Customer service auxiliary decision-making system based on knowledge base

Abstract

The invention discloses a knowledge base-based customer service auxiliary decision-making system, and relates to the technical field of customer service auxiliary decision-making, and the system comprises the steps: carrying out the voice activity detection of incoming call audio, and calculating the signal-to-noise ratio and one-way time delay of a current session; based on a pre-calibrated selection time parameter set and a logarithmic relationship, obtaining selection time under different candidate suggested number; calculating a re-speaking time cost in combination with the signal-to-noise ratio, and determining a time delay penalty according to the one-way time delay; in a preset suggested number range, the selection time, the re-speaking time cost, the one-way time delay and the time delay penalty are synthesized into a unified time cost, and the hit income which is progressively decreased from margins along with the increase of the suggested number is introduced to determine the optimal suggested number; and shrinking the optimal number of suggestions under the session comfort time delay constraint to obtain a final number of suggestions, and retrieving and intercepting a suggestion set consistent with the final number of suggestions from the knowledge base document. According to the method, the suggestion scale can be adaptively controlled when the noise environment and the time delay condition dynamically change, and the customer service handling efficiency is improved.

Description

Technical Field

[0001] This invention relates to the field of customer service decision support technology, and more particularly to a knowledge base-based customer service decision support system. Background Technology

[0002] In existing intelligent customer service systems, users primarily initiate business inquiries via voice. After the voice signal is collected at the front end, it is transcribed into text by a speech recognition module. A knowledge-based retrieval module then returns several candidate answers for agents or automated processes to choose from. This process is typically deployed in complex acoustic environments such as call centers, service halls, and open-plan offices. Interference sources such as air conditioning noise, keyboard clicks, background voices, and equipment noise cause the signal-to-noise ratio of the voice signal to fluctuate continuously during the conversation. Simultaneously, customer service systems often operate in multi-link distributed network environments. Network congestion and routing changes can lead to fluctuations in end-to-end transmission latency, making it difficult to maintain stable single-trip latency. Under these circumstances, agents need to read, compare, and confirm multiple suggestions within a limited conversation time. Too many suggestions can prolong selection time, while too few suggestions may reduce the hit rate of responses and increase the number of supplementary retrieval rounds, directly impacting service efficiency and conversation coherence.

[0003] Current common customer service assistance solutions often control the return scale by using a fixed number of suggestions or simple rules based on experience thresholds. They typically only refer to historical hit rates or business configuration parameters, lacking a unified characterization of the current conversation's signal-to-noise ratio, speech recognition quality, and network one-way latency. On the one hand, when environmental noise increases and the signal-to-noise ratio decreases, speech recognition errors and repetitions increase, yet the system still returns a large number of suggestions at a fixed rate. This places an additional reading and judgment burden on agents under unfavorable noise conditions, leading to a significant increase in overall processing time. On the other hand, when one-way latency approaches or exceeds the comfort limit of the conversation, existing solutions have not yet transformed latency factors into quantifiable latency penalties, nor have they constrained or reduced the number of suggestions accordingly, making it difficult to improve the user's subjective perception of slow response. Summary of the Invention

[0004] The purpose of this invention is to solve the problem in the prior art that it is impossible to adaptively determine a reasonable number of suggestions when there are noisy environments and network latency changes dynamically, and to propose a customer service auxiliary decision-making system based on a knowledge base.

[0005] To address the problems existing in the prior art, the present invention adopts the following technical solution: Knowledge-based customer service decision support systems include: The observation and measurement module is used to perform speech segmentation on the incoming audio stream, calculate the signal-to-noise ratio based on the speech segmentation results, and determine the one-way delay based on the end-to-end time difference. The selection time module is used to calculate the selection time for the number of candidate suggestions based on a pre-calibrated set of selection time parameters. The re-speech cost module is used to calculate the re-speech time cost based on the signal-to-noise ratio and a pre-calibrated set of identification parameters. The delay penalty module is used to set the out-of-bounds delay based on the one-way delay and determine the delay penalty based on the out-of-bounds delay; The optimal number of suggestions module is used to determine the optimal number of suggestions based on one-way delay, selection time, repetition time cost, and delay penalty. The budget contraction module is used to calculate the maximum number of suggestions within the time budget, and determine the final number of suggestions based on the maximum number of suggestions and the optimal number of suggestions. The retrieval and cropping module is used to truncate documents in the knowledge base based on the final number of suggestions, forming a suggestion set.

[0006] Preferably, the incoming audio stream is segmented into speech segments, the signal-to-noise ratio is calculated based on the speech segmentation results, and the one-way delay is determined according to the end-to-end time difference, including: The incoming audio stream is subjected to voice activity detection by a preset energy determination threshold, and the audio stream is divided into a voice segment sample set and a silence segment sample set. Calculate the mean square energy of the speech segment sample set and the mean square energy of the silence segment sample set respectively; The signal-to-noise ratio of the current session is obtained by taking the logarithm of the ratio of the mean square energy of the speech segment sample set to the mean square energy of the silence segment sample set. Collect the transmission time and corresponding reception time of the incoming call audio; The difference between the sending time and the receiving time is used as the one-way delay.

[0007] Preferably, the selection time for calculating the number of candidate suggestions based on a pre-defined set of selection time parameters includes: Obtain a pre-calibrated set of selection time parameters, including: fixed overhead, time slope baseline, noise expansion coefficient, and reference signal-to-noise ratio; Based on the difference between the current signal-to-noise ratio and the reference signal-to-noise ratio, the time slope baseline is monotonically adaptively corrected by combining the noise expansion coefficient to obtain the noise-corrected time slope. Construct a selection time function based on fixed overhead and a noise-corrected time slope; Within the preset number of suggestions, for each number of candidate suggestions, the selection time corresponding to each number of candidate suggestions is calculated using a selection time function.

