature vector; wherein, based on the autoregressive mechanism, when the next semantic feature vector is decoded, the decoded text information is also input into the decoder network to participate in decoding; judging whether the text information corresponding to the extracted semantic feature vector contains the preset end information: if so, the text information generated by the semantic feature is arranged according to the time order of generation to form a conversation summary; otherwise, the next semantic feature vector is extracted from the semantic feature in sequence, and input into the decoder network again for decoding.

[0007] In one embodiment of the present invention, the summary extraction model is a BERT model. For each recruitment communication record, the recruitment communication record is input into the summary extraction model to obtain a conversation summary of the recruitment communication record, including: preprocessing the recruitment communication record; encoding each conversation sentence in the preprocessed recruitment communication record to obtain a sentence-level vector for each conversation sentence; performing average pooling processing on the sentence-level vectors of all conversation sentences to obtain a text vector; calculating the similarity between each sentence-level vector and the text vector, and screening out the sentence-level vector with the largest similarity, and using the conversation sentence corresponding to the sentence-level vector as the conversation summary.

[0008] In one embodiment of the present invention, the preprocessing of the recruitment communication records includes: text cleaning of the recruitment communication records; unifying the time tags in the cleaned recruitment communication records into a preset format to obtain preprocessed recruitment communication records.

[0009] In one embodiment of the present invention, the merging of active behaviors and sorted conversation summaries to form a synthetic text includes: converting active behaviors into standardized text features based on a preset format specification; and merging the standardized text features and the sorted conversation summaries to form a synthetic text.

[0010] In one embodiment of the present invention, the synthesized text is input into a large model for intent recognition. After the driver's intention to join the recruitment platform is identified, the method further includes: predicting the probability of the driver joining the recruitment platform and the corresponding reason analysis.

[0011] In one embodiment of the present invention, a driver intention recognition system is also provided, the system comprising: an information acquisition module, used to obtain multiple recruitment communication records of the driver and the driver's active behavior on the recruitment platform within a preset time period; wherein the recruitment communication record includes the unilateral conversation content between the driver and the recruitment platform and its corresponding time stamp; a summary generation module, used to input each recruitment communication record into a summary extraction model to obtain a conversation summary of each recruitment communication record; wherein the summary extraction model is a natural language processing model; a merging module, used to sort the conversation summaries of each time according to their respective time sequence, and merge the active behavior and the sorted conversation summaries to form a synthetic text; an intention recognition module, used to input the synthetic text into a large model for intention recognition, and identify the driver's intention to be recruited.

[0012] In one embodiment of the present invention, an electronic device is also provided, comprising: one or more processors; a storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the electronic device implements any of the above-mentioned methods for identifying the driver's intention.

[0013] In one embodiment of the present invention, a computer-readable storage medium is further provided, on which a computer program is stored. When the computer program is executed by a processor of a computer, the computer executes any of the above-mentioned methods for identifying the driver's intention.

[0014] As described above, the driver intention recognition method, system, device and medium of the present invention have the following beneficial effects: by extracting conversation summaries from the driver's multiple rounds of recruitment communication records, the amount of data that needs to be analyzed by the model can be effectively reduced, and the interference of noise in the processing process can be reduced, thereby greatly improving the processing rate and accuracy. The central idea of ​​each recruitment communication record is extracted in the form of a summary, and a synthetic text is generated in combination with its active behavior information. The intention recognition of the synthetic text is performed through a large model, and the driver's intention can be more accurately identified. The present invention combines the driver's historical data and recent active behavior for comprehensive analysis, and can more accurately and quickly predict the driver's intention to join the platform. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 A schematic flow chart of a method for identifying driver intention provided by an embodiment of the present invention;

[0016] Figure 2 Shown is a structural block diagram of a driver intention recognition system provided by an embodiment of the present invention;

[0017] Figure 3 Shown is a structural schematic diagram of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0018] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict.

[0019] It should be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present invention, and thus the drawings only show components related to the present invention rather than being drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component may be changed arbitrarily, and the component layout may also be more complicated.

