A driver-passenger dispute identification method of a network car-hailing platform

Abstract

The application provides a driver-passenger dispute identification method for a network car-hailing platform, comprising the following steps: S1: performing ASR voice transcription on the driver-passenger conversation voice of the driver and passenger, and arranging the voice according to the speaking order; S2: performing semantic identification on the driver-passenger voice sequence text obtained through voice transcription; S3: establishing a rule system, performing rule processing on the semantic identification of the driver-passenger multi-label, and realizing initial responsibility determination of the driver-passenger dispute; and S4: establishing a supplementary system, processing the initial responsibility determination of the driver-passenger dispute, and obtaining final responsibility determination of the driver-passenger dispute. Through the close coupling of multiple levels of subsystems, the application finally realizes extremely high accuracy of responsibility determination, realizes automatic and reasonable responsibility determination in the application of the network car-hailing platform, saves a large amount of human cost, improves the accuracy of identification, and improves the service quality.

Description

TECHNICAL FIELD

[0001] The application relates to a network car-hailing platform management method, in particular to a driver-passenger dispute identification method for a network car-hailing platform. BACKGROUND

[0002] Since the wide popularity of mobile payment, network car-hailing services have emerged, and a large number of dispute cases have been generated in daily network car-hailing transactions. Driver-passenger disputes and even conflicts frequently occur in the network car-hailing market, which greatly hinders the development process of the network car-hailing market. Based on various news reports, it is not difficult to find that most of the driver-passenger disputes and conflicts are caused by problems in the language communication between the vehicle driver and the passenger. The objective and fair judgment of the network car-hailing platform on the dispute cases reflects the correct service attitude of the platform to the driver and the passenger. SUMMARY

[0003] The application provides a driver-passenger dispute identification method for a network car-hailing platform, which solves the problem of judging the responsibility of the driver-passenger dispute through the platform, and the technical scheme is as follows:

[0004] A driver-passenger dispute identification method for a network car-hailing platform, comprising the following steps:

[0005] S1: performing ASR voice transcription on the driver-passenger conversation voice of the driver and the passenger, and arranging the voice in the order of speaking;

[0006] S2: performing semantic identification on the driver-passenger voice sequence text obtained through voice transcription;

[0007] S3: establishing a rule system, performing rule processing on the semantic identification of the driver-passenger multi-label, and realizing the initial responsibility judgment of the driver-passenger dispute;

[0008] S4: establishing a supplementary system, processing the initial responsibility judgment of the driver-passenger dispute, and obtaining the final responsibility judgment of the driver-passenger dispute.

[0009] Further, in step S2, the semantic identification includes a keyword model, an FBA model and a milvus similarity model, and the corresponding semantic is obtained by sequentially processing the text content using the three models.

[0010] The keyword model is used to determine whether the driver-passenger text contains a word that can directly determine the semantic, and if so, the intention of a sentence is directly given.

[0011] The FBA model comprises the following steps:

[0012] S11: Extract the semantic features of the text using the pre-trained language model fastbert, and use a sliding window to truncate the original document to address the problem of long text;

[0013] S12: Input the collected multi-class intent data into fastbert for training, with at least 200 examples for each class;

[0014] S13: Use a weighted fusion model to predict the sample label after each layer of fastbert. If the confidence exceeds the threshold in the first few layers, the label is directly output without calculating the subsequent layers. The more difficult to predict will be predicted in the last 12th layer.

[0015] The milvus similarity model encodes the training data of step S13 using simbert, then inputs it into the milvus vector database, and realizes that the sentences with low confidence output by fastbert are sent to milvus for prediction to output the most similar label.

[0016] Further, in step S3, artificial custom dispute rules are used to determine the language semantics parsed from the driver-passenger conversation to determine the dispute initiator and dispute responsible party.

[0017] Map the driver and passenger intentions obtained in step S2 to them and return the risk level, where A represents the driver is responsible, B represents the driver is probably responsible, C represents the driver is probably not responsible, and D represents the driver is completely not responsible.

[0018] Further, in step S4, for disputes that exceed the rule system, other set rules are developed, and only need to determine which set rules are hit, then use a set of supplementary rules to correct the intention, to obtain a more accurate risk level.

