//
Editor's note: In the telecommunications industry, telephone services are used in all walks of life. All enterprises that need marketing and customer service have such service needs. However, as the telecom industry continues to develop, various pain points continue to emerge. In response to these problems, Qingdao Dongting Intelligent Technology Co., Ltd. developed the intelligent outbound call robot GO, which provides new solutions in various scenarios for the credit field.
Text/Chen Liang
Organize/LiveVideoStack
Hello everyone, my name is Chen Liang, from Qingdao Dongting Intelligent Technology Co., Ltd. The theme I want to share with you today is intelligent outbound calling: leading the future of credit services.
-01-
New ideas for solving traditional customer service scenarios in the credit field
There are several major pain points in the current industry, and the specific manifestations are as follows:
First, the problems of high labor costs and staff turnover are significant. Newly recruited operators need to be trained, but the daily work content is relatively simple, resulting in a high employee turnover rate.
Secondly, the operator's mood swings are large, and the long hours of working on the phone lead to a decrease in work efficiency.
Third, there are difficulties in statistics and recording, and duplicate records are prone to occur, resulting in waste and loss of customer resources.
The traditional outbound call process usually requires manual dialing operations using software. Faced with heavy call duration tasks every day, sometimes customers will even hang up immediately after being connected, forcing staff to work overtime to complete the task.
In the financial industry, collection is an indispensable link. However, with the development of big data, cloud computing and artificial intelligence technology, the collection field is gradually undergoing an intelligent transformation.
Intelligent conversational robots are not limited to the collection field, they have a wide range of applications in multiple scenarios. For example, in the field of customer service, this kind of intelligent outbound call robot can provide support to customers and answer frequently asked questions. In addition, in daily life, it can also be used for appointments, registrations, and even market research to quickly collect a large amount of user feedback data. Intelligent conversational robots can handle the entire process from listening, responding to recording, while humans only need to be responsible for management. Therefore, robots play an important role in assisting humans in various tasks.
The image above illustrates the value provided by intelligent conversational robot products.
By comparing human customer service and intelligent robots, robots show far superior advantages over human calls in terms of call volume. An intelligent robot can efficiently complete the workload equivalent to 5 to 10 human customer service personnel every day, thus reducing labor costs.
In addition, the intelligent robot can operate continuously 24/7 and 365 days a year. Secondly, intelligent robots can always maintain a stable working state because they are not affected by emotions, unlike human customer service personnel who are easily affected by emotional fluctuations and affect work efficiency. In terms of business training, you only need to import the speaking skills directly into the robot, without the need for excessive business training, thus saving training costs and time. Finally, in terms of operating costs, intelligent robots are less subject to site restrictions and have lower operating costs, which also brings economic advantages to enterprises. It can be seen that intelligent conversation robot products bring practical value to enterprises in many aspects.
-02-
Little Go makes communication more efficient and smarter
The Little Go robot is an intelligent robot that integrates multiple artificial intelligence technologies such as ASR (automatic speech recognition), NLP (natural language processing) and TTS (text to speech synthesis) . It has the ability to accurately determine customer intent and answer questions in a flexible manner.
During use, the little Go robot adopts a human-machine integrated working mode. First, the robot will filter large batches of calls, identify which calls require human intervention, and then flexibly integrate these calls into the manual processing process. The little Go robot has a convenient chat configuration function, allowing administrators to easily modify and edit chats. The assistance of robots can help humans complete boring and repetitive tasks, thereby reducing labor costs and improving work efficiency.
The picture above shows the complete workflow of the outbound call robot. First, the administrator manages the calls based on information such as dialing objects, dialing time periods, follow-up frequencies, strategies, and lines, and can also define multiple speech versions. After the settings of certain call skills versions are completed, the agent will import the calls that need to be made into the system, select the corresponding strategy and call skills, and then start making calls. Afterwards, the little Go robot will answer the call, perform speech recognition, extract dialogue information, and then generate reply text through natural language processing technology. Finally, it will synthesize the voice and broadcast it through TTS technology, and finally complete the call. Call results can also be displayed in the form of an intelligent data dashboard, making it easy for agents to view and export the call results for policy adjustments.
In short, the little Go robot is an intelligent robot that integrates multiple artificial intelligence technologies. Through the combination of man and machine, it can achieve an efficient dialing workflow, improve work efficiency and reduce costs.
The following is a detailed introduction to outbound call robot products:
1. Flexible dialogue management configuration : This outbound call robot allows users to flexibly configure dialogue management, including the definition of dial results and conversation skills. Users can customize the conversation process according to their needs and configure appropriate answering techniques according to different situations.
2. Call center configuration : In the call center, users can select different inbound and outbound lines, and adjust the dialing priority. This helps to better manage call traffic and improve call efficiency.
3. Intelligent policy configuration : The outbound call robot supports flexible policy configuration. Users can configure the call frequency and time period to increase the call completion rate and improve user experience.
4. Call task import : This product supports single call and batch import of call tasks. Users can easily import call tasks into the system and import customer information with one click to realize intelligent outbound call operations.