[0020] In the following description, numerous details are discussed to provide a more thorough explanation of the embodiments of the present invention. However, it is obvious to those skilled in the art that the embodiments of the present invention can be implemented without these specific details. In other embodiments, well-known structures and devices are shown in the form of block diagrams rather than in detail to avoid making the embodiments of the present invention difficult to understand.

[0021] The inventors found that when recruiting drivers, the current recruitment platform mainly analyzes single or multiple conversations between the inviter and the user, identifies the user's intentional expression, and selects high-quality clues based on the user's discrete features. However, there are the following disadvantages: on the one hand, due to the length limitation of text processing technology, existing solutions are often unable to handle large or long conversation records, resulting in only relying on the analysis of a single round of conversations to judge the user's intentions, or having to rely on a rule system, which reduces the accuracy of intention analysis. On the other hand, the practice of sacrificing the recall rate in order to improve the quality of leads has resulted in insufficient recall of effective leads, which is not conducive to improving recruitment efficiency. In addition, traditional solutions use natural language processing technology combined with discrete machine learning features to train models for prediction. This model is often limited by feature quality and unexplainability, resulting in low business acceptance and poor results.

[0022] The present invention provides a method for identifying driver intentions. By extracting conversation summaries from the driver's multiple rounds of recruitment communication records, the amount of data that needs to be analyzed by the model can be effectively reduced, and the interference of noise during the processing process can be reduced, thereby greatly improving the processing rate and accuracy. The central idea of ​​each recruitment communication record is extracted in the form of a summary, and a synthetic text is generated in combination with its active behavior information. The intention of the synthetic text is performed through a large model to identify the intention, which can more accurately identify the driver's intention. The present invention extracts summaries from the driver's multiple rounds of recruitment communication records, and generates synthetic text in combination with active behavior information, which not only effectively reduces the amount of data and noise interference, but also improves processing efficiency and accuracy, thereby more accurately identifying the driver's intention to join the recruitment platform.

[0023] See also Figure 1 ,The driver intention recognition method includes the following steps:

[0024] S1. Acquire multiple recruitment communication records of the driver and the driver's active behavior on the recruitment platform within a preset time period; wherein the recruitment communication records include the unilateral conversation content between the driver and the recruitment platform and its corresponding time stamp.

[0025] Recruitment communication records refer to the interactive information between the driver and the recruitment platform, where the recruitment communication records include but are not limited to call records, text chat records or voice chat records. Specifically, if the platform wants to recruit a driver to its company, the platform's inviter will communicate with the driver, and the communication information will be saved in the platform's server as a recruitment communication record, and each recruitment communication record includes the driver ID, the event of the conversation and the content of the conversation. It is understandable that, considering the data calculation rate, it is not possible to obtain multiple recruitment communication records without limit, but it is possible to obtain a preset number of recruitment communication records, or recruitment communication records within a preset time period, and determine whether the driver is willing to join the recruitment platform by analyzing historical data over a period of time.

[0026] Furthermore, considering that only the driver's one-way conversation will be the object of intent analysis during the recruitment process, in another embodiment of the present invention, the process of obtaining the recruitment communication record includes: obtaining the initial communication record; wherein the initial communication record is a two-way conversation record between the driver and the recruitment platform; and filtering out the driver's conversation content from the initial communication record as the recruitment communication record. Specifically, since the initial communication record also records the driver ID and the ID of the inviter of the recruitment platform, the entire content of the driver ID and the corresponding timestamp can be extracted from it, and all the one-way conversation content of the driver during a conversation and the timestamp of each time when no one is there are saved as a recruitment communication record.

[0027] It should be noted that when the recruitment communication record is a call record, the call recording between the driver and the recruitment platform can be saved first, and the recording can be converted into text through a speech recognition tool. A corresponding timestamp can be added to each sentence, and the speaker can be distinguished, thereby forming an initial recruitment communication record, which can be used to extract the driver's unilateral conversation content and time identification information later.

[0028] S2. Input each recruitment communication record into a summary extraction model to obtain a conversation summary of each recruitment communication record; wherein the summary extraction model is a natural language processing model.

