Logistics delivery notification method and device, and storage medium
By automatically dialing the recipient's phone number and using deep learning models to identify the intent of receiving, the problem of low format limitation and accuracy in the policy verification method is solved, improving the accuracy and efficiency of policy verification, and improving the user experience.
Patent Information
- Application Number
- CN202410984930.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-22
- Publication Date
- 2024-11-08
AI Technical Summary
The existing policy verification methods have the problem of large policy format limitations and low calibration accuracy, which is difficult to effectively improve the accuracy and efficiency of policy verification.
By obtaining logistics information to be notified from the express service system, automatically dialing the recipient's phone number, identifying the ringback tone to determine the connection type, if connected, the voice data of the call is obtained through the preset speech process, and the recipient is identified using the deep learning model The purpose of receiving the receipt is to determine the logistics genre method, and to achieve rapid statistics of recipient intention information and timely synchronization.
It improves the efficiency of express delivery, reduces the possibility of user complaints, improves the user experience, and improves the accuracy and efficiency of insurance policy verification.
Smart Images

Figure CN118917770A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of logistics technology, and in particular to a logistics delivery notification method, device and storage medium. Background Art
[0002] Policy verification refers to verifying the authenticity of insurance documents to ensure the legitimacy and validity of the policy. With the continuous development of the insurance field, the importance of verifying the authenticity of insurance policies has become increasingly prominent for users and corporate institutions. By verifying the information, terms and relevant qualifications of the insurance company on the policy, risks can be effectively reduced and the rights and interests of the policyholder can be protected.
[0003] The current method of verifying the authenticity of insurance policies is mainly to store encrypted information in the PDF format of the insurance policy, and then upload the PDF format of the insurance policy to the verification website for verification. However, the current method of verifying the authenticity of insurance policies has the problem of large restrictions on the insurance policy format and low verification accuracy.
[0004] Therefore, how to improve the accuracy of policy verification and reduce the limitations of policy verification is an urgent problem that needs to be solved. Summary of the invention
[0005] The present application provides a logistics delivery notification method, device and storage medium to solve the problem of ...
[0006] In a first aspect, the present application provides a logistics delivery notification method, comprising:
[0007] Obtaining logistics information to be notified from the express delivery service system, wherein the logistics information to be notified includes the recipient's contact information;
[0008] Automatically dial the recipient's phone number according to the recipient's contact information;
[0009] Identify the ringback tone of the recipient's phone during the automatic dialing process, and determine the connection type of the recipient's phone according to the ringback tone, wherein the connection type includes connected and unconnected, and the types of the ringback tone include ringing type, no answer type, call in progress type, power off type, out of service type, and unavailable number type;
[0010] If the connection type is not connected, the recipient's phone number is automatically replayed according to a preset time interval and / or replay cycle number;
[0011] If the connection type is connected, the call voice data of the recipient is obtained through a preset speech process;
[0012] Identify the call voice data and determine the recipient's receiving intention;
[0013] The delivery method of the logistics is determined according to the recipient's receiving intention.
[0014] Optionally, judging the connection type of the recipient's phone according to the ringback tone includes:
[0015] If the ringback tone is of the middle ringing type and the call is connected within the first preset duration, the connection type is connected;
[0016] If the ringback tone is the ringing type and the call is still not connected after reaching the first preset time, or the ringback tone is any one of the no-answer type, call in progress type, power-off type, out-of-service type, and unavailable number type, then the connection type is not connected.
[0017] Optionally, the identifying the call voice data and determining the recipient's receiving intention includes:
[0018] Processing the call voice data through a trained deep learning model to obtain a spectrogram of the user in the call voice data;
[0019] Extracting a speech feature sequence of the user's speech from the spectrogram, and decoding the speech feature sequence to generate a phoneme sequence of the user's speech, wherein the phoneme sequence is used to determine the content of the user's speech;
[0020] Combining the phoneme sequence into text information corresponding to the call voice data using a language model;
[0021] Inputting the text information into an intention recognition model to extract an intention recognition feature vector of the text information;
[0022] Determine the receiving intention corresponding to the text information by performing classification processing on the intention recognition feature vector, wherein the receiving intention includes a delivery method and a delivery time;
[0023] The step of processing the call voice data by using a trained deep learning model to obtain a spectrogram of the user in the call voice data includes:
[0024] The trained deep learning model is used to extract audio features of the call voice data, wherein the audio features include frequency features, time features, and amplitude features. The frequency features represent energy distribution of the audio signal of the call voice data at different frequencies, the time features represent temporal changes of the audio signal of the call voice data, and the amplitude features represent the intensity of the audio signal of the call voice data.
[0025] The frequency feature, the time feature, and the amplitude feature are concatenated to generate an audio feature matrix;
[0026] Converting the time domain signal in the audio feature matrix into a frequency domain signal based on discrete cosine transform to obtain the frequency spectrum feature of the call voice data;
[0027] Mapping the characteristic values in the frequency spectrum characteristics to the horizontal axis of the initial spectrogram to generate an initial spectrogram of the user in the call voice data, wherein the horizontal axis of the initial spectrogram represents time, the vertical axis represents frequency, and the brightness or color of each pixel represents the energy at the corresponding time and frequency;
[0028] The initial spectrogram is smoothed based on Gaussian filtering to generate a spectrogram of the user in the call voice data.
[0029] Optionally, before inputting the text information into the intent recognition model, the method further includes:
[0030] Detecting the text information using a deep learning model and extracting error recognition features of the text information, wherein the error recognition features include text sequence features, context association features, and semantic features;
[0031] Based on the text sequence features, contextual features, and semantic features, output the error probability of each text segment in the text information, wherein the text segment includes a word, a sentence, and a paragraph of the text information;
[0032] Obtaining an adjustment weight of each text segment according to the appearance frequency and importance ratio of each text segment in the text information;
[0033] According to the error probability of the word segmentation and the adjusted probability generated by the adjustment weight of the word segmentation, as well as the word segmentation error probability threshold, the error probability of the sentence and the adjusted probability generated by the adjustment weight of the sentence, as well as the sentence error probability threshold, the error probability of the paragraph and the adjusted probability generated by the adjustment weight of the paragraph, and the paragraph error probability threshold, respectively determine the type of the word segmentation, the type of the sentence, and the type of the paragraph; the types include a type to be corrected and a correct type;
[0034] According to the type of the word segmentation, the type of the sentence, and the type of the paragraph, a fusion calculation is performed to obtain the field to be corrected in the text information;
[0035] The fields to be corrected are corrected through a preset field correction knowledge graph to ensure the accuracy and appropriateness of the text information.
[0036] Optionally, the step of obtaining the field to be corrected in the text information by fusing and calculating the type of the word segmentation, the type of the sentence, and the type of the paragraph includes:
[0037] According to the internal connection between the participle, the sentence, and the paragraph, a first correlation feature between the participle and the sentence, a second correlation feature between the sentence and the paragraph, and a third correlation feature between the participle and the paragraph are obtained, wherein the correlation features are used to indicate whether there is a logical conflict between the participle, the sentence, and the paragraph;
[0038] Constructing a correlation matrix of the text information through the first correlation feature, the second correlation feature, and the third correlation feature;
[0039] Calculate the association strength between the word segmentation, the sentence, and the paragraph according to the association matrix;
[0040] If the correlation strength is lower than a preset strength threshold, any correlation feature in the correlation matrix is deleted to generate sub-correlation matrices;
[0041] Calculating the sub-correlation strength corresponding to the sub-correlation matrix according to each of the sub-correlation matrices;
[0042] Using the correlation feature deleted from the sub-correlation matrix with the highest sub-correlation strength as the target correlation feature;
[0043] Taking the text segment corresponding to the target correlation feature as the field to be corrected;
[0044] The obtaining, according to the intrinsic relationship between the segmentation, the sentence, and the paragraph, a first correlation feature between the segmentation and the sentence includes:
[0045] Using a word vector model, converting the word segmentation included in the sentence into a first word segmentation vector;
[0046] Generate a sentence vector corresponding to the sentence according to the first word segmentation vector, where the sentence vector is generated based on an average value of the first word segmentation vector;
[0047] Using a word vector model, converting the word segmentation into a second word segmentation vector;
[0048] The cosine similarity between the sentence vector and the second word segmentation vector is calculated based on a similarity analysis method to generate a first correlation feature between the word segmentation and the sentence.