[0008] Preferably, the time cost of repeating the message is calculated based on the signal-to-noise ratio and a pre-calibrated set of identification parameters, including: Obtain a pre-calibrated set of recognition parameters, including: a recognition quality threshold and a recognition cost coefficient; If the signal-to-noise ratio is lower than the recognition quality threshold, the difference between the recognition quality threshold and the signal-to-noise ratio is used as the gap; otherwise, the gap is set to zero. The time cost of repeating the information is obtained by multiplying the gap amount by the identification cost coefficient.

[0009] Preferably, the out-of-bounds delay is set based on the one-way delay, and the delay penalty is determined based on the out-of-bounds delay, including: If the one-way latency is greater than the preset session comfort latency limit, the difference between the one-way latency and the session comfort latency limit will be used as the out-of-bounds latency; otherwise, the out-of-bounds latency will be set to zero. The delay penalty is obtained by multiplying the out-of-bounds delay by the preset delay out-of-bounds penalty coefficient.

[0010] Preferably, the optimal number of suggestions is determined based on one-way delay, selection time, repetition time cost, and delay penalty, including: Add the selection time to the repetition time cost, one-way delay, and delay penalty to obtain the total time cost corresponding to the number of candidate suggestions; Based on the preset suggestion hit benefit coefficient and benefit saturation speed parameters, the hit benefit that marginally decreases as the number of candidate suggestions increases is calculated. Subtracting the hit benefit from the total time cost yields the target value corresponding to the number of candidate suggestions; Within the preset number of suggestions, the number of candidate suggestions with the smallest target value is taken as the optimal number of suggestions.

[0011] Preferably, within the time budget, the maximum number of suggestions is calculated, and the final number of suggestions is determined based on the maximum number of suggestions and the optimal number of suggestions, including: Subtract one-way latency, repetition time cost, and fixed overhead from the upper limit of comfortable conversation latency to obtain the time budget; The maximum number of suggestions within the time budget is calculated using the inverse transformation of the logarithmic relationship. The maximum number of suggestions is compared with the optimal number of suggestions, and the smaller value is selected as the final number of suggestions.

[0012] Preferably, the documents in the knowledge base are truncated according to the final number of suggestions to form a suggestion set, including: Retrieve the user's speech-to-text query content and all documents in the knowledge base; A probabilistic relevance scoring method based on term frequency and inverse document frequency combined with document length normalization is used to calculate the relevance score of each document. Sort all knowledge base documents from highest to lowest relevance score; From the sorted document sequence, extract the preceding documents with the same number as the final number of suggestions to form a suggestion set.

[0013] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention achieves this by segmenting incoming audio into speech, measuring signal-to-noise ratio and one-way latency, calculating the time required to select candidate suggestions, estimating the time cost of repeating, quantifying latency penalties based on one-way latency, and optimizing and shrinking the number of suggestions under time budget constraints. Ultimately, it retrieves and extracts documents from the knowledge base based on this information. This transforms the determination of the number of suggestions from a static configuration into a dynamic decision-making process based on time cost and hit benefit. It unifies the impact of noise environment, speech recognition performance, and network latency on customer service interaction efficiency within the same time measurement framework, effectively solving the problem of not being able to adaptively determine a reasonable number of suggestions when noise environment and network latency change dynamically. This ensures dialogue continuity while improving overall processing efficiency.

[0014] 2. This invention divides incoming audio into speech segments and silence segments through speech activity detection. It calculates the signal-to-noise ratio (SNR) of the current session based on the logarithm of the ratio of the mean square energy of the speech and silence segments. This SNR is used as a common input for the selection time function and the calculation of the repetition time cost. On one hand, based on pre-calibrated fixed overhead, time slope baseline, noise expansion coefficient, and reference SNR, the time slope is monotonically adaptively corrected to obtain a selection time that reflects changes in decision-making speed under noisy conditions. On the other hand, based on the recognition quality threshold and recognition cost coefficient, SNR gaps below the threshold are converted into linear repetition time costs. This automatically increases the time cost weight when noise increases and reduces the repetition penalty when noise is low. This ensures that the optimization of the number of candidate suggestions truly reflects the comprehensive impact of the speech environment on recognition quality and selection efficiency, avoiding increased repetition and reading burden caused by returning suggestions with a fixed number of suggestions.

[0015] 3. This invention defines the difference between the single-trip latency and a set comfortable latency limit as the out-of-bounds latency by setting a limit. It then calculates the latency penalty using a pre-calibrated latency out-of-bounds penalty coefficient. Simultaneously, it constructs a hit-benefit model that marginally decreases with the number of candidate suggestions based on the suggestion hit benefit coefficient and benefit saturation speed parameter. The optimal number of suggestions is obtained by comparing the target values ​​of all candidate suggestions. Within the time budget, the maximum suggestion limit is calculated using the inverse transformation of the logarithmic relationship. Finally, the smaller of the two values ​​is taken as the final suggestion number to control the scale of the knowledge base retrieval results. This achieves adaptive control of the suggestion set size under dynamically changing noise and latency conditions, while ensuring that the latency of a single interaction does not exceed the comfort threshold, thus improving the hit rate and selection time control. This enhances customer service response speed and user consultation experience. Attached Figure Description