[0029] Since each recruitment communication record may include multiple rounds of dialogue, which also involves some meaningless information, it is time-consuming and laborious to directly analyze such recruitment communication records manually, and the analysis results are highly subjective; if such recruitment communication records are analyzed by machine, due to the large amount of noise in them, the robustness of the model is poor, resulting in poor final recognition effect. In order to improve this problem, the present invention achieves a more efficient analysis by extracting summaries from recruitment communication records. Specifically, for each acquired recruitment communication record: the present invention inputs each recruitment communication into a pre-trained summary extraction model, analyzes the deep semantic information of the recruitment communication record, pays attention to the correlation between each input word and other words, thereby extracting representative key information and generating a dialogue summary that can characterize the content of the entire recruitment communication record. This dialogue summary can succinctly and effectively characterize the main content of the entire recruitment communication record, and retains the time sensitivity of the driver's intention expression, which is convenient for subsequent analysis. It can be understood that the summary extraction model includes but is not limited to the T5 model, the BERT model, the BART model, etc., as long as it can process context information in the form of long text.

[0030] In one embodiment of the present invention, the summary extraction model is a T5 model. For each recruitment communication record, the recruitment communication record is input into the summary extraction model to obtain a dialogue summary of the recruitment communication record, including:

[0031] Pre-process the recruitment communication records;

[0032] Input the preprocessed recruitment communication records into the encoder network of the T5 model, extract features layer by layer based on multiple cascaded self-attention layers, and obtain semantic features of the recruitment communication records;

[0033] The semantic features are input into the decoder network of the T5 model, and text information is gradually generated based on the autoregressive mechanism. When the preset end conditions are met, the text information generated each time is arranged in sequence to generate a conversation summary.

[0034] For each recruitment communication record that needs to be analyzed, the following processing is performed: Considering that there is a lot of noise information in the record, in order to improve the robustness of the model, the recruitment communication record needs to be preprocessed first. Since the T5 model can effectively process multi-round dialogue information, has excellent performance in text generation, and can accurately extract deep semantic information, the T5 model is used as the summary extraction model in this embodiment. The T5 model includes an encoder and a decoder. The preprocessed recruitment communication record is input into the encoder network of the T5 model, and the preprocessed recruitment communication record is word segmented and word embedded to generate the corresponding vectorized recruitment communication record. The vectorized recruitment communication record is input into multiple cascaded self-attention layers. Each layer calculates the relationship between the current word and the remaining words based on the self-attention mechanism for the input information, and dynamically adjusts the weight value of each word, so as to capture the semantic information more accurately. The output of each layer is used as the input of the next layer. Through this layer-by-layer transmission processing method, the deep semantic features of the recruitment communication record are gradually extracted. This semantic feature can characterize the key information in the input recruitment communication record, so that the subsequent decoder can further generate the content of the dialogue summary. The deep semantic features generated by the last self-attention layer of the encoder are input into the decoder. Based on the autoregressive mechanism, each time text information is generated, the previously generated text information is referenced. Based on the previously generated text information, the next most likely word is predicted and used as the new input to continue generating the next word. When the decoder generates an end symbol, it means that the conversation summary has been generated.

[0035] It can be understood that the semantic feature includes a plurality of semantic feature vectors arranged in sequence, and each semantic feature vector is arranged according to the order of extraction to form a semantic feature. Each semantic feature vector represents certain semantic information in the recruitment communication record. For example, the recruitment communication record is "I think the membership fee is too high, but I am worried that the number of orders is small and I can't get any orders. I need to think about it again." Then the first semantic feature vector can be the concept representing "the membership fee is too high", the second semantic feature vector can be the concept representing "the number of orders is small and I can't get any orders", and the third semantic feature vector can be the concept representing "think about it again". These three semantic feature vectors are arranged in sequence to form a set, which is a semantic feature.