5. Powerful conversational robot : The conversational robot is the core component of the product, providing the ability to accurately judge customer intentions and flexibly answer questions. By integrating technologies such as ASR, NLP and TTS, an intelligent dialogue process is realized.
6. Data dashboard display : The product presents the dialing results in the form of a data dashboard, making it convenient for users to view the overall dialing process and statistical data. This helps users better understand the performance of their calls and make adjustments accordingly.
In short, this outbound call robot product provides flexible configuration and functions in many aspects, including dialogue management, call center configuration, policy configuration, call task import, intelligent dialogue robot, etc., and at the same time displays the results through the data dashboard, bringing users Efficient, smart and convenient outbound calling experience.
-03-
Technical features
Speech recognition is an interdisciplinary field covering multiple disciplines, including acoustics, linguistics, phonetics, digital signal theory and other disciplines. Dongting Intelligence independently developed a speech recognition engine driver and adopted the internationally leading Transformer algorithm. This algorithm has achieved remarkable results in accuracy, reaching a high accuracy level of more than 93%.
The structure of the model is very clear and concise, and is mainly divided into three parts: one Encoder and two Decoder. The CTC decoding results are scored by the CTC decoder and the Attention decoder. This model is suitable for both streaming and non-streaming speech recognition. During the recognition process, the processing time window chunk size can be adjusted as needed. If the chunk is large, the model will decode after obtaining the complete segment, which is a non-streaming recognition. If the chunk is small, the model will support streaming recognition. The real-time recognition accuracy of the model can reach below 0.1, which means it takes less than 1 second to parse 10 seconds of audio.
In terms of model training, Dongting Intelligence uses LianSignal Group’s customer service voice call data for many years. The data has been annotated by a professional annotation team, with a total of approximately 1,400 hours of voice data annotated. This data provides a solid foundation for model training.
The overall structure of the model framework:
The first layer is based on the PyTorch ecosystem . During the training phase, the model relies entirely on PyTorch for training, without the need for complex installation and configuration tools.
The second layer covers both R&D and product aspects . Convert the model trained by PyTorch through TorchScript, so that the model can run efficiently in the C++ environment and obtain superior performance. LibTorch is a C++ interface provided by PyTorch, which is very important to meet the needs of industrial scenarios because LibTorch is mainly composed of C++ code.
We provide two deployment options, public cloud and private deployment. This means that users can choose to deploy their models in a public cloud environment, or choose to deploy them in a private environment, depending on their needs. This flexibility helps meet the different deployment needs of different users.
The main purpose of male and female timbre recognition is to identify the speaker's identity. The main reason why voiceprint recognition is not used is that facing the massive amount of audio data every day will lead to high storage costs. Therefore, male and female voice recognition was chosen as a solution to meet the needs of debtor information identification. The model is based on the Librispeech data set, uses Facebook's Wav2Vec2 model, and converts the model into a C++ framework through TorchScript and embeds it into speech recognition. This solution helps improve the accuracy of call recognition and responses, thereby enhancing user experience.
After converting audio data into text, a large amount of unstructured text data needs to be processed. In order to extract useful information from it, we chose the ERNIE 3.0 model to help us extract relevant information from speaker text effectively.
Addresses, personal information, repayment amounts, phone numbers, etc. are extracted from the text, and then this information will be automatically integrated into a reminder for managers to view at any time. In addition, whenever the robot obtains new information, the system will also generate a reminder of new clues to ensure timely follow-up and processing.
Intention understanding is an important research direction in the field of natural language processing, whose goal is to identify the speaker's intention. By deeply understanding the speaker's intent, we can respond and interact more effectively. In this field, a Wenxin-based NLP model is adopted, which covers about more than 100 different intentions. It is worth noting that in terms of data identification in the credit field, the accuracy rate has exceeded 85%.
The recognition of multiple intentions can be completed at the same time.
Intent recognition uses the ernie3.0 model. At the same time, it also supports highly customized intent recognition for other scenarios.
Dialog management is a mathematical model using finite state machines. By transitioning between different states, the flow of the conversation can be effectively managed.
The above examples can help us better understand:
The initial state of the light is off. When the state transition of turning on the light is triggered, the state of the light will change from off to on. In addition, this light has some attributes, such as medium brightness, high brightness, and low brightness. When entering the light-on state, it may also trigger other conditions, such as a malfunction of the light.
Similarly, we can apply this state transition to actual scenarios, such as when little Go is talking to a customer. Initially, he needs to confirm the identity of the client, which corresponds to an unidentified state. When the customer starts speaking, the system will perform intent recognition. If the customer's intention is confirmed to be positive, the system will convert the status from unconfirmed identity to confirmed identity. In this way, the robot can answer questions in a targeted manner and become more intelligent.
This example clearly shows the role of state transition and intent recognition in practical applications, and how to generate corresponding replies based on different states and attributes.
The TTS (Text to Speech) speech synthesis module uses the FastSpeech2 model. The FastSpeech2 model gives speech more diverse tones by predicting and controlling accent pauses in timbre, thereby making the robot's replies more vivid. In addition, customized voice services are provided, allowing users to customize the pronunciation of specific characters.