[0016] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0017] Figure 1 This is a functional module diagram of a knowledge base-based customer service auxiliary decision-making system provided in an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0019] Example: This example provides a knowledge-based customer service decision support system. See [link to example]. Figure 1 Specifically, including: The observation and measurement module is used to perform speech segmentation on the incoming audio stream, calculate the signal-to-noise ratio based on the speech segmentation results, and determine the one-way delay based on the end-to-end time difference. In an embodiment of the present invention, the incoming audio stream is segmented into speech segments, the signal-to-noise ratio is calculated based on the speech segmentation results, and the one-way delay is determined based on the end-to-end time difference, including: The incoming audio stream is subjected to voice activity detection by a preset energy determination threshold, and the audio stream is divided into a voice segment sample set and a silence segment sample set. Specifically, the incoming audio stream is preprocessed, with uniform sampling rate and frame length parameters set. The sampling rate is 16kHz, a common practice in speech processing, and the frame length is set to 20 to 50 milliseconds. The frame shift is half the frame length to ensure inter-frame continuity. Audio data is extracted frame by frame using a sliding window traversal method, and the mean square energy of each frame is calculated as the energy feature of that frame. The preset energy judgment threshold is determined based on the statistical characteristics of silent segment audio data in historical customer service scenarios. Specifically, a large number of background noise samples from customer service environments without effective speech are collected, and their mean and standard deviation are calculated. The threshold is set to the mean plus 3 to 5 times the standard deviation. This setting is based on the ability to balance noise suppression effect and speech false negative rate, ensuring accurate differentiation between speech and silence under common customer service environment noise levels. The energy characteristics of each frame of audio data are compared with a preset energy threshold. If the energy characteristics of a frame are greater than the threshold, it is marked as a speech frame; if they are less than or equal to the threshold, it is marked as a silence frame. Subsequently, consecutively marked speech frames are merged to form consecutive speech segments, and consecutively marked silence frames are merged to form consecutive silence segments. At the same time, isolated speech frames or silence frames are smoothed and merged into adjacent longer segments to avoid segmentation errors caused by single-frame judgment errors. Finally, a speech segment sample set composed of consecutive speech frames and a silence segment sample set composed of consecutive silence frames are obtained.

[0020] Calculate the mean square energy of the speech segment sample set and the mean square energy of the silence segment sample set respectively; The signal-to-noise ratio of the current session is obtained by taking the logarithm of the ratio of the mean square energy of the speech segment sample set to the mean square energy of the silence segment sample set. Collect the transmission time and corresponding reception time of the incoming call audio; The difference between the sending time and the receiving time is used as the one-way delay; Specifically, the signal-to-noise ratio (SNR) is a technical indicator that quantifies the difference in intensity between the speech signal and the ambient noise. It is calculated based on the speech segment sample set and the silence segment sample set after the incoming audio stream is segmented by speech activity detection. It is obtained by performing a ten-fold logarithmic operation on the ratio of the mean square energy of the speech segment sample set to the mean square energy of the silence segment sample set. It can objectively reflect the noise interference level of the current conversation environment and is a core input parameter for correcting the selection time slope and calculating the repeat time cost, directly affecting the dynamic optimization result of the number of candidate suggestions. One-way latency refers to the one-way transmission time of incoming audio data from the client to the server. It is calculated by the system time after the client and server are synchronized by the network time protocol. Specifically, it is the arithmetic mean of the difference between the reception time and the corresponding transmission time of each frame of audio data after outlier removal. It can accurately characterize the network transmission performance and is a key basis for calculating latency penalties and determining the perceptible latency budget. It is directly related to the rationality of the final suggestion set size and the naturalness of the interaction.

[0021] Specifically, the process iterates through all consecutive speech segments in the speech segment sample set, extracts all audio frame sample values ​​contained in each speech segment, calculates the square of each sample value, sums them, and then divides this sum by the total number of sample points in the speech segment sample set to obtain the mean square energy of the speech segment sample set. The same calculation logic is used to process the silence segment sample set. All consecutive silence segments are iterated through, all audio frame sample values ​​are extracted and their squares are calculated, and then divided by the total number of sample points in the silence segment sample set to obtain the mean square energy of the silence segment sample set. The mean square energy of the speech segment sample set is used as the numerator, and the mean square energy of the silence segment sample set is used as the denominator. The ratio between the two is calculated, and then a logarithmic operation is performed on this ratio at a factor of ten. The base of the logarithmic operation is set to 10, which conforms to the standard calculation specifications for signal-to-noise ratio in the field of audio signal processing, accurately quantifying the intensity difference between the speech signal and environmental noise, and finally obtaining the signal-to-noise ratio (SNR) in dB for the current session.

[0022] Specifically, the client and server synchronize their times using a network time protocol to ensure consistent time bases and eliminate delay calculation errors caused by time base deviations. When the client sends each frame of incoming audio data, it simultaneously records the transmission time of that frame, taken from the client's locally synchronized system time. Upon receiving each frame of incoming audio data, the server records the reception time of the corresponding frame, taken from the server's locally synchronized system time. Simultaneously, a one-to-one correspondence between transmission and reception times is established using frame sequence numbers or data checksums to avoid frame matching errors. For the current conversation turn, the sending and receiving times of 10 to 20 consecutive frames of audio data are collected, and the time difference of each frame is calculated as the instantaneous value of single-frame single-way latency. The preset outlier judgment range is 0 to 1 second. This range is determined based on the latency statistics characteristics of common customer service network environments and can cover normal latency fluctuations in most scenarios such as wired and wireless. Single-frame single-way latency instantaneous values ​​that exceed this range are judged as outliers and removed. Then, the arithmetic average of the remaining valid instantaneous values ​​is performed to finally obtain the single-way latency of the current session.