[0036] In one embodiment of the present invention, the semantic features are input into the decoder network of the T5 model, and text information is gradually generated based on an autoregressive mechanism, and when a preset end condition is met, the text information generated each time is arranged in sequence to generate a conversation summary, including:

[0037] Extracting a first semantic feature vector of the semantic feature;

[0038] The extracted semantic feature vector is input into the decoder network, and is decoded based on the self-attention mechanism to obtain text information corresponding to the extracted semantic feature vector; wherein, based on the autoregressive mechanism, when the next semantic feature vector is decoded, the decoded text information is also input into the decoder network to participate in decoding;

[0039] Determine whether the text information corresponding to the extracted semantic feature vector contains the preset end information:

[0040] If so, the text information generated by the semantic features is arranged in a chronological order to form a conversation summary;

[0041] Otherwise, the next semantic feature vector is sequentially extracted from the semantic features and input into the decoder network again for decoding.

[0042] When the decoder generates a dialogue summary, it first extracts the first semantic feature vector from the semantic features and combines it with a preset start symbol (such as <start>) are input into the decoder together to generate the text information corresponding to the semantic feature vector. Among them, the text information represents the words representing the semantic feature vector. Then, based on the autoregressive mechanism, all the current and previously generated text information and the next semantic feature vector are input into the decoder together. The decoder is based on the self-attention mechanism and decodes the input semantic feature vector by paying attention to the contextual information, and predicts the text information corresponding to the input semantic feature vector. In this way, each generated text information will be input into the decoder and participate in the decoding as input, forming an autoregressive generation process. It should be noted that after each text information is generated, it will be determined whether the currently generated text information is a preset end symbol (such as <end>), if the terminator is generated, it means that the entire conversation summary has been generated. At this point, the decoder stops the generation process and arranges the previously generated text information in the order of generation to form a complete conversation summary. On the contrary, if the currently generated text information is not a terminator, it means that the summary has not been generated yet, and the decoder will continue to decode the next semantic feature vector until the terminator is generated.

[0043] In one embodiment of the present invention, the summary extraction model is a BERT model. For each recruitment communication record, the recruitment communication record is input into the summary extraction model to obtain a conversation summary of the recruitment communication record, including:

[0044] Pre-process the recruitment communication records;

[0045] Encode each conversation sentence in the preprocessed recruitment communication record to obtain a sentence-level vector for each conversation sentence;

[0046] Perform average pooling on the sentence-level vectors of all dialogue sentences to obtain text vectors;

[0047] The similarity between each sentence-level vector and the text vector is calculated, and the sentence-level vector with the greatest similarity is screened out, and the conversation sentence corresponding to the sentence-level vector is used as the conversation summary.

[0048] Considering that the BERT model has powerful context modeling capabilities and the advantages of pre-training tasks, it is particularly suitable for processing complex natural language tasks. Therefore, in this embodiment, the summary extraction model is a BERT model. Each time the recruitment communication record is processed, it needs to be pre-processed to prevent the model from being disturbed by irrelevant items. The pre-processed recruitment communication record is input into the BERT model for encoding. The BERT model captures the semantic relationship between each word and the context based on the self-attention mechanism, and converts each dialogue sentence into a corresponding sentence-level vector. This sentence-level vector is a high-dimensional embedding representation that contains the deep semantic information of the sentence, which can provide richer corpus support for subsequent summary generation, so as to generate a more accurate dialogue summary. After obtaining the sentence-level vectors of all dialogues in the recruitment communication record, these vectors are averaged and pooled to obtain a text vector. Among them, the text vector represents the global semantic information of the entire recruitment communication record, and can cover the key content of all dialogue sentences to the maximum extent. By calculating the similarity between each sentence-level vector and the text vector, this similarity can represent the degree of matching between each sentence and the entire dialogue record. The sentence-level vector with the highest similarity is selected, and the corresponding dialogue sentence is used as the dialogue summary of the entire recruitment communication record. As the most relevant text in the entire recruitment communication record, this conversation summary can effectively reflect the core content of the conversation and facilitate the subsequent accurate identification of the driver's intention. It can be understood that the similarity calculation method includes but is not limited to cosine similarity, Euclidean distance, Manhattan distance, etc. Preferably, in order to better implement text measurement, this implementation uses cosine similarity.