[0019] The driver-passenger dispute identification method of the online car-hailing platform tightly couples multiple levels of subsystems to achieve high accuracy in identifying responsibility, and realizes automatic and reasonable identification of responsibility in the application of the online car-hailing platform, saving a large amount of manpower cost, improving the accuracy of identification, and improving the service quality. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 is a flowchart of the driver-passenger dispute identification method of the online car-hailing platform;

[0021] Figure 2 is a flowchart of the language identification process;

[0022] Figure 3 is a schematic diagram of the FBA model;

[0023] Figure 4 is an example schematic diagram of the business rule. DETAILED DESCRIPTION

[0024] Normal service process of online car-hailing: A1. Passenger places an order; A2. Driver grabs the order or accepts the assigned order; A3. Driver calls the passenger to inform the driver of the arrival; A4. Passenger answers and walks to the pickup location; A5. Driver picks up the passenger and drives; A6. Driver arrives at the destination and completes the order.

[0025] Possible dispute scenarios in the service process:

[0026] B1 Induced cancellation: In the A3 scenario, the driver is reluctant to take the order and suggests the passenger to cancel for reasons such as rest, wrong order grabbing, and traffic accident.

[0027] B2 Private transaction: In the A3 or A5 scenario, the driver does not want to complete the order through private transactions instead of taking a commission from the platform.

[0028] B3 Driver and passenger abuse: In the A3 scenario, the driver and passenger abuse each other for various reasons, resulting in the failure to complete the service.

[0029] Driver and passenger disputes usually occur in the A3 step, and the present application needs to collect the voice of the driver and passenger, process it, and determine the responsibility.

[0030] As shown in Figure 1 , the driver and passenger dispute determination method of the online car-hailing platform includes the following steps:

[0031] S1: Transcribe the driver and passenger conversation voice of the driver and passenger into ASR voice, and arrange it according to the speaking order;

[0032] S2: Perform semantic recognition on the driver and passenger voice sequence text obtained by voice transcription;

[0033] The purpose of semantic recognition is to recognize the intentions of both parties through the content of their communication, so that each language semantic can be accurately understood in the subsequent process, so as to determine the initiator of the dispute, decompose the dispute process, and confirm the responsible party of the dispute.

[0034] The purpose of this step is to convert the voice text into multi-label language intent, such as the following example voice text:

[0035]

[0036] When performing semantic recognition, as shown in Figure 2 , semantic recognition includes three models: keyword model, fastbert text classification model, and milvus text similarity model. By using the three models to process the text content in turn, the corresponding semantic is obtained.

[0037] In this stage, the patent analyzes the meaning contained in each sentence in the driver-passenger dialogue through multiple models. The total number can be summarized into a limited number of intentions, and some of the intention classification patterns are as follows:

[0038] 1. The driver has a strong intention to be in place

[0039] 2. The driver has already been in place

[0040] 3. The driver indicates that he cannot pick up the passenger

[0041] 4. The driver indicates that he forgets to close the order

[0042] 5. The driver eats, charges, etc.

[0043] 6. The driver indicates that the passenger does not wear a mask and cannot be picked up

[0044] 7. The passenger indicates that he has not made a car call mistake

[0045] 8. The passenger forgets to cancel

[0046] 9. The passenger calls another car

[0047] 10. The passenger's additional demand

[0048] 11. The passenger indicates that he has canceled

[0049] 12. The passenger's positioning is problematic

[0050] 13. The passenger postpones the pickup time ...

[0052] In the present application, the models for different meanings are optimized, and the following is a description of several models:

[0053] 1. Keyword model

[0054] The keyword model is to determine whether the text contains words that can directly determine the meaning, and if so, the intention of a sentence is directly given.

[0055] 2. FBA model (fastbert adaptive mechanism)

[0056] The FBA model is an optimization based on fastbert, as shown in Figure 3

[0057] S11: The pre-training language model of fastbert is used to extract text semantic features. Fastbert is different from BERT and other pre-training language models, which uses relative position encoding and does not limit the length of the input text. At the same time, a sliding window is used to truncate the original document to solve the problem of long text.​

[0058] S12: The collected multi-class intent data is input into fastbert for training, and at least 200 examples are input for each class.

[0059] S13: The fastbert of each layer of the weighted fusion model is predicted after the Transformer, and if the confidence exceeds the threshold in the previous few layers, the label is directly output without calculating the later layers. The more difficult to predict will get the prediction output in the last 12th layer, which not only improves the accuracy, but also improves the running speed.