[0023] The selection time module is used to calculate the selection time for the number of candidate suggestions based on a pre-calibrated set of selection time parameters. In an embodiment of the present invention, the selection time for calculating the number of candidate suggestions based on a pre-defined set of selection time parameters includes: Obtain a pre-calibrated set of selection time parameters, including: fixed overhead, time slope baseline, noise expansion coefficient, and reference signal-to-noise ratio; Specifically, the selection time parameter set is obtained through offline calibration using historical customer service conversation data. The calibration process is based on at least one thousand customer service interaction records containing different noise environments, network conditions, and agent selection behaviors. Fixed overhead is the basic time cost for agents to perform the selection operation. It is determined by statistically analyzing the average preparation time from when an agent sees the suggestion set to when they begin selecting, including the time spent on basic actions such as eye movement to focus on the suggestion set and operating the device. The value ranges from 0.1 to 0.3 seconds, determined based on the physiological reaction time characteristics of human visual focus and simple operations. The time slope baseline is the basic rate at which selection time increases with the number of candidates when there is no significant noise interference. It is obtained by linearly fitting the logarithmic relationship between agent selection time and the number of candidates in historical data under high signal-to-noise ratio (above 30dB) scenarios. The value ranges from 0.1 to 0.3 seconds, conforming to Hick's law that selection time is proportional to the logarithm of the number of candidates. The noise expansion coefficient is used to quantify the amplification effect of noise on the selection time slope. It is obtained by comparing the deviations of measured seat selection times with baseline predicted values ​​under different signal-to-noise ratios (SNRs) using nonlinear least squares fitting, with a value ranging from 0.5 to 1.2. This range is determined based on the statistical characteristics of selection time extension in low SNR scenarios; the stronger the noise, the more significant the amplification effect of this coefficient on the slope. The reference SNR is a benchmark value for determining the degree of noise impact. A typical SNR of 25 to 30 dB in a noise-free laboratory environment is selected. This range corresponds to an environment with speech recognition accuracy higher than 95% and stable seat selection efficiency, ensuring a clear benchmark for quantifying the noise impact. These parameters are calibrated and stored in the system parameter configuration module, and can be directly read and obtained through the module interface during actual applications.

[0024] Based on the difference between the current signal-to-noise ratio and the reference signal-to-noise ratio, the time slope baseline is monotonically adaptively corrected by combining the noise expansion coefficient to obtain the noise-corrected time slope. Specifically, the difference between the current signal-to-noise ratio (SNR) and the reference SNR is calculated. When this difference is less than or equal to zero, it indicates that the current session environment noise is higher than the reference level, and the time slope baseline needs to be corrected. When the difference is greater than zero, it indicates that the current noise is lower than the reference level, and the time slope baseline remains unchanged. For scenarios requiring correction, the absolute difference between the reference SNR and the current SNR is first calculated. Then, this absolute difference is divided by the reference SNR to obtain the relative noise gap, which is used to quantify the proportion by which the current noise exceeds the reference level. The noise expansion coefficient is multiplied by the relative noise gap to obtain the amplification ratio of the noise on the slope. This amplification ratio is then added to 1 to form the complete correction coefficient. This correction coefficient increases monotonically with the increase of the relative noise gap, which conforms to the actual characteristic of prolonged agent selection time under low SNR. The time slope baseline is multiplied by the above correction coefficient to obtain the noise-corrected time slope, ensuring that the corrected slope can accurately reflect the rate of change of selection time with the increase of the number of candidates in a noisy environment. For scenarios that do not require correction, the time slope baseline is directly used as the noise-corrected time slope. The entire correction process is achieved through a monotonically increasing functional relationship, ensuring that the stronger the noise, the greater the slope after correction, which is consistent with the behavior of agents making decisions more slowly in high-noise environments.

[0025] Construct a selection time function based on fixed overhead and a noise-corrected time slope; Specifically, the selection time function uses the number of candidate suggestions as an input variable to quantify the time cost for agents to select target content from the suggestion set. This function consists of two parts: fixed overhead and variable terms. The fixed overhead, as a constant, represents the basic preparation time before the agent begins selection and does not change with the number of candidate suggestions. The selection time function is as follows:

[0026] In the formula, The number of candidate suggestions is When to choose the time For fixed expenses, The time slope is corrected for noise. The logarithmic term is based on Hick's law, reflecting the growth of selection time as the number of candidate suggestions increases; This function is designed based on Hick's Law, which states that the selection time increases logarithmically with the number of candidates. It can accurately reflect the impact of the increase in the number of candidate suggestions on the selection decision time. The function can output the corresponding selection time according to different numbers of candidate suggestions, and the dependent variable includes a time slope corrected for noise. It can adapt to the changes in the seat selection speed under different noise environments, ensuring that the estimated selection time matches the actual scenario.

[0027] Within the preset number of suggestions, for each number of candidate suggestions, the selection time corresponding to each number of candidate suggestions is calculated using a selection time function; Specifically, the preset number of suggestions ranges from a positive integer of 1 to 15. This range is determined based on the correlation between the number of candidate suggestions and the hit probability and selection time in historical customer service conversation data. Statistical analysis shows that when the number of suggestions exceeds 15, the marginal increase in hit probability is less than 0.5%, while the selection time increases by an average of more than 30%. Therefore, the range of 1 to 15 can cover the vast majority of effective candidate scenarios while avoiding invalid calculations. Each candidate suggestion number within this range is iterated over, and each number is sequentially substituted as an input variable into the constructed selection time function. The fixed overhead in the function is the pre-defined base preparation time, the noise-corrected time slope is an adaptive parameter under the current environment, and the logarithmic term is dynamically calculated as the number of candidate suggestions changes. For each number of candidate suggestions, the selection time is obtained by adding the fixed cost of the selection time function to the variable term. Selection time refers to the time consumed by the customer service agent to filter and determine the target suggestion that matches the user's needs from the set of candidate suggestions provided by the system. It is one of the core indicators for measuring the efficiency of customer service interaction. The time slope in the variable term, which is corrected for noise, ensures that the calculation results under different noise environments match the actual selection efficiency. Finally, a mapping relationship between the number of candidate suggestions and the corresponding selection time is formed, providing basic data for solving the optimal number of suggestions in the subsequent process.