[0049] Optionally, the method further includes:
[0050] Acquire the user's historical interaction record, wherein the historical interaction record includes the user's historical receiving intention;
[0051] Inputting the historical interaction records and the call voice data into a trained intention inference model, and outputting probability data of multiple user intentions corresponding to the historical interaction records and the call voice data;
[0052] If the number of probability data of the multiple user intentions that are within the target probability interval is greater than or equal to a preset number threshold, the user's deep receiving intention will be determined based on similar features of the multiple user intentions;
[0053] The training process of the intention reasoning model specifically includes:
[0054] Acquire multiple historical interaction record samples of multiple users, and convert the historical interaction record samples into standardized text, wherein the historical interaction record samples include interaction content, timestamp, user behavior, and user real intention, wherein the user real intention includes explicit intention and / or implicit intention;
[0055] Generate a training set for the intention inference model according to the standardized text, using the user's real intention as label data;
[0056] The standardized text is input into the intention reasoning model to train the intention reasoning model.
[0057] Optionally, the method further includes:
[0058] Obtaining the user's historical interaction records and historical logistics information;
[0059] Extracting the user's preference features according to the historical interaction records and the historical logistics information, wherein the preference features include time features, communication mode features, and delivery mode features. The time features include communication time preference and delivery time preference. The communication mode features include language style and communication habits.
[0060] Determining a time period for automatically dialing the user's phone according to the communication time preference;
[0061] According to the receiving time preference, determining the receiving time inquiry process in the preset speech process;
[0062] According to the characteristics of the receiving method, determining the receiving method inquiry process in the preset speech process;
[0063] Determining the communication tone and / or communication emotion adopted by the preset speech process according to the language style and the communication habits;
[0064] Adjust the preset speech process according to the receipt time inquiry process, the receipt method inquiry process, and the communication tone and / or communication emotion;
[0065] The recipient's call voice data is obtained through the adjusted preset speech process.
[0066] In a second aspect, the present application provides an electronic device, comprising: a processor, a communication interface and a memory, wherein the processor is communicatively connected to the communication interface and the memory respectively;
[0067] The memory stores computer-executable instructions;
[0068] The communication interface communicates and interacts with external devices;
[0069] The processor executes the computer-executable instructions stored in the memory to implement the method as described in any one of the first aspects.
[0070] In a third aspect, the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores computer-executable instructions, and when the computer-executable instructions are executed by a processor, they are used to implement the method as described in any one of the first aspects.
[0071] In a fourth aspect, the present application provides a computer program product, comprising a computer program, which implements any one of the methods in the first aspect when executed by a processor.
[0072] The logistics delivery notification method, device and storage medium provided in the present application automatically dial the recipient's phone according to the recipient's contact information by obtaining the logistics information to be notified from the express service system. In the automatic dialing process, the ringback tone of the recipient's phone is identified, and the connection type of the recipient's phone is determined according to the ringback tone. If the connection type is not connected, the recipient's phone is automatically replayed in a loop according to the preset time interval and / or the number of replay cycles; if the connection type is connected, the recipient's call voice data is obtained through the preset speech process. The call voice data is identified to determine the recipient's receiving intention, and the logistics delivery method is determined according to the recipient's receiving intention, so that the recipient's mobile phone number can be dialed in batches. The information is extracted through voice interaction technology, and the recipient's intention information is quickly counted and synchronized to the courier's work mobile phone in a timely manner, thereby improving the efficiency of express delivery, reducing the possibility of user complaints, and improving user experience. BRIEF DESCRIPTION OF THE DRAWINGS
[0073] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.
[0074] Figure 1 A flow chart of a logistics delivery notification method provided in an embodiment of the present application;
[0075] Figure 2 A flowchart of another logistics delivery notification method provided in an embodiment of the present application;
[0076] Figure 3 A flowchart of another logistics delivery notification method provided in an embodiment of the present application;
[0077] Figure 4 A flowchart of another logistics delivery notification method provided in an embodiment of the present application;
[0078] Figure 5 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application.
[0079] The above drawings have shown clear embodiments of the present application, which will be described in more detail later. These drawings and text descriptions are not intended to limit the scope of the present application in any way, but to illustrate the concept of the present application to those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION
[0080] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The implementations described in the following exemplary embodiments do not represent all implementations consistent with the present application. Instead, they are merely examples of devices and methods consistent with some aspects of the present application as detailed in the appended claims.
[0081] The technical solution of the present application and how the technical solution of the present application solves the above-mentioned technical problems are described in detail below in conjunction with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. The embodiments of the present application will be described below in conjunction with the accompanying drawings.
[0082] Figure 1 A flow chart of a logistics delivery notification method provided in an embodiment of the present application. Figure 1 As shown, the method may include:
[0083] S101. Obtain logistics information to be notified from the express delivery service system.
[0084] The logistics information to be notified includes the recipient's contact information.
[0085] The server connects to the data interface of the express business system to establish a stable communication connection. The server sends requests regularly to obtain the latest logistics information to be notified. This information is usually transmitted in a structured data format, such as JSON or XML format. After receiving this data, the server parses and stores it, extracts key information and stores it in the local database for subsequent processing.
[0086] To ensure the accuracy and integrity of the data, the server may perform data validation and error correction. It checks the format, consistency, and completeness of the data, and takes appropriate measures when errors or omissions are found, such as requesting resending or using an alternative data source.
[0087] For example, suppose the express delivery service system updates a package from region A to region B, and the recipient of the package is user A. When the server retrieves the logistics information to be notified through the data interface, it will store the information in the local database. The database may record the following information: express number 123456789, recipient user A, recipient address region B, contact information 123-4567-890, etc.
[0088] Optionally, the server may periodically synchronize data with the express business system to obtain the latest logistics information to be notified.
[0089] S102: Automatically dial the recipient's phone number according to the recipient's contact information.
[0090] The server calls the phone dialing service and passes the recipient's phone number as a parameter to the service. The phone dialing service uses the pre-configured communication protocol and line to initiate a call to the recipient's phone. The server monitors the status of the phone dialing in real time, including whether it is connected or busy. Once the call is connected, the server switches to call mode and prepares for subsequent interactions.
[0091] During the call, the server may use voice prompts or speech synthesis technology to play some notifications or prompts to the recipient, such as reminding the recipient to answer the phone or informing the recipient of relevant information about the express delivery.
[0092] For example, the server can automatically dial the recipient's phone number based on the recipient's phone number. Once the call is connected, the server will play a prompt tone to inform the recipient that there is a courier waiting to be signed for, and ask the recipient to choose the delivery method and delivery time.
[0093] S103: Identify the ringback tone of the recipient's phone during the automatic dialing process, and determine the connection type of the recipient's phone according to the ringback tone.
[0094] Among them, the connection types include connected and unconnected, and the types of ringback tones include ringing type, no answer type, call in progress type, power off type, out of service type, and empty number type.
[0095] During the call, the server uses audio processing technology to monitor and analyze the ringback tone in real time. The server collects the audio signal of the ringback tone and uses an audio feature extraction algorithm to extract the characteristics of the ringback tone, such as frequency, duration, rhythm, etc.
[0096] The server then matches and judges the extracted features with pre-trained audio recognition models. These audio recognition models are usually built based on machine learning or deep learning algorithms and can classify and identify different types of ringback tones.
[0097] Based on the judgment result, the server determines the connection type of the recipient's phone, such as ringing, no one answered, in a call, turned off, out of service or unavailable.
[0098] For example, when the server dials the recipient's phone, if it hears a ringing type ringback tone and the call is connected within a first preset time, it is determined as a connected type. If the call is still not connected after the first preset time, or if it hears a ringback tone of a no answer type, a call in progress type, a power off type, a shutdown type, or an unavailable number type, it is determined as a not connected type.
[0099] Specifically, this step can be implemented through the following sub-steps:
[0100] S1031. If the ringback tone is of the middle ringing type and is connected within a first preset duration, the connection type is connected.