[0060] 3, milvus similarity model

[0061] In this stage, the training data in S13 is encoded by simbert and then input into the milvus vector database. The sentences that fastbert cannot output with high confidence are sent to milvus to predict the most similar label. Finally output:

[0062]

[0063]

[0064] S3: Establish a rule system for semantic recognition of driver and passenger multi-label, and perform rule processing to realize initial responsibility determination of driver and passenger disputes;

[0065] Use artificially customized dispute rules to judge the language semantics parsed in the driver and passenger dialogue to determine the dispute initiator and dispute responsible party.

[0066] The application adopts a rule system of a set of business rules, respectively maps the driver and passenger intents obtained in S2 to the rule system, and returns a risk level, and the mapping rules are as follows Figure 4 Wherein A represents that the driver is responsible, B represents that the driver is probably responsible, C represents that the driver is probably not responsible, and D represents that the driver is completely not responsible.

[0067] For example, the driver's strong intention to just be or the driver has just been, the level is C;

[0068] The driver indicates that there are other passengers in the car or the driver forgets to close the order or the driver eats and charges, etc., the level is A;

[0069] The driver expresses apology and thanks, and the level is B.

[0070] S4: Establish a supplementary system to process the initial responsibility determination of driver and passenger disputes to obtain the final responsibility determination of driver and passenger disputes.

[0071] Through the initial responsibility judgment of driver-passenger disputes, the dispute responsibility is determined, but in some cases, the rule system is still insufficient to cover all dispute cases, and it is necessary to continue to judge after hitting other set rules, only to judge which set rules are hit, and then use a set of supplementary rules to correct the intention, so that a more accurate risk control level can be obtained.

[0072] The model deployment method used in the application is as follows:

[0073] 1. An online web service is built through fastAPI and K8s, and an interface program is embedded therein to realize web calling of the online service.

[0074] 2. The service is accessed through a UI end, and relevant dispute types and dispute responsibility judgment levels are returned

[0075] Through experiments, compared with the previous use of a simple risk control identification of a bert model, the accuracy of the application is improved from 64% to 92%.

[0076] The application designs a new online car-hailing dispute risk control model analysis framework, designs three independent models, and closely cooperates to realize high-accuracy judgment, and the application also designs a rule system for use with a semantic model to avoid the low accuracy of a multi-classification or multi-relation extraction model.

Claims

1. A method for identifying disputes between drivers and passengers on a ride-hailing platform, comprising the following steps: S1: Transcribe the speech of the driver and passenger through ASR, and arrange the speech in the order of speaking; S2: Perform semantic identification on the text sequence obtained by transcribing the speech; the semantic identification includes a keyword model, an FBA model, and a milvus similarity model. The text content is processed using the three models in sequence to obtain the corresponding semantic meaning. The keyword model determines whether the text contains words that directly determine the semantic meaning. If there are, the intent of a sentence is directly given; The FBA model includes the following steps: S11: Use the fastbert pre-training language model to extract text semantic features, and use a sliding window to truncate the original document to handle the problem of long text; S12: Input the collected multi-class intent data into fastbert for training, and input at least 200 samples for each class; S13: Use a weighted fusion model to predict the sample label after each layer of fastbert. If the confidence level exceeds the threshold in the first few layers, the label is directly output without calculating the subsequent layers. The more difficult to predict will be predicted in the last 12th layer; The milvus similarity model encodes the training data of step S13 using simbert, and then inputs it into the milvus vector database to achieve the output of the most similar label in milvus for the statement that fastbert cannot output with high confidence; S3: Establish a rule system to identify the semantic meaning of the driver and passenger, and perform rule processing to determine the initial responsibility of the driver and passenger dispute; use artificially customized dispute rules to determine the language semantic meaning parsed from the driver and passenger conversation to determine the dispute initiator and dispute responsible party; S4: Establish a supplementary system to process the initial responsibility determination of the driver and passenger dispute to obtain the final responsibility determination of the driver and passenger dispute, The model used in the method for identifying disputes between drivers and passengers on a ride-hailing platform is built online using fastAPI and K8s, and the interface program is embedded to realize web calling of online services, access to services through the UI, and return of relevant dispute types and dispute responsibility determination levels. 2.The method of claim 1, wherein: The intent of the driver and passenger obtained in step S2 is mapped to it, and the risk level is returned, where A represents the driver is responsible, B represents the driver is probably responsible, C represents the driver is probably not responsible, and D represents the driver is completely not responsible. 3.The method of claim 1, wherein: In step S4, for disputes that exceed the rule system, other setting rules are developed, and only the hit setting rules need to be determined, and then a set of supplementary rules is used to correct the intent to obtain a more accurate risk level.

Images