[0028] The re-speech cost module is used to calculate the re-speech time cost based on the signal-to-noise ratio and a pre-calibrated set of identification parameters. In embodiments of the present invention, the calculation of the repeat time cost based on the signal-to-noise ratio and a pre-calibrated set of identification parameters includes: Obtain a pre-calibrated set of recognition parameters, including: a recognition quality threshold and a recognition cost coefficient; Specifically, the recognition parameter set is obtained through offline calibration using historical speech recognition data and customer service interaction records. The calibration process is based on at least 5,000 incoming audio data points containing different signal-to-noise ratios, user accents, and business scenarios. Each data point corresponds to a manually annotated speech recognition result and the actual re-speech time. The recognition quality threshold is a signal-to-noise ratio benchmark for determining whether the speech recognition quality meets basic interaction requirements. It is determined by analyzing the change curve of the speech recognition error rate under different signal-to-noise ratios. When the signal-to-noise ratio is higher than this threshold, the speech recognition error rate (word error rate) is stable below 5%, which can ensure that the customer service representative accurately understands the user's intent. When the signal-to-noise ratio is lower than this threshold, the error rate increases significantly as the signal-to-noise ratio decreases, leading to an increased probability of re-speech. Based on this, the recognition quality threshold is set to 15 to 25 dB, which covers the critical point of sudden change in the recognition error rate in most customer service scenarios. The recognition cost coefficient is a parameter that quantifies the repeat time cost corresponding to a unit signal-to-noise ratio (SNR) gap. It is obtained by statistically analyzing the average repeat time corresponding to different SNR gap amounts (the difference between the recognition quality threshold and the actual SNR) in historical data, and then using a linear regression method to fit the proportional relationship between the gap amount and the repeat time. The value ranges from 0.05 to 0.2 seconds per dB. This range is determined based on the actual growth rate of repeat time with increasing noise in low SNR scenarios, ensuring that the calculated repeat time cost is consistent with the actual interaction cost. These parameters, after calibration, are stored in the system's parameter configuration module and can be directly read through the module interface during practical applications, providing a foundation for real-time calculation of the recognition cost.

[0029] If the signal-to-noise ratio is lower than the recognition quality threshold, the difference between the recognition quality threshold and the signal-to-noise ratio is used as the gap; otherwise, the gap is set to zero. The time cost of repeating the story is obtained by multiplying the gap amount by the identification cost coefficient. Specifically, the signal-to-noise ratio (SNR) of the current session is first compared with a pre-defined recognition quality threshold, which is set between 15 and 25 dB. This threshold is determined based on the abrupt changes in speech recognition error rates in historical data. When the SNR is higher than this threshold, the recognition error rate stabilizes below 5%, while below it, the error rate increases significantly as the SNR decreases. If the current SNR is less than the recognition quality threshold, it indicates that the speech recognition quality does not meet basic interaction requirements. In this case, the difference between the recognition quality threshold and the current SNR is calculated, and this difference is defined as the gap. The size of the gap directly reflects the degree of recognition quality deficiency caused by noise. If the current SNR is greater than or equal to the recognition quality threshold, it indicates that the recognition quality is stable and the probability of repeating is extremely low. In this case, the gap is set to zero. After obtaining the gap amount, it is multiplied by a pre-calibrated identification cost coefficient, which ranges from 0.05 to 0.2 dB per second. This coefficient is obtained by linear regression fitting of the gap amount and the actual repeat time in historical data, which can accurately quantify the additional repeat time cost corresponding to a unit signal-to-noise ratio gap. The product of the two is the repeat time cost, which is used to characterize the time loss required for the user or customer service to repeat the statement due to the identification error, and provides identification-related quantitative parameters for the subsequent calculation of the total time cost.

[0030] The delay penalty module is used to set the out-of-bounds delay based on the one-way delay and determine the delay penalty based on the out-of-bounds delay; In an embodiment of the present invention, setting an out-of-bounds delay based on a one-way delay and determining a delay penalty based on the out-of-bounds delay includes: If the one-way latency is greater than the preset session comfort latency limit, the difference between the one-way latency and the session comfort latency limit will be used as the out-of-bounds latency; otherwise, the out-of-bounds latency will be set to zero. Specifically, the preset upper limit for comfortable conversation latency is set at 150 milliseconds. This value is determined according to the ITU-T G.114 international standard, which has been verified through numerous voice interaction user experience experiments. This standard clearly states that a single-trip latency within 150 milliseconds can ensure the natural fluency of the conversation. Exceeding this value will lead to a noticeable delay perceived by the user and affect interaction efficiency. The single-trip latency of the current session is compared with the 150-millisecond upper limit for comfortable conversation latency. If the single-trip latency is greater than 150 milliseconds, it indicates that the network transmission latency has exceeded the user's acceptable comfort range. In this case, the difference between the single-trip latency and 150 milliseconds is calculated, and this difference is defined as the out-of-bounds latency, used to quantify the degree to which the latency exceeds the comfort range. If the single-trip latency is less than or equal to 150 milliseconds, it indicates that the latency is within the user's perceived comfort range. In this case, the out-of-bounds latency is set to zero, representing no additional impact from exceeding the latency limit.

[0031] The delay penalty is obtained by multiplying the out-of-bounds delay by the preset delay out-of-bounds penalty coefficient; Specifically, the preset latency violation penalty coefficient is obtained through user experience surveys and historical interaction data calibration. This coefficient is used to quantify the negative impact of unit latency violation on interaction smoothness. During the calibration process, user subjective satisfaction ratings and customer service handling efficiency data under different latency violations were collected. By analyzing the rate of rating decline and the proportion of efficiency loss, the penalty coefficient was determined to be within the range of 1.0 to 3.0 seconds / second. A lower value (1.0 to 1.5 seconds / second) is used when the latency violation is small, corresponding to a slight impact of perceived delay; a higher value (2.0 to 3.0 seconds / second) is used when the latency violation is large, corresponding to significant interaction stuttering and efficiency loss caused by delay. When calculating the latency penalty, the determined latency violation is first obtained, then the preset latency violation penalty coefficient is read, and the two are multiplied. The result is the latency penalty. This penalty value directly reflects the additional loss of overall interaction cost due to the single-trip latency exceeding the comfort limit, providing quantitative parameters related to network latency for subsequent calculation of total time cost, ensuring that the objective function accurately weighs the negative impact of latency exceeding the limit.