[0049] In one embodiment of the present invention, the pre-processing of the recruitment communication records includes:

[0050] The recruitment communication records were cleaned textually;

[0051] In the recruitment communication records after the text cleaning, the time tags therein are unified into a preset format to obtain the pre-processed recruitment communication records.

[0052] The recruitment communication records to be processed are cleaned to remove irrelevant interference information and redundant content to make the text more concise. Specifically, the text cleaning methods include but are not limited to removing irrelevant characters, deleting extra spaces and line breaks, correcting spelling errors, etc. Among them, irrelevant characters include but are not limited to special symbols, emoticons, etc. Considering that drivers may have different ways of expressing the same time, such as "May 3", "5-03", "May 3", etc., in order to ensure that the subsequent model can accurately recognize this information, it is necessary to convert the time mark into the same preset standard format, such as "YYYY-MM-NN", so as to obtain the pre-processed recruitment communication records.

[0053] S3. Sort the conversation summaries of each session according to their respective chronological order, and merge the active behaviors and the sorted conversation summaries to form a synthetic text.

[0054] After extracting the summary of each conversation through the summary extraction model, the summary of each conversation is sorted according to the time sequence of the actual conversation to obtain the conversation summary sequence. As the driver's historical data, the conversation summary sequence is a complete historical conversation context in a certain period of time in the past, which can analyze the driver's intention to join the platform from one aspect. In addition, in order to more accurately identify the driver's intention, the driver's recent operation data on the platform (i.e., active behavior) will also be extracted. Among them, active behavior includes but is not limited to the number of times the driver logs into the recruitment platform, the active duration, the most recent active time, etc. This information is formed into a text form and spliced ​​with the conversation summary sequence to form a synthetic text. The synthetic text not only integrates the driver's historical conversation content, but also combines his recent active behavior, thereby providing more comprehensive data support for subsequent intention recognition.

[0055] In one embodiment of the present invention, merging the active behavior and the sorted conversation summaries to form a synthetic text includes:

[0056] Based on the preset format specifications, active behaviors are converted into standardized text features;

[0057] The standardized text features and the sorted conversation summaries are combined to form a synthetic text.

[0058] In order to unify the format of active behavior data and conversation summaries for subsequent processing, it is first necessary to convert the active behavior data according to the preset format specifications to generate standardized text features. Specifically, for each active behavior, a format specification is pre-defined, and the extracted active behavior is converted according to the corresponding format specification. After all active behaviors are converted, the corresponding standardized text features are obtained. Exemplarily, the format specification of the driver's last active time is "last active at [time]". After all format specifications are converted, the standardized text features finally generated are "The driver was last active on 2023-11-6 18:31:37, logged into the APP 10 times in the past week, and the total active time in a week was 54 minutes". The standardized text features and the sorted conversation summary are spliced ​​together to generate a synthetic text, which can provide more comprehensive contextual information for subsequent intent recognition.

[0059] As an example, there are three historical conversation summaries extracted, and the resulting conversation summary sequence is "[2023-11-1 10:10:20: The driver said he was driving and it was inconvenient to answer the call; 2023-11-2 13:30:52: The driver felt that the membership fee was too high, and was worried that the number of orders was small and he could not get any orders, so he needed to reconsider; 2023-11-3 15:30:37: The driver asked about the insurance fee and felt it was too high, so he would reconsider]", and the standardized text feature is "The driver was last active on 2023-11-6 18:31:37, logged into the APP 10 times in the past week, and was active for a total of 54 minutes a week". After splicing the two, the generated synthetic text is "[2023-11-1 10:10:20: The driver said that he was driving and it was inconvenient to answer the call; 2023-11-2 13:30:52: The driver felt that the membership fee was too high, and was worried that the number of orders was small and he could not get any orders, so he needed to reconsider; 2023-11-3 15:30:37: The driver asked about the insurance fee and felt it was too high, so he would reconsider; the driver was last active on 2023-11-6 18:31:37, and logged into the APP 10 times in the past week, with a total active time of 54 minutes a week]".