[0101] When the server is performing automatic dialing, it will continue to monitor the ringback tone in the phone line. If the recognized ringback tone is of the "ringing" type, it means that the phone is trying to connect and the other party's phone is ringing. At this time, the server will start a timer, which is used to measure the time from the start of the phone ringing to the connection. This time is set to a first preset duration, such as 20 seconds. If the call is answered by the other party within this first preset duration, that is, the line status changes from "ringing" to a state where both parties can talk, then the server will determine that the connection type of the call is "connected".
[0102] Exemplarily, the server automatically dials the phone based on the contact information of user B in the express delivery information. After the call is made, the server detects that the ringback tone is of the "ringing" type through voice recognition technology, and then starts timing. After 15 seconds, user B answers the call, and the server detects that the line status has changed, and it is no longer "ringing", but has become a call status. Since 15 seconds is less than the preset first preset duration of 20 seconds, the server determines that the connection type of this call is "connected". Subsequently, the server will interact with user B according to the preset speech process to obtain the necessary logistics information.
[0103] S1032. If the ringback tone is the ringing type and the call is still not connected after reaching the first preset time, or the ringback tone is any one of the no-answer type, call in progress type, power-off type, out-of-service type, and unavailable number type, then the connection type is not connected.
[0104] When the server makes a call and identifies the ringback tone as the "ringing" type, it will wait until the first preset duration ends. If the call is not answered within this duration, that is, the line status remains "ringing", the server will determine that the connection type of the call is "unanswered". In addition, if after making a call, the server identifies the ringback tone as any of the "no answer" type, "on call", "power off" type, "out of service" type, or "unavailable number" types, then there is no need to wait for the first preset duration to end, and the server will immediately determine that the connection type of the call is "unanswered".
[0105] Exemplarily, the server attempts to call user C to notify him of the express delivery information. After the call is made, the server recognizes that the ringback tone is "ringing", so it starts a 20-second countdown. However, after 20 seconds, the call is still not answered, and the line status is always "ringing". Therefore, the server determines that the connection type of this call is "unconnected". In another case, if the server immediately recognizes that the ringback tone is "no answer" or "out of service" or other non-"ringing" types after making the call, it will also immediately determine that the call is "unconnected" without waiting for the 20-second countdown to end. After determining that it is "unconnected", the server will decide whether to replay the call according to the preset rules.
[0106] S104: If the connection type is not connected, the recipient's phone number is automatically replayed in a loop according to a preset time interval and / or a replay cycle number.
[0107] If the server determines that the recipient's phone is not connected, it will replay the message according to the preset strategy. The preset strategy may include the replay time interval and the number of replay cycles.
[0108] The server will wait for a certain time interval before each replay and then initiate the call again. The time interval setting can be adjusted according to the actual situation to avoid frequent replays causing trouble to the recipient.
[0109] The replay loop count setting can limit the number of times the server attempts to replay, avoiding infinite replay. Once the preset number of loops is reached, the server will stop replaying and take other measures, such as sending SMS notifications or recording missed calls.
[0110] For example, if the server fails to dial the recipient's phone for the first time, it will automatically dial again after a period of time according to the preset time interval. If the call is still not connected, the server will continue to automatically dial according to the setting of the number of replay cycles until the call is connected or the maximum number of replays is reached.
[0111] S105. If the connection type is connected, the recipient's call voice data is obtained through a preset speech process.
[0112] When the server detects that the recipient's phone is connected, it will immediately start the preset conversation process. The conversation process usually includes a series of pre-recorded voice prompts and questions. The server will play these voice prompts in sequence and wait for the recipient's answer.
[0113] The server can use speech recognition technology to convert the recipient's answer into text data, and record and store the text data in real time. During the conversation, the server can dynamically adjust subsequent voice prompts and questions based on the recipient's answer to guide the recipient to provide more relevant information.
[0114] For example, after the server connects to the recipient's phone, it plays a preset line, such as "Hello, you have a courier to sign for, please choose the delivery method and time". Then, the server obtains the recipient's answer through voice recognition technology and converts it into text data. Based on the recipient's choice, the server can further provide relevant options and confirmation information.
[0115] S106: Identify the call voice data and determine the recipient's receiving intention.
[0116] The server converts the recipient's call voice data into text information using speech recognition technology. Speech recognition technology is usually based on deep learning models that can understand and analyze natural language.
[0117] The server performs natural language processing and semantic analysis on the converted text information to extract the recipient's key information and intentions. It identifies the recipient's needs, requirements, and preferences, such as the time of delivery, location of delivery, special instructions, etc.
[0118] The server may use machine learning algorithms or pattern matching techniques to classify and identify the recipient's intent. It matches the recipient's intent with pre-set intent categories to determine the recipient's specific intent.
[0119] For example, the server analyzes the recipient's answer and identifies the delivery method (such as door-to-door delivery, self-collection, etc.) and delivery time selected by the recipient. The server can also further analyze the recipient's other requirements or special circumstances and handle and record them accordingly.
[0120] S107: Determine the logistics delivery method according to the recipient's delivery intention.
[0121] The server adjusts and arranges the delivery method of the logistics according to the recipient's delivery intention. If the recipient specifies a specific delivery time and location, the server will convey this information to the courier or logistics system so that they can deliver the goods as required.
[0122] The server will communicate with the courier or logistics system in real time, inform them of the recipient's requirements and preferences, and coordinate the appropriate delivery time and route. The courier or logistics system will adjust its work plan and itinerary according to the recipient's requirements to ensure that the package can be delivered to the designated location on time.
[0123] For example, the server determines the delivery method of logistics as door-to-door delivery based on the recipient's choice, and passes the recipient's address and delivery time to the courier for delivery. At the same time, the server will also update the order status in the express business system and notify relevant personnel for subsequent processing.
[0124] The method provided in the embodiment of the present application automatically dials the recipient's phone according to the recipient's contact information by obtaining the logistics information to be notified from the express delivery service system. In the automatic dialing process, the ringback tone of the recipient's phone is identified, and the connection type of the recipient's phone is determined according to the ringback tone. If the connection type is not connected, the recipient's phone is automatically replayed in a loop according to the preset time interval and / or the number of replay cycles; if the connection type is connected, the recipient's call voice data is obtained through the preset speech process. The call voice data is identified to determine the recipient's receiving intention, and the logistics delivery method is determined according to the recipient's receiving intention, so that the recipient's mobile phone number can be dialed in batches. The information is extracted through voice interaction technology, and the recipient's intention information is quickly counted and synchronized to the courier's work mobile phone in a timely manner, thereby improving the efficiency of express delivery, reducing the possibility of user complaints, and improving user experience.
[0125] In a possible implementation, the method of step S106 can be implemented by the following steps:
[0126] S1061. Process the call voice data using a trained deep learning model to obtain a spectrogram of the user in the call voice data.
[0127] After collecting the call voice data, the server will use a pre-trained deep learning model to process the data. This model may be an automatic speech recognition (ASR) model based on an architecture such as a recurrent neural network (RNN) or a convolutional neural network (CNN). During the processing, the model will first convert the call voice data into a visual representation, namely a spectrogram. A spectrogram is a graph that can show the frequency of a voice signal changing over time, which can help us better understand the characteristics of a voice signal. Through the complex calculations of the deep learning model, the server can accurately extract the user's spectrogram from the call voice data, laying the foundation for subsequent voice feature extraction and voice content recognition.
[0128] For example, the server receives a piece of voice data from a conversation between a user and a courier. The server calls the trained deep learning model to process the voice data. After complex calculations of the model, the server successfully extracts the user's spectrogram from the voice data. This spectrogram clearly shows the frequency changes and time distribution of the user's voice, providing strong data support for subsequent voice content analysis.
[0129] Specifically, this step can be implemented through the following sub-steps:
[0130] S1. Using the trained deep learning model to extract the audio features of the call voice data, the audio features include frequency features, time features, and amplitude features.
[0131] Among them, the frequency feature represents the energy distribution of the audio signal of the call voice data at different frequencies, the time feature represents the temporal change of the audio signal of the call voice data, and the amplitude feature represents the intensity of the audio signal of the call voice data.