[0032] The optimal number of suggestions module is used to determine the optimal number of suggestions based on one-way delay, selection time, repetition time cost, and delay penalty. In embodiments of the present invention, the optimal number of suggestions is determined based on one-way delay, selection time, repetition time cost, and delay penalty, including: Based on the preset suggestion hit benefit coefficient and benefit saturation speed parameters, the hit benefit that marginally decreases as the number of candidate suggestions increases is calculated. Specifically, the suggestion hit benefit coefficient and benefit saturation rate parameter are obtained through offline calibration using suggestion hit data from historical customer service conversations. The calibration process is based on at least one thousand interaction records containing different numbers of candidate suggestions, hit results, and corresponding processing times. The suggestion hit benefit coefficient represents the average reduction in processing time after a correct suggestion is hit. It is determined by statistically analyzing the difference between the actual processing time when a suggestion is hit and the average processing time when a suggestion is not hit in historical data. The value ranges from 0.3 to 0.8 seconds, which is determined based on the improvement in decision-making efficiency of customer service with effective suggestion assistance, ensuring accurate quantification of the time benefit brought by a hit. The benefit saturation rate parameter is used to control the saturation rate of hit benefit as the number of candidate suggestions increases. It is determined by analyzing the growth curve of the hit probability under different numbers of candidate suggestions. When the number of candidate suggestions increases to 10 to 15, the marginal increase in hit probability is usually less than 5%. Based on this, the value range of this parameter is set to 0.2 to 0.5. The larger the value, the faster the benefit saturation, which is consistent with the actual law that the effect of new suggestions on the hit probability gradually weakens after the number of suggestions increases to a certain scale. When calculating the hit reward, the number of candidate suggestions is used as input. An exponential saturation model is adopted. First, the product of the reward saturation rate parameter and the number of candidate suggestions is calculated. Then, the negative of the natural exponent of the product is taken and 1 is added. Finally, the result is multiplied by the suggestion hit reward coefficient. The value obtained is the hit reward. This calculation process ensures that the reward decreases marginally with the increase of the number of candidate suggestions, which is consistent with the influence of the number of suggestions on the hit probability in the actual scenario. This provides a quantitative basis for the subsequent calculation of the target value.

[0033] Add the selection time to the repetition time cost, one-way delay, and delay penalty to obtain the total time cost corresponding to the number of candidate suggestions; Subtracting the hit benefit from the total time cost yields the target value corresponding to the number of candidate suggestions; Specifically, for each candidate suggestion within a preset suggestion range, the selection time corresponding to that number of suggestions is extracted. This selection time is calculated using a selection time function, reflecting the time consumed by the agent in selecting from the suggestion set for that number of suggestions. Simultaneously, the re-speech time cost of the current session is obtained, derived from the recognition cost estimation step, representing the time spent repeating statements due to speech recognition errors. The one-way latency of the current session is also obtained, derived from the voice and network observation step, reflecting the one-way transmission time of audio data. Finally, the latency penalty of the current session is obtained, derived from the latency penalty calculation step, reflecting the additional impact of one-way latency exceeding the comfort limit. The four parameters—selection time, re-speech time cost, one-way latency, and latency penalty—are summed to obtain the total time cost corresponding to that number of candidate suggestions. This cost comprehensively quantifies the interaction time loss at that number of suggestions. Based on this, the hit benefit corresponding to that number of candidate suggestions is extracted, exhibiting a diminishing marginal return as the number of candidate suggestions increases, reflecting the processing time saved after a suggestion is hit. Subtracting the hit benefit from the total time cost, the difference is the target value corresponding to the number of candidate suggestions. The smaller the target value, the higher the overall interaction efficiency under that number of suggestions, providing a quantitative basis for selecting the optimal number of suggestions from the candidate range.

[0034] Within the preset number of suggestions, the number of candidate suggestions with the smallest target value is taken as the optimal number of suggestions; Specifically, the preset range for the number of suggestions is a positive integer from 1 to 15. This range is determined based on historical customer service conversation data. Statistical analysis shows that when the number of suggestions exceeds 15, the marginal increase in hit probability is less than 0.5%, while the average selection time increases by more than 30%. Therefore, a range of 1 to 15 can cover effective scenarios while avoiding ineffective calculations. Each candidate suggestion within this range is iterated through, and the target value corresponding to each number is extracted. The target value is the difference between the total time cost and the hit benefit. The total time cost includes selection time, repetition time cost, one-way latency, and latency penalty. The hit benefit decreases marginally with the increase in the number of candidate suggestions. The target values ​​of all candidate suggestion numbers are compared one by one, and the target value with the smallest value is selected as the optimal number of suggestions. If multiple candidate suggestion numbers have the same target value, i.e., a tie occurs, the candidate suggestion number with the smaller value is selected as the optimal number of suggestions. This rule is determined based on the efficiency principle of reducing selection time, ensuring that interaction efficiency is improved with fewer suggestions while maintaining the same target value. Through the above process, the optimal number of suggestions that balances time cost and hit rate is finally obtained, providing a core reference for determining the final number of suggestions.

[0035] The budget contraction module is used to calculate the maximum number of suggestions within the time budget, and determine the final number of suggestions based on the maximum number of suggestions and the optimal number of suggestions. In an embodiment of the present invention, within a time budget, the maximum number of suggestions is calculated, and the final number of suggestions is determined based on the maximum number of suggestions and the optimal number of suggestions, including: Subtract one-way latency, repetition time cost, and fixed overhead from the upper limit of comfortable conversation latency to obtain the time budget; Specifically, the maximum comfortable latency for a conversation is set at 150 milliseconds. This value is determined according to the ITU-TG.114 international standard, which has been verified through extensive voice interaction experience experiments, clearly defining this value as the maximum single-way latency threshold for users to perceive a natural and smooth dialogue. From this maximum comfortable latency, the following subtractions are made: first, the single-way latency of the current conversation, reflecting the one-way transmission time of audio data from the client to the server; then, the cost of repeating speech, quantifying the time lost due to repeated statements caused by speech recognition errors; and finally, a fixed overhead, which is the basic preparation time before the agent begins selecting a suggestion, ranging from 0.1 to 0.3 seconds, determined based on the physiological reaction time characteristics of human visual focus and simple operations. The result of these continuous subtractions is the time budget, which limits the variable portion of the selection time that varies with the number of candidate suggestions, ensuring that the total latency of a suggestion display and selection does not exceed the maximum comfortable latency for the conversation, thus guaranteeing the smoothness of the user's interactive experience.