[0060] S4. Inputting the synthesized text into a large model for intent recognition to identify the driver's intention to join the recruitment platform.

[0061] By calling the interface of the big model, the synthesized text is input into the big model. By extracting features from each part of the synthesized text, on the one hand, the driver's historical communication records are considered to identify the driver's demand changes and emotional tendencies at different historical time points; on the other hand, the driver's recent active behavior is also considered, and the driver's willingness to join the platform is inferred through information such as the number of logins and active time. Combining the above two, the driver's intention to join the recruitment platform is finally obtained. Among them, the big model includes but is not limited to chatglm, baichuan, qwen, etc. It can be understood that the driver's intention to join the recruitment platform can be divided into positive intention, uncertain intention and negative intention, which makes it easier for the recruitment platform to screen out target drivers for recruitment.

[0062] In one embodiment of the present invention, the synthesized text is input into the big model for intent recognition, and after the driver's intention to join the recruitment platform is identified, it also includes: predicting the probability of the driver joining the recruitment platform and the corresponding reason analysis. In order to facilitate the recruitment platform to more accurately identify the driver's intention, facilitate subsequent recruitment decisions, and improve the trust in the model results, after obtaining the driver's intention, the present invention will further predict the probability of the driver joining the recruitment platform through the big model and analyze the corresponding reasons. Exemplarily, the output of the big model is "Driver's intention: positive intention, probability of joining the platform: 0.8; reason analysis: The driver was initially worried that he would not be able to receive orders after joining the platform, and thought that the membership fee was too high, so he had certain concerns, but his attitude changed to a certain extent after being informed that the insurance premium was reduced."

[0063] See also Figure 2 , the driver intention recognition system 100 includes: an information acquisition module 110, a summary generation module 120, a merging module 130 and an intention recognition module 140. The above-mentioned information acquisition module 110 is used to obtain multiple recruitment communication records of the driver and the active behavior of the driver on the recruitment platform within a preset time period; wherein the recruitment communication record includes the unilateral conversation content between the driver and the recruitment platform and its corresponding time mark. The summary generation module 120 is used to input each recruitment communication record into the summary extraction model to obtain the conversation summary of each recruitment communication record; wherein the summary extraction model is a natural language processing model. The merging module 130 is used to sort the conversation summaries of each time according to their respective time sequence, and merge the active behavior and the sorted conversation summary to form a synthetic text. The intention recognition module 140 is used to input the synthetic text into the large model for intention recognition, and identify the intention of the driver to be recruited.

[0064] The specific definition of the driver intention recognition system can be found in the above definition of the driver intention recognition method, which will not be repeated here. Each module in the above driver intention recognition system can be implemented in whole or in part by software, hardware and a combination thereof. The above modules can be embedded in or independent of the processor in the computer device in hardware format, or stored in the memory of the computer device in software format, so that the processor can call the corresponding operations of each of the above modules.

[0065] It should be noted that, in order to highlight the innovative part of the present invention, the present embodiment does not introduce modules that are not closely related to solving the technical problem proposed by the present invention, but this does not mean that there are no other modules in the present embodiment.

[0066] See also Figure 3 The electronic device 1 may include a memory 12, a processor 13 and a bus, and may also include a computer program stored in the memory 12 and executable on the processor 13, such as a driver intention recognition program.

[0067] Among them, the memory 12 includes at least one type of readable storage medium, and the readable storage medium includes flash memory, mobile hard disk, multimedia card, card-type memory (for example: SD or DX memory, etc.), magnetic memory, disk, optical disk, etc. In some embodiments, the memory 12 can be an internal storage unit of the electronic device 1, such as a mobile hard disk of the electronic device 1. In other embodiments, the memory 12 can also be an external storage device of the electronic device 1, such as a plug-in mobile hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), etc. equipped on the electronic device 1. Further, the memory 12 can also include both an internal storage unit of the electronic device 1 and an external storage device. The memory 12 can not only be used to store application software and various types of data installed in the electronic device 1, such as the code for identifying the driver's intention, etc., but can also be used to temporarily store data that has been output or is to be output.