[0132] During the extraction process, the trained deep learning model can analyze the audio signal in the call voice data and extract frequency features, time features, and amplitude features. Frequency features mainly reflect the energy distribution of the audio signal at different frequencies, which helps to identify different phonemes and pronunciation methods in the speech. Time features capture the changes in the audio signal over time, which is crucial for identifying the rhythm and speaking speed of the speech. The amplitude feature represents the strength of the audio signal and can reflect the volume of the speech. These features together constitute a comprehensive description of the audio data and provide rich information for the subsequent generation of the spectrogram.
[0133] For example, the server receives a piece of voice data from a call between a user and customer service, and then calls the trained deep learning model for analysis. Through complex calculations, the model successfully extracts frequency features, time features, and amplitude features from the voice data. For example, the model recognizes that when the user pronounces, the energy distribution at certain specific frequencies is higher, which may be related to the user's pronunciation habits or voice content. At the same time, the model also captures the ups and downs of the user's voice over time, as well as the volume at different time points. These features provide key data for the subsequent spectrogram generation.
[0134] S2. Concatenate the frequency feature, the time feature, and the amplitude feature to generate an audio feature matrix.
[0135] After extracting the frequency, time, and amplitude features of the audio, the server will then perform splicing on these features. The purpose of splicing is to integrate these features into a unified data structure, namely the audio feature matrix. Each row or column of this matrix may represent a specific time frame or frequency band, and the elements in the matrix reflect the feature values in that time frame or frequency band. Through such splicing, the server can easily manage and process these feature data and prepare for the subsequent generation of the spectrogram.
[0136] For example, the server can splice these features according to specific rules to generate an audio feature matrix. This matrix not only contains the energy distribution information of the voice at different frequencies, but also reflects the changes in time and volume of the voice. Through this matrix, the server can have a more comprehensive understanding of the characteristics of the voice data, which facilitates subsequent operations.
[0137] S3. Convert the time domain signal in the audio feature matrix into a frequency domain signal based on discrete cosine transform to obtain the frequency spectrum feature of the call voice data.
[0138] After obtaining the audio feature matrix, the server will use discrete cosine transform (DCT) to convert the time domain signal into a frequency domain signal. DCT is a widely used transformation technology that can convert time domain signals into frequency domain, thereby revealing the spectral characteristics of the signal. Through DCT processing, the server can obtain the spectral characteristics of the call voice data, which can more intuitively reflect the frequency composition and energy distribution of the voice signal. This is of great significance for the subsequent spectrogram generation and voice content analysis.
[0139] S4. Mapping the characteristic values in the frequency spectrum characteristics to the horizontal axis of the initial spectrogram to generate an initial spectrogram of the user in the call voice data.
[0140] The horizontal axis in the initial spectrogram represents time, the vertical axis represents frequency, and the brightness or color of each pixel represents the energy at the corresponding time and frequency.
[0141] After obtaining the spectral features, the server maps these feature values to the horizontal axis of the initial spectrogram. The initial spectrogram is a two-dimensional image representation, where the horizontal axis represents time and the vertical axis represents frequency. The brightness or color of each pixel represents the energy at the corresponding time and frequency. Through the mapping operation, the server can match each feature value in the spectral features with the pixel point in the spectrogram to generate an initial spectrogram. This initial spectrogram intuitively shows the energy distribution of the call voice data at different times and frequencies.
[0142] S5. Smoothing the initial spectrogram based on Gaussian filtering to generate a spectrogram of the user in the call voice data.
[0143] After generating the initial spectrogram, the server will further smooth it to improve the image quality. Smoothing is usually achieved using Gaussian filtering technology. Gaussian filtering is a linear filter that reduces noise and detail levels in an image by taking a weighted average of each pixel in the image and the pixels in its neighborhood. In this way, the server can generate a clearer and smoother spectrogram, providing a more accurate data basis for subsequent speech content recognition and analysis.
[0144] S1062: Extract a speech feature sequence of the user's speech from the spectrogram, and decode the speech feature sequence to generate a phoneme sequence of the user's speech.
[0145] The phoneme sequence is used to determine the content of the user's voice.
[0146] After obtaining the spectrogram, the server will further extract the speech feature sequence from it. These features may include pitch, intensity, timbre, and formants, which can reflect the basic properties of the speech signal. The server uses specific algorithms and technologies, such as Mel-Frequency Cepstral Coefficients (MFCC), to extract these features. The server will then decode the extracted speech feature sequence and convert it into a phoneme sequence. Phonemes are the basic units that make up syllables. By analyzing the phoneme sequence, the server can determine the specific content of the user's voice.
[0147] For example, the server has obtained the user's spectrogram, and then it begins to extract the voice feature sequence from this spectrogram. Through precise calculation and analysis, the server successfully extracts a series of feature values that reflect the characteristics of the user's voice. Then, the server uses a specific decoding algorithm to convert these feature sequences into phoneme sequences. By analyzing these phoneme sequences, the server accurately recognizes the content of the user's voice, such as the specific information that the user said "Please deliver it tomorrow afternoon".
[0148] S1063: Use a language model to combine the phoneme sequence into text information corresponding to the call voice data.
[0149] After obtaining the phoneme sequence, the server will use the language model to combine these phoneme sequences into coherent text information. A language model is a model that can predict and generate text sequences based on language rules. It can help the server combine discrete phoneme sequences into meaningful sentences or phrases. Through the processing of the language model, the server can accurately restore the actual expression of the user during the call and provide accurate text input for subsequent intent recognition.
[0150] For example, the server has obtained the user's phoneme sequence through decoding, and then it uses the language model to combine these phonemes into text information. The language model successfully converts the phoneme sequence into a coherent text based on the grammatical rules and statistical laws of the language: "Please deliver to my home at 3 pm tomorrow." This text accurately reflects the user's actual intention and expression during the call.
[0151] S1064: Input the text information into an intention recognition model to extract an intention recognition feature vector of the text information.
[0152] After receiving the text information, the server will input it into an intent recognition model. This model may be a deep learning model based on natural language processing (NLP) technology, such as long short-term memory network (LSTM) or Transformer. The model will conduct in-depth analysis and processing of the text information and extract feature vectors that can reflect the user's intent. These feature vectors capture the key information and semantic structure in the text, providing an important data basis for subsequent intent classification.
[0153] For example, the server inputs the text message "Please deliver to my home tomorrow at 3 pm" into the intent recognition model. Through complex calculations and analysis, the model successfully extracts intent recognition feature vectors from the text. These feature vectors accurately capture the user's intention and key information that he wants the goods to be delivered to his home at 3 pm tomorrow.
[0154] S1065: Determine the recipient intent corresponding to the text message by performing classification processing on the intent recognition feature vector.
[0155] The receiving intention includes the delivery method and delivery time.
[0156] After extracting the intention recognition feature vectors, the server will use a classification algorithm to classify these feature vectors. The classification algorithm may be a machine learning algorithm such as support vector machine (SVM), random forest or neural network. Through classification processing, the server can accurately determine the user's delivery intention, including specific information such as delivery method and delivery time. This information is crucial for subsequent logistics distribution and customer service.
[0157] For example, the server has extracted the intention recognition feature vectors, and then it uses the classification algorithm to classify these feature vectors. After precise calculation and judgment, the server successfully determines the user's intention to receive the package: the user wants the package to be delivered to his home by express delivery at 3 pm tomorrow. This information will be immediately passed to the logistics system to ensure that the package can be delivered on time according to the user's requirements.
[0158] The method provided in the embodiment of the present application processes the call voice data through a trained deep learning model, obtains the spectrogram of the user in the call voice data, extracts the voice feature sequence of the user's voice from the spectrogram, and decodes the voice feature sequence to generate a phoneme sequence of the user's voice. The phoneme sequence is combined into text information corresponding to the call voice data using a language model, and the text information is input into an intent recognition model to extract an intent recognition feature vector of the text information. By performing classification processing on the intent recognition feature vector, the receiving intent corresponding to the text information is determined, thereby improving the accuracy of extracting the user's receiving intent.