[0036] The maximum number of suggestions within the time budget is calculated using the inverse transformation of the logarithmic relationship. Specifically, when calculating the maximum number of suggestions, the time budget is first divided by the noise-corrected time slope to obtain... The maximum allowable value is then calculated by performing an exponential operation with base 2 on that allowable value. The maximum theoretical value is calculated, and then 1 is subtracted from this theoretical value to obtain the theoretical upper limit of the number of candidate suggestions. Finally, the theoretical upper limit is rounded down to obtain the maximum upper limit of the number of suggestions in integer form. The rounding operation is determined based on the actual requirement that the number of suggestions must be a positive integer. At the same time, this upper limit is compared with the preset upper limit of the number of suggestions. If it exceeds the limit, the upper limit of the suggestion range is used as the maximum upper limit of the number of suggestions to ensure that the result is within the effective calculation range. The final maximum upper limit of the number of suggestions can ensure that the corresponding selection time does not exceed the time budget, taking into account both interaction efficiency and user experience.

[0037] The maximum number of suggestions is compared with the optimal number of suggestions, and the smaller value is selected as the final number of suggestions. Specifically, the maximum number of suggestions calculated under time budget constraints is obtained. This maximum number of candidate suggestions does not exceed the time budget, ensuring that the variable portion of the selection time does not exceed the remaining time within the comfortable latency range of the session. Simultaneously, the optimal number of suggestions determined by minimizing the target value is obtained. This number represents the optimal solution within the preset suggestion number range, considering both time cost and hit benefit. The maximum suggestion limit and the optimal suggestion number are compared numerically. If the maximum suggestion limit is less than the optimal suggestion number, it indicates that the selection time corresponding to the optimal suggestion number has exceeded the time budget. In this case, the maximum suggestion limit is selected as the final suggestion number to meet the user's comfortable perception of interaction latency. If the maximum suggestion limit is greater than or equal to the optimal suggestion number, it indicates that the selection time corresponding to the optimal suggestion number is within the time budget. In this case, the optimal suggestion number is selected as the final suggestion number to balance interaction efficiency and hit benefit. By selecting the smaller of the two values, the final determined number of suggestions avoids latency exceeding the limit due to too many suggestions while maximizing interaction efficiency within the time constraint, ensuring that the interaction between customer service and users is both smooth and efficient.

[0038] The retrieval and cropping module is used to truncate documents in the knowledge base based on the final number of suggestions, forming a suggestion set; In an embodiment of the present invention, documents in the knowledge base are truncated according to the final number of suggestions to form a suggestion set, including: Retrieve the user's speech-to-text query content and all documents in the knowledge base; A probabilistic relevance scoring method based on term frequency and inverse document frequency combined with document length normalization is used to calculate the relevance score of each document. Specifically, the voice signal input by the user is converted into text form through a voice recognition engine to obtain the query content. This conversion process adopts a method of integrating an acoustic model and a language model based on a deep neural network to ensure that the voice-to-text conversion accuracy is not less than 95%. This accuracy is determined according to the common voice recognition performance standards in the customer service field. At the same time, all documents are read from the knowledge base. The knowledge base documents include structured and unstructured text data such as pre-organized business Q&A, product descriptions, and process specifications, covering all possible query ranges of users in the current business scenario. The query content and each document are pre-processed respectively, including using the bidirectional maximum matching word segmentation method for word segmentation and removing the words in the stop word set. The stop word set is constructed based on the corpus statistics in the customer service field and contains words such as "de", "le", "zai" with a frequency of occurrence exceeding 30% of the total number of words and no actual semantic meaning to reduce the interference of irrelevant information. Calculate the word frequency of the query word in each document. The word frequency is the ratio of the number of times the query word appears in the document to the total number of words in the document, reflecting the occurrence density of the query word in the document. Calculate the inverse document frequency of each query word. The inverse document frequency is the logarithm to the base 10 of the ratio of the total number of documents in the knowledge base to the number of documents containing the query word. The choice of the base 10 is based on the common practice in the information retrieval field, which can balance the weight differences between high-frequency and low-frequency words. When a query word appears in all documents, the inverse document frequency is 0 to avoid the influence of non-discriminatory words on the scoring. Multiply the word frequency of each query word by its inverse document frequency to obtain the weight of the word in the corresponding document, and sum the weights of all query words in the document to obtain the initial relevance score. Perform document length normalization on the initial relevance score. Specifically, divide the initial score by the square root of the total number of words in the document. This normalization method is determined according to the characteristic that long documents are likely to contain more irrelevant words, which can eliminate the interference of document length differences on the scoring. Finally, obtain the relevance scores of each document. The higher the score, the higher the degree of association between the document and the query content.

[0039] Sort all the knowledge base documents in descending order of the relevance score; From the sorted document sequence, intercept the preface documents with the same quantity as the final recommended number to form a recommendation set; Specifically, the relevance scores of all documents in the knowledge base are numerically compared, and the documents are sorted in descending order of score. If documents with the same score exist, a secondary sort is performed based on their update time, with more recent documents appearing first. This rule is determined based on the timeliness priority characteristic of customer service information, ensuring that the latest business content is displayed first. After sorting, the final number of suggestions is determined, which is the optimal value after balancing time cost, hit rate benefits, and time delay constraints. The length of the sorted document sequence is compared with the final number of suggestions. If the document sequence length is greater than or equal to the final number of suggestions, the preceding documents in the sorted document sequence with the same number as the final number of suggestions are extracted. If the document sequence length is less than the final number of suggestions, all documents are extracted. This processing method is determined based on the actual need to ensure that the suggestion set is not empty in customer service interactions, avoiding situations where no suggestions are available due to insufficient document quantity. The document set obtained from the above extraction is the suggestion set. This suggestion set contains documents that are most relevant to the user's query and whose quantity is appropriate for the interaction efficiency. These documents are then displayed to customer service agents to help them respond to user needs quickly.