[0068] In some embodiments, the processor 13 may be composed of an integrated circuit, for example, a single packaged integrated circuit, or a plurality of packaged integrated circuits with the same or different functions, including one or more central processing units (CPUs), microprocessors, digital processing chips, graphics processors, and combinations of various control chips. The processor 13 is the control core (Control Unit) of the electronic device 1, and uses various interfaces and lines to connect various components of the entire electronic device 1, and executes or executes programs or modules (such as driver intention recognition programs, etc.) stored in the memory 12, and calls data stored in the memory 12 to execute various functions of the electronic device 1 and process data.

[0069] The processor 13 executes the operating system and various installed applications of the electronic device 1. The processor 13 executes the applications to implement the steps in the above-mentioned method for identifying the driver's intention.

[0070] Exemplarily, the computer program may be divided into one or more modules, which are stored in the memory 12 and executed by the processor 13 to complete the present application. The one or more modules may be a series of computer program instruction segments capable of completing specific functions, which are used to describe the execution process of the computer program in the electronic device 1. For example, the computer program may be divided into an information acquisition module 110, a summary generation module 120, a merging module 130, and an intent recognition module 140.

[0071] The above-mentioned integrated unit implemented in the form of a software function module can be stored in a computer-readable storage medium, and the computer-readable storage medium can be non-volatile or volatile. The above-mentioned software function module is stored in a storage medium, including several instructions for enabling a computer device (which can be a personal computer, a computer device, or a network device, etc.) or a processor to perform part of the functions of the driver intention recognition method described in various embodiments of the present application.

[0072] In summary, the present invention discloses a method, system, device and medium for identifying driver intentions. By extracting dialogue summaries from multiple rounds of recruitment communication records of drivers, the amount of data that needs to be analyzed by the model can be effectively reduced, and the interference of noise in the processing process can be reduced, thereby greatly improving the processing rate and accuracy. The central idea of ​​each recruitment communication record is extracted in the form of a summary, and a synthetic text is generated in combination with its active behavior information. The intention of the synthetic text is recognized by a large model, so that the driver's intention can be more accurately identified. The present invention combines the driver's historical data and recent active behavior for comprehensive analysis, and can more accurately and quickly predict the driver's intention to join the platform. Furthermore, the present invention can effectively process a large amount of historical dialogue data, retain the importance of time series, and integrate real-time features with dialogue text information, making full use of the powerful processing capabilities of the large model to improve the accuracy of the prediction probability and the intention recall rate. At the same time, providing more explainable results is conducive to improving the business team's trust and acceptance of the model, and further optimizing the efficiency and cost of driver recruitment. The problem of low quality of clues and invitation efficiency in existing driver recruitment methods is effectively improved, the problem of existing technologies being limited by text length and feature quality is improved, and the prediction accuracy and recall rate of the model are improved. Therefore, the present invention effectively overcomes various shortcomings in the existing technologies and has high industrial utilization value.

[0073] The above embodiments are merely illustrative of the principles and effects of the present invention, and are not intended to limit the present invention. Anyone familiar with the art may modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by a person of ordinary skill in the art without departing from the spirit and technical ideas disclosed by the present invention shall still be covered by the claims of the present invention.< / end> < / start>

Claims

1. A method for identifying driver intention, characterized in that: Applied to a recruitment platform, the method comprises: Obtaining multiple recruitment communication records of the driver and the driver's active behavior on the recruitment platform within a preset time period; wherein the recruitment communication records include the unilateral conversation content between the driver and the recruitment platform and its corresponding time stamp; Input each recruitment communication record into a summary extraction model to obtain a conversation summary of each recruitment communication record; wherein the summary extraction model is a natural language processing model; Sort the conversation summaries of each session according to their respective chronological order, and merge the active behaviors and the sorted conversation summaries to form a synthetic text; The synthesized text is input into a large model for intent recognition to identify the driver's intention to join the recruitment platform.