[0159] In a possible implementation, before inputting the text information into the intent recognition model, the method may further include correcting the content of the text information. Figure 2 A flow chart of another logistics delivery notification method provided in an embodiment of the present application. Figure 2 As shown, the method may also include:
[0160] S201. Detect the text information using a deep learning model to extract error recognition features of the text information.
[0161] The error recognition features include text sequence features, context association features, and semantic features.
[0162] The server first loads a trained deep learning model that is specifically designed for text error detection. The model accepts text information as input and performs multi-level feature extraction on it. During the feature extraction process, the model focuses on the sequential features of the text, that is, the order and arrangement of words. It learns the patterns and regularities of the language, capturing the dependencies and contextual information between words. At the same time, the model also considers contextual features, that is, the relationship between words and sentences before and after the text. By understanding the context, the model can better judge whether a word or phrase fits the context.
[0163] In addition, the model analyzes semantic features, that is, the meaning and semantic representation of the text. It tries to understand the semantics of words and judge the correctness of the text based on semantic relationships. By extracting these features, the model can conduct a comprehensive analysis of the text information to find possible errors or unreasonableness.
[0164] For example, assume the text information is: "Please deliver it to my home at 3 pm tomorrow. My home is about ten li away from location XX." The deep learning model will extract features for each word and analyze the relationships between them. The model will notice that the word "li" (a traditional Chinese unit of distance) does not make semantic sense and is inconsistent with common expressions. At the same time, the model will also consider the context information and find that the description "ten li" does not match well with the word "about".
[0165] S202. Output the error probability of each text segment in the text information based on the text sequence features, context association features, and semantic features.
[0166] Wherein, the text segments include word segmentation, sentences, and paragraphs of the text information.
[0167] After extracting the error recognition features of the text, the server can use these features to calculate the error probability of each word segmentation, sentence, and paragraph in the text. This is usually achieved by inputting the features into a probability estimation model. The probability estimation model can be based on statistical methods or machine learning algorithms, such as the Naive Bayes classifier or Support Vector Machine (SVM). The model will calculate the probability of each text segment being incorrect based on the input features.
[0168] The calculation of the error probability will consider multiple factors, such as the degree of deviation of the text segment from the normal language pattern and the abnormality of the features. The model will evaluate the error possibility of each text segment based on the training data and the learned patterns.
[0169] S203. Obtain the adjustment weights of the text segments according to the occurrence frequency and importance ratio of each text segment in the text information.
[0170] The server will determine the adjustment weights of the text segments according to their occurrence frequency and importance ratio in the text. Text segments with higher occurrence frequencies may be considered more important and thus will be assigned higher weights.
[0171] The importance ratio can be determined according to the structure and semantics of the text. For example, titles, abstracts, or key paragraphs may be given higher importance. In addition, the server can also consider other factors, such as the length and position of the text segment, to further adjust the weights.
[0172] By comprehensively considering the occurrence frequency and importance ratio, the server can assign a reasonable adjustment weight to each text segment.
[0173] S204. Determine the type of the word segmentation, the type of the sentence, and the type of the paragraph according to the error probability of the word segmentation and the adjusted probability generated by the adjustment weight of the word segmentation, the word segmentation error probability threshold, the error probability of the sentence and the adjusted probability generated by the adjustment weight of the sentence, the sentence error probability threshold, the error probability of the paragraph and the adjusted probability generated by the adjustment weight of the paragraph, and the paragraph error probability threshold.
[0174] The types include a type to be corrected and a correct type.
[0175] The calculation of the adjusted probability can be achieved by multiplying the error probability with the adjustment weight. This adjusted probability reflects the possibility of error in the text segment and is adjusted accordingly according to the weight.
[0176] Based on the adjustment probability and the preset threshold, the server will classify each text segment as "to be corrected type" or "correct type". The threshold setting can be adjusted according to specific application requirements and data characteristics. If the adjustment probability exceeds the threshold, the text segment is marked as "to be corrected type"; otherwise, it is marked as "correct type".
[0177] S205 , obtaining the field to be corrected in the text information by fusion calculation according to the type of the word segmentation, the type of the sentence, and the type of the paragraph.
[0178] After determining the type of each text segment, the server will enter the fusion calculation stage. The main task of this stage is to accurately locate the specific fields that need to be corrected by comprehensively analyzing and processing all text segments marked as "types to be corrected". To achieve this goal, the server will use complex algorithms and models to integrate various information, including the error probability of the text segment, adjustment weights, and relationships with other text segments. In this way, the server can accurately identify the fields in the text that really need to be corrected, providing strong support for subsequent correction work.
[0179] Specifically, this step can be implemented through the following sub-steps:
[0180] S2051. According to the internal relationship between the participle, the sentence, and the paragraph, obtain a first correlation feature between the participle and the sentence, a second correlation feature between the sentence and the paragraph, and a third correlation feature between the participle and the paragraph.
[0181] The correlation feature is used to indicate whether there is a logical conflict between the word segmentation, the sentence, and the paragraph.
[0182] When processing text information, the server will deeply analyze the internal connections between word segments, sentences, and paragraphs. First, the server will identify each word segment in the text and understand the grammatical role and semantic meaning of these word segments in the sentence. Then, the server will analyze each sentence, explore the logical relationship between the word segments within the sentence, and the connection between sentences. This analysis helps the server obtain the first correlation features between word segments and sentences, which reflect the rationality and logic of the word segment in the sentence.
[0183] The server then further examines the relationship between sentences and paragraphs, analyzing how sentences form paragraphs and the role and position of each sentence in a paragraph. In this way, the server is able to extract the second correlation features between sentences and paragraphs, which reveal the logic and coherence of sentences in the paragraph structure.
[0184] Finally, the server will directly start from the relationship between the word segmentation and the paragraph, explore the distribution, frequency and importance of the word segmentation in the paragraph, and thus obtain the third correlation features between the word segmentation and the paragraph. These features help to identify abnormalities or unreasonableness of the word segmentation in the paragraph.
[0185] Specifically, the method for obtaining the first correlation feature in step S2051 can be implemented by the following sub-steps:
[0186] Using the word vector model, the word segments included in the sentence are converted into the first word segment vector. When processing text data, the server will first use a pre-trained word vector model, such as Word2Vec or GloVe, to convert each word segment in the sentence into a corresponding word vector. These word vectors are points in a high-dimensional space that can capture the semantic relationship between words. For each word segment in the sentence, the server will look up the corresponding vector representation in the word vector model, and these vectors are called first word segment vectors. These vectors not only contain the semantic information of the word segment, but also reflect the relationship between the word segment.
[0187] According to the first word segmentation vector, a sentence vector corresponding to the sentence is generated, and the sentence vector is generated based on the average value of the first word segmentation vector. After obtaining the first word segmentation vectors of all word segments in the sentence, the server calculates the average value of these vectors to generate a sentence vector representing the semantics of the entire sentence, so as to integrate the semantic information of all word segments in the sentence and obtain a unified vector representation.
[0188] The word segmentation is converted into a second word segmentation vector using a word vector model. In order to subsequently calculate the correlation features between the word segmentation and the sentence, the server will again use the word vector model to convert the word segmentation into a second word segmentation vector. The second word segmentation vector here is essentially the same as the first word segmentation vector, both of which are semantic representations of the word segmentation, but they play different roles in subsequent calculations. The first word segmentation vector is used to generate a sentence vector, while the second word segmentation vector is used for similarity analysis with the sentence vector.
[0189] Based on the similarity analysis method, the cosine similarity between the sentence vector and the second word segmentation vector is calculated to generate the first correlation feature between the word segmentation and the sentence. The server uses a similarity analysis method, such as cosine similarity calculation, to measure the similarity between the sentence vector and the second word segmentation vector. Cosine similarity is a commonly used method to measure the angle between vectors. Its value range is [-1,1]. The closer the value is to 1, the more similar the two vectors are. By calculating the cosine similarity between each word segmentation and the sentence, the server can generate the first correlation feature between the word segmentation and the sentence. This feature reflects the semantic importance of the word segmentation in the sentence and its consistency with the overall semantics of the sentence.