[0040] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A customer service decision support system based on a knowledge base, characterized in that, include: The observation and measurement module is used to perform speech segmentation on the incoming audio stream, calculate the signal-to-noise ratio based on the speech segmentation results, and determine the one-way delay based on the end-to-end time difference. The selection time module is used to calculate the selection time for the number of candidate suggestions based on a pre-calibrated set of selection time parameters. The re-speech cost module is used to calculate the re-speech time cost based on the signal-to-noise ratio and a pre-calibrated set of identification parameters. The delay penalty module is used to set the out-of-bounds delay based on the one-way delay and determine the delay penalty based on the out-of-bounds delay; The optimal number of suggestions module is used to determine the optimal number of suggestions based on one-way delay, selection time, repetition time cost, and delay penalty. The budget contraction module is used to calculate the maximum number of suggestions within the time budget, and determine the final number of suggestions based on the maximum number of suggestions and the optimal number of suggestions. The retrieval and cropping module is used to truncate documents in the knowledge base based on the final number of suggestions, forming a suggestion set.

2. The customer service auxiliary decision-making system based on a knowledge base according to claim 1, characterized in that, The incoming audio stream is segmented into speech segments. The signal-to-noise ratio (SNR) is calculated based on the speech segmentation results, and the one-way delay is determined according to the end-to-end time difference, including: The incoming audio stream is subjected to voice activity detection by a preset energy determination threshold, and the audio stream is divided into a voice segment sample set and a silence segment sample set. Calculate the mean square energy of the speech segment sample set and the mean square energy of the silence segment sample set respectively; The signal-to-noise ratio of the current session is obtained by taking the logarithm of the ratio of the mean square energy of the speech segment sample set to the mean square energy of the silence segment sample set. Collect the transmission time and corresponding reception time of the incoming call audio; The difference between the sending time and the receiving time is used as the one-way delay.

3. The customer service auxiliary decision-making system based on a knowledge base according to claim 1, characterized in that, The selection time for calculating the number of candidate suggestions based on a pre-defined set of selection time parameters includes: Obtain a pre-calibrated set of selection time parameters, including: fixed overhead, time slope baseline, noise expansion coefficient, and reference signal-to-noise ratio; Based on the difference between the current signal-to-noise ratio and the reference signal-to-noise ratio, the time slope baseline is monotonically adaptively corrected by combining the noise expansion coefficient to obtain the noise-corrected time slope. Construct a selection time function based on fixed overhead and a noise-corrected time slope; Within the preset number of suggestions, for each number of candidate suggestions, the selection time corresponding to each number of candidate suggestions is calculated using a selection time function.

4. The customer service auxiliary decision-making system based on a knowledge base according to claim 1, characterized in that, Based on the signal-to-noise ratio and a pre-calibrated set of identification parameters, the time cost of repeating the audio is calculated, including: Obtain a pre-calibrated set of recognition parameters, including: a recognition quality threshold and a recognition cost coefficient; If the signal-to-noise ratio is lower than the recognition quality threshold, the difference between the recognition quality threshold and the signal-to-noise ratio is used as the gap; otherwise, the gap is set to zero. The time cost of repeating the information is obtained by multiplying the gap amount by the identification cost coefficient.

5. The customer service auxiliary decision-making system based on a knowledge base according to claim 1, characterized in that, An out-of-bounds delay is set based on the one-way delay, and a delay penalty is determined based on the out-of-bounds delay, including: If the one-way latency is greater than the preset session comfort latency limit, the difference between the one-way latency and the session comfort latency limit will be used as the out-of-bounds latency; otherwise, the out-of-bounds latency will be set to zero. The delay penalty is obtained by multiplying the out-of-bounds delay by the preset delay out-of-bounds penalty coefficient.

6. The customer service auxiliary decision-making system based on a knowledge base according to claim 1, characterized in that, Based on one-way delay, selection time, repetition time cost, and delay penalty, the optimal number of suggestions is determined, including: Add the selection time to the repetition time cost, one-way delay, and delay penalty to obtain the total time cost corresponding to the number of candidate suggestions; Based on the preset suggestion hit benefit coefficient and benefit saturation speed parameters, the hit benefit that marginally decreases as the number of candidate suggestions increases is calculated. Subtracting the hit benefit from the total time cost yields the target value corresponding to the number of candidate suggestions; Within the preset number of suggestions, the number of candidate suggestions with the smallest target value is taken as the optimal number of suggestions.

7. The customer service auxiliary decision-making system based on a knowledge base according to claim 1, characterized in that, Within the time budget, calculate the maximum number of suggestions, and determine the final number of suggestions based on the maximum number of suggestions and the optimal number of suggestions, including: Subtract one-way latency, repetition time cost, and fixed overhead from the upper limit of comfortable conversation latency to obtain the time budget; The maximum number of suggestions within the time budget is calculated using the inverse transformation of the logarithmic relationship. The maximum number of suggestions is compared with the optimal number of suggestions, and the smaller value is selected as the final number of suggestions.

8. The customer service auxiliary decision-making system based on a knowledge base according to claim 1, characterized in that, Based on the final number of suggestions, documents in the knowledge base are truncated to form a suggestion set, including: Retrieve the user's speech-to-text query content and all documents in the knowledge base; A probabilistic relevance scoring method based on term frequency and inverse document frequency combined with document length normalization is used to calculate the relevance score of each document. Sort all knowledge base documents from highest to lowest relevance score; From the sorted document sequence, extract the preceding documents with the same number as the final number of suggestions to form a suggestion set.

Images