2. The method for identifying driver intention according to claim 1, characterized in that: The summary extraction model is a T5 model. For each recruitment communication record, the recruitment communication record is input into the summary extraction model to obtain a dialogue summary of the recruitment communication record, including: Pre-process the recruitment communication records; Input the preprocessed recruitment communication records into the encoder network of the T5 model, extract features layer by layer based on multiple cascaded self-attention layers, and obtain semantic features of the recruitment communication records; The semantic features are input into the decoder network of the T5 model, and text information is gradually generated based on the autoregressive mechanism. When the preset end conditions are met, the text information generated each time is arranged in sequence to generate a conversation summary.

3. The method for identifying driver intention according to claim 2, characterized in that: The semantic features are input into the decoder network of the T5 model, and text information is gradually generated based on the autoregressive mechanism. When the preset end condition is met, the text information generated each time is arranged in sequence to generate a dialogue summary, including: Extracting a first semantic feature vector of the semantic feature; The extracted semantic feature vector is input into the decoder network, and is decoded based on the self-attention mechanism to obtain text information corresponding to the extracted semantic feature vector; wherein, based on the autoregressive mechanism, when the next semantic feature vector is decoded, the decoded text information is also input into the decoder network to participate in decoding; Determine whether the text information corresponding to the extracted semantic feature vector contains the preset end information: If so, the text information generated by the semantic features is arranged in a chronological order to form a conversation summary; Otherwise, the next semantic feature vector is sequentially extracted from the semantic features and input into the decoder network again for decoding.

4. The method for identifying driver intention according to claim 1, characterized in that: The summary extraction model is a BERT model. For each recruitment communication record, the recruitment communication record is input into the summary extraction model to obtain a dialogue summary of the recruitment communication record, including: Pre-process the recruitment communication records; Encode each conversation sentence in the preprocessed recruitment communication record to obtain a sentence-level vector for each conversation sentence; Perform average pooling on the sentence-level vectors of all dialogue sentences to obtain text vectors; The similarity between each sentence-level vector and the text vector is calculated, and the sentence-level vector with the greatest similarity is screened out, and the conversation sentence corresponding to the sentence-level vector is used as the conversation summary.

5. The method for identifying driver intention according to any one of claims 2 or 4, characterized in that: The recruitment communication records are pre-processed, including: The recruitment communication records were cleaned textually; In the recruitment communication records after the text cleaning, the time tags therein are unified into a preset format to obtain the pre-processed recruitment communication records.

6. The method for identifying driver intention according to claim 1, characterized in that: The active behavior and the sorted conversation summary are merged to form a synthetic text, including: Based on the preset format specifications, active behaviors are converted into standardized text features; The standardized text features and the sorted conversation summaries are combined to form a synthetic text.

7. The method for identifying driver intention according to claim 1, characterized in that: The synthesized text is input into the large model for intention recognition. After the intention of the driver to join the recruitment platform is recognized, the method further includes: predicting the probability of the driver joining the recruitment platform and the corresponding reason analysis.

8. A driver intention recognition system, characterized in that: The system comprises: An information acquisition module, used to acquire multiple recruitment communication records of the driver and the driver's active behavior on the recruitment platform within a preset time period; wherein the recruitment communication record includes the unilateral conversation content between the driver and the recruitment platform and its corresponding time mark; A summary generation module is used to input each recruitment communication record into a summary extraction model to obtain a conversation summary of each recruitment communication record; wherein the summary extraction model is a natural language processing model; A merging module is used to sort the conversation summaries of each session according to their respective chronological order, and merge the active behaviors and the sorted conversation summaries to form a synthetic text; The intention recognition module is used to input the synthesized text into the large model for intention recognition, so as to identify the intention of the driver to be recruited.

9. An electronic device, characterized in that: The electronic device comprises: one or more processors; A storage device for storing one or more programs, which, when executed by the one or more processors, enables the electronic device to implement the method for identifying the driver's intention as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and when the computer program is executed by a processor of a computer, the computer is caused to execute the method for identifying the driver's intention as described in any one of claims 1 to 7.

Images