[0190] Correspondingly, the method of obtaining the second correlation feature and the third correlation feature in step S2051 is similar to the method of obtaining the first correlation feature, and will not be repeated here.
[0191] S2052: Construct a correlation matrix of the text information by using the first correlation feature, the second correlation feature, and the third correlation feature.
[0192] After obtaining the correlation features between words, sentences, and paragraphs, the server will use these features to build a correlation matrix. This matrix is a multidimensional data structure, in which each dimension represents a different component of the text (word, sentence, paragraph), and the value in the matrix represents the correlation strength or logical rationality between these components.
[0193] The server quantifies the first correlation feature, the second correlation feature and the third correlation feature through an algorithm and fills them into the corresponding positions of the correlation matrix. This process requires complex mathematical calculations and data processing capabilities to ensure the accuracy and effectiveness of the matrix.
[0194] S2053: Calculate the association strength between the word segmentation, the sentence, and the paragraph according to the association matrix.
[0195] After building the correlation matrix, the server can use specific algorithms to calculate the correlation strength between words, sentences, and paragraphs. These algorithms may include matrix operations, statistical analysis, and other methods, aiming to extract meaningful numerical indicators from the correlation matrix to quantify the logical closeness between the various parts of the text. Through these calculations, the server can obtain a comprehensive assessment of the logical coherence of the text.
[0196] S2054: If the correlation strength is lower than a preset strength threshold, any correlation feature in the correlation matrix is deleted respectively to generate sub-correlation matrices.
[0197] When the server finds that the correlation strength between certain words, sentences or paragraphs is lower than the preset strength threshold, it will take further action to identify the problem. Specifically, the server will delete a certain correlation feature in the correlation matrix (such as the correlation between words and sentences, the correlation between sentences and paragraphs, etc.), thereby generating multiple sub-correlation matrices. Each sub-correlation matrix represents the logical structure of the text after removing a specific correlation feature.
[0198] S2055. Calculate the sub-correlation strength corresponding to the sub-correlation matrix according to each of the sub-correlation matrices.
[0199] For each generated sub-correlation matrix, the server recalculates its sub-correlation strength. The purpose of this step is to evaluate whether the logical coherence of the rest of the text is improved after removing a certain correlation feature. By comparing the sub-correlation strengths of different sub-correlation matrices, the server can identify which correlation features have the greatest impact on the logical coherence of the text.
[0200] S2056: Use the correlation feature deleted from the sub-correlation matrix with the highest sub-correlation strength as the target correlation feature.
[0201] After comparing the sub-correlation strengths of all sub-correlation matrices, the server will select the sub-correlation matrix with the highest sub-correlation strength. The correlation feature deleted from this sub-correlation matrix will be considered the target correlation feature because it has the greatest impact on the logical coherence of the text. After identifying the target correlation feature, the server has found the key factor that may cause the text logic conflict.
[0202] S2057: Use the text segment corresponding to the target correlation feature as the field to be corrected.
[0203] Finally, the server will mark the text segments corresponding to the target relevance features as fields to be corrected. These fields are the parts of the text where logical conflicts are most obvious and need to be corrected the most. Through this step, the server successfully locates the errors or unreasonableness in the text, providing a clear target for subsequent correction work.
[0204] S206. Correct the field to be corrected through a preset field correction knowledge graph to ensure the accuracy and appropriateness of the text information.
[0205] After determining the fields that need to be corrected, the server will automatically correct them using a preset field correction knowledge graph. This knowledge graph contains rich language rules and corpora, which can provide accurate correction suggestions to the server. For example, when the server detects a typo or grammatical error, it will find the most appropriate replacement word or correction method based on the rules and corpora in the knowledge graph. In this way, the server can automatically correct the erroneous fields in the text to ensure the accuracy and appropriateness of the text information.
[0206] The method provided in the embodiment of the present application obtains the first correlation feature between the word segment and the sentence, the second correlation feature between the sentence and the paragraph, and the third correlation feature between the word segment and the paragraph according to the internal connection between the word segment, the sentence, and the paragraph. The correlation matrix of the text information is constructed by the first correlation feature, the second correlation feature, and the third correlation feature, and the correlation strength between the word segment, the sentence, and the paragraph is calculated according to the correlation matrix. If the correlation strength is lower than the preset strength threshold, any correlation feature in the correlation matrix is deleted respectively to generate each sub-correlation matrix. The correlation feature deleted from the sub-correlation matrix with the highest sub-correlation strength is used as the target correlation feature, and the text segment corresponding to the target correlation feature is used as the field to be corrected, thereby improving the accuracy of the extracted text information.
[0207] Optional, Figure 3 A flow chart of another logistics delivery notification method provided in an embodiment of the present application. Figure 3 As shown, the method may also include:
[0208] S301: Obtain the historical interaction records of the user.
[0209] The historical interaction record includes the user's historical receiving intentions.
[0210] The server can retrieve and extract the user's historical interaction records from the database. These records may include the user's previous phone call records, SMS interactions, online chat records, etc., which contain the user's historical delivery intentions. For example, the user may have clearly stated in a past call with customer service that he or she wants to receive the package at a specific time, or prefers a specific delivery method. The server will organize these historical interaction records into structured data for subsequent analysis. In addition, the server will clean and preprocess these historical interaction records, remove redundant information, and extract key data to ensure the accuracy and effectiveness of the data.
[0211] For example, suppose user A has interacted with the customer service of a logistics company over the phone many times in the past few months, and these call records are properly stored in the database by the server. Now, the server needs to obtain user A's historical interaction records in order to analyze his historical delivery intentions. The server retrieves all of user A's call records from the database and extracts key information such as call time, call content, and user requirements. By analyzing these records, the server can learn that user A usually likes to receive packages in the afternoon and prefers to pick up the packages in person rather than have them delivered to the door.
[0212] S302. Input the historical interaction records and the call voice data into a trained intention inference model, and output the probability data of multiple user intentions corresponding to the historical interaction records and the call voice data.
[0213] The server can input the user's historical interaction records and current call voice data into a trained intent inference model. This model is trained with a large amount of historical data and can identify and analyze the user's intent. The model processes the input data and outputs multiple possible user intents and their corresponding probability data. These data reflect the model's prediction and judgment of the user's current intent.
[0214] For example, the server inputs user A's historical interaction records and current call voice data into the intention inference model. After calculation, the model outputs several possible user intentions and their probabilities, such as the intention probability of "pick up the package in the afternoon" is 0.7, the intention probability of "home delivery" is 0.2, and the intention probability of "change the delivery address" is 0.1. These probability data provide the server with clues about user A's current possible intentions.
[0215] Specifically, the training process of the intention reasoning model includes:
[0216] Acquire multiple historical interaction record samples of multiple users, and convert the historical interaction record samples into standardized text, wherein the historical interaction record samples include interaction content, timestamp, user behavior, and user real intention, wherein the user real intention includes explicit intention and / or implicit intention.
[0217] The server can collect a large number of historical interaction record samples from the system. These samples cover the interaction data of different users in different scenarios, including text conversations, voice records, etc. The server will pre-process these raw data and convert them into a unified standardized text format. During the conversion process, the server will extract and record the interaction content, timestamp, user's specific behavior, and the user's true intention. It is particularly important that the user's true intention includes not only the explicitly expressed display intention, but also the implicit intention implicitly inferred through user behavior and interaction content.
[0218] A training set for the intent inference model is generated based on the standardized text, and the user's real intent is used as label data. The server can use these standardized texts and the user's real intent to build a training data set. In this data set, each standardized text will be matched with its corresponding user's real intent, and these real intents will be used as label data in the training process. Doing so can help the model learn the mapping relationship between text and intent during training.
[0219] The standardized text is input into the intent inference model to train the intent inference model. The server can input this training data set into the intent inference model, and through repeated iterations and optimization of the model's parameters, the model can accurately predict the user's intent from the input text. This process usually involves complex machine learning algorithms, such as deep learning, through which the model can gradually learn how to extract useful information from text features and judge the user's intent accordingly.
[0220] S303: If the number of probability data of the multiple user intentions that is within the target probability interval is greater than or equal to a preset number threshold, the user's deep receiving intention will be determined based on similar features of the multiple user intentions.
[0221] The server can set a target probability interval, such as between 0.6 and 1, and a preset quantity threshold, such as 2. When the intent inference model outputs multiple user intents with two or more intent probabilities falling within the target probability interval, the server will further analyze the similar features of these high-probability intents. By comparing the commonalities and differences of these intents, the server can comprehensively judge and determine the user's deep intention to receive the email. This deep intention may not be expressed directly, but is inferred through the user's words, deeds, and historical habits.
[0222] For example, in the case of user A, if the probabilities of the two intentions "pick up the package in the afternoon" and "deliver to home" are both within the target probability range (assuming they are both above 0.7), and the number reaches the preset threshold (here is 2), then the server will further analyze the similar features of the two intentions. Although they appear to be opposite intentions (one for pick up and the other for delivery), the server may combine user A's historical habits (such as frequently choosing to pick up the package) and certain hints in the current call content (such as asking about the business hours of the pick-up point), and ultimately determine that user A's deep intention to receive the package may be "pick up the package at a certain time in the afternoon." In this way, the server can meet the user's needs more accurately.
[0223] The method provided in the embodiment of the present application obtains the historical interaction records of the user, inputs the historical interaction records and the call voice data into the trained intention inference model, and outputs the probability data of multiple user intentions corresponding to the historical interaction records and the call voice data. If the number of probability data of the multiple user intentions that are within the target probability interval is greater than or equal to the preset number threshold, the user's deep receiving intention will be determined based on the similar features of the multiple user intentions, thereby accurately identifying the user's deep receiving intention and improving the user interaction experience and efficiency.
[0224] Optional, Figure 4 A flow chart of another logistics delivery notification method provided in an embodiment of the present application. Figure 4 As shown, the method may also include:
[0225] S401: Obtain the user's historical interaction records and historical logistics information.
[0226] The server accesses the internal database to retrieve and obtain the historical interaction records and historical logistics information of a specific user. The historical interaction records may include chat records between the user and customer service, phone call records, complaints and suggestions, etc., while the historical logistics information covers the user's shipping and receiving records, delivery address, delivery time, etc. This information provides the server with basic data for analyzing user behavior and preferences.
[0227] S402: Extract the user's preference characteristics based on the historical interaction records and the historical logistics information.
[0228] Among them, the preference characteristics include time characteristics, communication method characteristics, and receiving method characteristics. The time characteristics include communication time preference and receiving time preference. The communication method characteristics include language style and communication habits.
[0229] The server uses data analysis algorithms to conduct in-depth mining of historical interaction records and historical logistics information to identify user preference characteristics. These characteristics include but are not limited to time characteristics (such as when users prefer to communicate or receive parcels), communication method characteristics (such as the user's language style and communication habits), and delivery method characteristics (such as whether the user prefers to pick up the parcel in person or have it delivered to the door).
[0230] S403: Determine a time period for automatically dialing the user's phone number according to the communication time preference.
[0231] Based on the user's communication time preference extracted in step S402, the server will set a time policy for automatically dialing the phone. This policy can be a time range, such as 6 pm to 9 pm, or a specific time point, such as 3 pm every Wednesday. The server will configure a task scheduling system, such as a task scheduler, to ensure that the call is automatically performed at the set time.
[0232] S404: Determine the receipt time inquiry process in the preset speech process according to the receipt time preference.
[0233] The server adjusts the conversation process based on the user's preferred delivery time. In particular, when asking about the delivery time, it will provide time options that match the user's preferences.
[0234] In the script process, the inquiry about the delivery time will be customized according to the user's delivery time preference. For example, if the user usually receives the delivery after 5 pm, the script may include a question like "Would you like to have the delivery after 5 pm as usual?" The server will dynamically generate this part of the script based on the user's preference.
[0235] S405. Determine the receiving method inquiry process in the preset speech process according to the receiving method characteristics.
[0236] The sales pitch will be adjusted accordingly based on the user's delivery method preferences. If the user is used to self-collection, the sales pitch will focus on the location of the self-collection point and the pickup process; if the user prefers home delivery, the sales pitch will emphasize the accuracy and timeliness of delivery. The server will dynamically adjust the content of this part of the sales pitch based on the user's delivery method characteristics.
[0237] S406: Determine the communication tone and / or communication emotion used in the preset speech process according to the language style and the communication habit.
[0238] In order to communicate more naturally with users, the server will adjust the tone and emotion of the speech according to the user's language style and communication habits. For example, for users with a formal language style, the speech will use more polite and professional wording; while for users with a casual language style, the speech may be more colloquial and friendly. This requires natural language processing and sentiment analysis technology to achieve.
[0239] S407: adjusting the preset speech process according to the receipt time inquiry process, the receipt method inquiry process, and the communication tone and / or communication emotion.
[0240] Based on the analysis results of the above steps, the server will make comprehensive adjustments to the preset speech process. This includes the order, content, tone, emotion and other aspects of the speech. The adjusted speech will be more in line with the user's personality and preferences, thereby improving communication effectiveness and user experience.
[0241] S408. Acquire the recipient's call voice data through the adjusted preset speech process.
[0242] When talking to a user, the server will enable the voice recording function to capture the voice data during the call. This data will be stored on the server and can be used for subsequent analysis and mining. For example, the voice can be converted into text through voice recognition technology to further analyze the user's needs and feedback. At the same time, this voice data can also be used to train and optimize the server's voice recognition and speech synthesis models to improve the quality and efficiency of future calls.
[0243] The method provided in the embodiment of the present application obtains the historical interaction records and historical logistics information of the user. Based on the historical interaction records and the historical logistics information, the preference characteristics of the user are extracted, and based on the preference characteristics, the time period for automatically dialing the user's phone, the receipt time inquiry process in the preset speech process, the receipt method inquiry process, and the communication tone and / or communication emotion used in the preset speech process are determined. According to the receipt time inquiry process, the receipt method inquiry process, and the communication tone and / or communication emotion, the preset speech process is adjusted, and the recipient's call voice data is obtained through the adjusted preset speech process, thereby improving user experience and communication efficiency.
[0244] Figure 5 The structure diagram of an electronic device provided in an embodiment of the present application. The electronic device is used to execute the logistics delivery notification method mentioned above, for example, it can be the server mentioned above. Figure 5 As shown, the electronic device 500 may include: at least one processor 501 , a memory 502 , and a communication interface 503 .
[0245] The memory 502 is used to store programs. Specifically, the programs may include program codes, and the program codes include computer operation instructions.
[0246] The memory 502 may include a high-speed RAM memory, and may also include a non-volatile memory, such as at least one disk memory.
[0247] The processor 501 is used to execute the computer-executable instructions stored in the memory 502 to implement the method described in the above method embodiment. The processor 501 may be a CPU, or an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of the present application.
[0248] The processor 501 can communicate and interact with external devices through the communication interface 503. The external device can be, for example, the user's terminal device mentioned above, or a server of the express service system. In a specific implementation, if the communication interface 503, the memory 502 and the processor 501 are implemented independently, the communication interface 503, the memory 502 and the processor 501 can be interconnected through a bus and communicate with each other. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component (PCI) bus or an Extended Industry Standard Architecture (EISA) bus. The bus can be divided into an address bus, a data bus, a control bus, etc., but it does not mean that there is only one bus or one type of bus.
[0249] Optionally, in a specific implementation, if the communication interface 503, the memory 502 and the processor 501 are integrated on a chip, the communication interface 503, the memory 502 and the processor 501 can communicate through an internal interface.
[0250] The present application also provides a computer-readable storage medium, which may include: a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a disk or an optical disk, and other media that can store program codes. Specifically, the computer-readable storage medium stores program instructions, and the program instructions are used for the methods in the above embodiments.
[0251] The present application also provides a program product, which includes an execution instruction, which is stored in a readable storage medium. At least one processor of a computing device can read the execution instruction from the readable storage medium, and at least one processor executes the execution instruction so that the computing device implements the above method.
[0252] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present application.
Claims
1. A logistics delivery notification method, characterized in that: The method comprises: Obtaining logistics information to be notified from the express delivery service system, wherein the logistics information to be notified includes the recipient's contact information; Automatically dial the recipient's phone number according to the recipient's contact information; Identify the ringback tone of the recipient's phone during the automatic dialing process, and determine the connection type of the recipient's phone according to the ringback tone, wherein the connection type includes connected and unconnected, and the types of the ringback tone include ringing type, no answer type, call in progress type, power off type, out of service type, and unavailable number type; If the connection type is not connected, the recipient's phone number is automatically replayed according to a preset time interval and / or replay cycle number; If the connection type is connected, the call voice data of the recipient is obtained through a preset speech process; Identify the call voice data and determine the recipient's receiving intention; The delivery method of the logistics is determined according to the recipient's receiving intention.
2. The method according to claim 1, characterized in that The step of determining the connection type of the recipient's phone according to the ringback tone includes: If the ringback tone is of the middle ringing type and the call is connected within the first preset duration, the connection type is connected; If the ringback tone is the ringing type and the call is still not connected after reaching the first preset time, or the ringback tone is any one of the no-answer type, call in progress type, power-off type, out-of-service type, and unavailable number type, then the connection type is not connected.
3. The method according to claim 1, characterized in that The identifying the call voice data and determining the recipient's receiving intention includes: Processing the call voice data through a trained deep learning model to obtain a spectrogram of the user in the call voice data; Extracting a speech feature sequence of the user's speech from the spectrogram, and decoding the speech feature sequence to generate a phoneme sequence of the user's speech, wherein the phoneme sequence is used to determine the content of the user's speech; Combining the phoneme sequence into text information corresponding to the call voice data using a language model; Inputting the text information into an intention recognition model to extract an intention recognition feature vector of the text information; Determine the receiving intention corresponding to the text information by performing classification processing on the intention recognition feature vector, wherein the receiving intention includes a delivery method and a delivery time; The step of processing the call voice data by using a trained deep learning model to obtain a spectrogram of the user in the call voice data includes: The trained deep learning model is used to extract audio features of the call voice data, wherein the audio features include frequency features, time features, and amplitude features. The frequency features represent energy distribution of the audio signal of the call voice data at different frequencies, the time features represent temporal changes of the audio signal of the call voice data, and the amplitude features represent the intensity of the audio signal of the call voice data. The frequency feature, the time feature, and the amplitude feature are concatenated to generate an audio feature matrix; Converting the time domain signal in the audio feature matrix into a frequency domain signal based on discrete cosine transform to obtain the frequency spectrum feature of the call voice data; Mapping the characteristic values in the frequency spectrum characteristics to the horizontal axis of the initial spectrogram to generate an initial spectrogram of the user in the call voice data, wherein the horizontal axis of the initial spectrogram represents time, the vertical axis represents frequency, and the brightness or color of each pixel represents the energy at the corresponding time and frequency; The initial spectrogram is smoothed based on Gaussian filtering to generate a spectrogram of the user in the call voice data.
4. The method according to any one of claims 1 to 3, characterized in that: Before inputting the text information into the intention recognition model, the method further includes: Detecting the text information using a deep learning model and extracting error recognition features of the text information, wherein the error recognition features include text sequence features, context association features, and semantic features; Based on the text sequence features, contextual features, and semantic features, output the error probability of each text segment in the text information, wherein the text segment includes a word, a sentence, and a paragraph of the text information; Obtaining an adjustment weight of each text segment according to the appearance frequency and importance ratio of each text segment in the text information; According to the error probability of the word segmentation and the adjusted probability generated by the adjustment weight of the word segmentation, as well as the word segmentation error probability threshold, the error probability of the sentence and the adjusted probability generated by the adjustment weight of the sentence, as well as the sentence error probability threshold, the error probability of the paragraph and the adjusted probability generated by the adjustment weight of the paragraph, and the paragraph error probability threshold, respectively determine the type of the word segmentation, the type of the sentence, and the type of the paragraph; the types include a type to be corrected and a correct type; According to the type of the word segmentation, the type of the sentence, and the type of the paragraph, a fusion calculation is performed to obtain the field to be corrected in the text information; The fields to be corrected are corrected through a preset field correction knowledge graph to ensure the accuracy and appropriateness of the text information.
5. The method according to claim 4, characterized in that The step of obtaining the field to be corrected in the text information by fusing and calculating according to the type of the word segmentation, the type of the sentence, and the type of the paragraph includes: According to the internal connection between the participle, the sentence, and the paragraph, a first correlation feature between the participle and the sentence, a second correlation feature between the sentence and the paragraph, and a third correlation feature between the participle and the paragraph are obtained, wherein the correlation features are used to indicate whether there is a logical conflict between the participle, the sentence, and the paragraph; Constructing a correlation matrix of the text information through the first correlation feature, the second correlation feature, and the third correlation feature; Calculate the association strength between the word segmentation, the sentence, and the paragraph according to the association matrix; If the correlation strength is lower than a preset strength threshold, any correlation feature in the correlation matrix is deleted to generate sub-correlation matrices; Calculating the sub-correlation strength corresponding to the sub-correlation matrix according to each of the sub-correlation matrices; Using the correlation feature deleted from the sub-correlation matrix with the highest sub-correlation strength as the target correlation feature; Taking the text segment corresponding to the target correlation feature as the field to be corrected; The obtaining, according to the intrinsic relationship between the segmentation, the sentence, and the paragraph, a first correlation feature between the segmentation and the sentence includes: Using a word vector model, converting the word segmentation included in the sentence into a first word segmentation vector; Generate a sentence vector corresponding to the sentence according to the first word segmentation vector, where the sentence vector is generated based on an average value of the first word segmentation vector; Using a word vector model, converting the word segmentation into a second word segmentation vector; The cosine similarity between the sentence vector and the second word segmentation vector is calculated based on a similarity analysis method to generate a first correlation feature between the word segmentation and the sentence.
6. The method according to any one of claims 1 to 3, characterized in that: The method further comprises: Acquire the user's historical interaction record, wherein the historical interaction record includes the user's historical receiving intention; Inputting the historical interaction records and the call voice data into a trained intention inference model, and outputting probability data of multiple user intentions corresponding to the historical interaction records and the call voice data; If the number of probability data of the multiple user intentions that are within the target probability interval is greater than or equal to a preset number threshold, the user's deep receiving intention will be determined based on similar features of the multiple user intentions; The training process of the intention reasoning model specifically includes: Acquire multiple historical interaction record samples of multiple users, and convert the historical interaction record samples into standardized text, wherein the historical interaction record samples include interaction content, timestamp, user behavior, and user real intention, wherein the user real intention includes explicit intention and / or implicit intention; Generate a training set for the intention inference model according to the standardized text, using the user's real intention as label data; The standardized text is input into the intention reasoning model to train the intention reasoning model.
7. The method according to any one of claims 1 to 3, characterized in that: The method further comprises: Obtaining the user's historical interaction records and historical logistics information; Extracting the user's preference features according to the historical interaction records and the historical logistics information, wherein the preference features include time features, communication mode features, and delivery mode features. The time features include communication time preference and delivery time preference. The communication mode features include language style and communication habits. Determining a time period for automatically dialing the user's phone according to the communication time preference; According to the receiving time preference, determining the receiving time inquiry process in the preset speech process; According to the characteristics of the receiving method, determining the receiving method inquiry process in the preset speech process; Determining the communication tone and / or communication emotion adopted by the preset speech process according to the language style and the communication habits; Adjust the preset speech process according to the receipt time inquiry process, the receipt method inquiry process, and the communication tone and / or communication emotion; The recipient's call voice data is obtained through the adjusted preset speech process.
8. An electronic device, characterized in that: include: A processor, a communication interface and a memory, wherein the processor is communicatively connected to the communication interface and the memory respectively; The memory stores computer-executable instructions; The communication interface communicates and interacts with external devices; The processor executes the computer-executable instructions stored in the memory to implement the method according to any one of claims 1 to 7.
9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-executable instructions, which are used to implement the method according to any one of claims 1 to 7 when executed by a processor.
10. A computer program product, characterized in that The invention comprises a computer program, which implements the method according to any one of claims 1 to 7 when being executed by a processor.