TTS speech synthesis technology not only makes the robot's response more natural and vivid, but also provides flexible pronunciation customization options to meet various speech synthesis needs. This technology can improve the conversational experience, making the bot’s responses more humane and natural.
The system architecture adopts the design of non-regressive encoder and decoder. Both the encoder and decoder are composed of multiple Transformer blocks, and a transformation adaptation layer is also included between the Encoder and Decoder. The main function of this layer is to associate the pauses between audio features with factors such as pitch, volume, etc., so that the model can better capture the characteristics of the audio.
Keyword recognition is a key task in the field of speech recognition and natural language processing. It is widely used in voice assistants, smart homes, and automotive intelligent systems. For example, the intelligent assistant Siri and Xiaoai classmates are among them. Dongting uses a hybrid model based on temporal convolution and ResNet, which combines the advantages of residual network and temporal convolution to improve efficiency and speed up processing.
Keyword recognition plays an important role in the field of telephone communications. Its main task is to determine the status of the phone, including various situations such as shutdown, shutdown, inability to dial, or ongoing calls. Traditional customer service centers usually use the SIP response code method, but its accuracy is low and the number of categories is limited. In contrast, Dongting uses self-developed keyword recognition technology, which can accurately identify multiple statuses, including 19 common categories, thus providing more accurate and detailed dialing results.
Keyword recognition inputs the MFCC features of each frame into the network as a time series. This method has obvious advantages. The original 3×3 convolution kernel is transformed into a 3×1 convolution kernel. This change increases the calculation amount by 40 times.
The used convolution and residual model network structure has significantly improved performance compared with traditional two-dimensional convolution, and this improvement is also reflected in the number of parameters. The first layer uses a 3×1 convolution kernel, while the subsequent modules use a 9×1 convolution kernel.
Intelligent early warning interception includes two key aspects:
The first part is about risk prevention and how to avoid risks during the process. Before making a call to the call center, we check to see if there are any previous complaints. If there is such a record, we will not use a robot to automatically make the call, but will be handled by a human agent to avoid further complaints.
During the robot's call, the debtor may display agitated emotions, which can be classified into different levels. Through intent identification, we can determine the level of risk. It is worth noting that the higher the risk level, when the risk level reaches 4, we will immediately connect a human agent to calm the debtor's emotions to prevent complaints from occurring. So far, none of the cases have been related to Go robots.
-04-
Customer case
Little Go robot is an intelligent collection robot developed for Lianxin Group. It has the following characteristics:
1. Automation : The little Go robot can automatically perform collection tasks without manual intervention, thereby significantly reducing labor costs.
2. Intelligent : The robot uses intelligent algorithms and intent recognition to identify the debtor's emotions and risk levels and respond to different situations in a targeted manner.
3. Highly customized : The little Go robot can be highly customized according to different needs and adapt to the requirements of different industries and collection strategies.
4. 24-hour operation: The robot can perform tasks all day and without interruption, without being restricted by time and location, improving the efficiency and coverage of collection.
5. Privacy protection: The little Go robot can strictly abide by privacy regulations when performing tasks and protect the security of customers' private information.
In short, the Little Go robot provides Lianxin Group with an efficient, flexible and safe collection solution, which helps improve collection efficiency and reduce costs.
Some business value points of the little Go robot:
Human-machine integrated working mode: The little Go robot adopts a human-machine integrated working mode, combining the automated execution of the machine with artificial intelligent intervention, giving full play to the advantages of both parties and improving work efficiency.
Intelligent risk interception: Through intelligent algorithms and data analysis, the robot can identify potential risks, thereby avoiding interactions with high-risk debtors and reducing unnecessary risks.
Automatically generate dialing strategies: The little Go robot can automatically generate dialing strategies based on data and situations, and flexibly adjust the strategies according to different collection needs, improving the pertinence and efficiency of collections.
Operation interface display: The graphical operation interface enables users to intuitively monitor and manage the work of the robot, and provides convenient management tools to help users better control the collection process.
In general, the Little Go robot provides businesses with efficient, automated, and intelligent solutions through its intelligent collection function, which helps improve collection efficiency, reduce risks, and provides a user-friendly operation interface.
Taking Lianxin Group's full-hand case as an example, the experiment was carried out. In the experiment, the Little Go group used the cooperation between the Little Go robot and humans to collect collections, while the manual group relied only on manual collections. Experimental results show that compared with the artificial group, the small Go group has achieved significant improvements in many aspects:
The per capita collection rate increased by 15.3%, which shows that the Little Go robot and human workers can work together to collect debts more effectively. The proportion of daily dialed cases that can be contacted increased by 18.2%, which means that the robot's automatic dialing and strategies are more targeted and can increase the chance of successful contact.
The number of calls answered per person per hour increased by 36.9%, which shows that the automated dialing and strategy generation of robots can improve collection efficiency in a short period of time.
Taken together, the experimental results show that the small Go robot and manual collection method have obvious advantages in improving the recovery rate, contact rate and efficiency. This proves the potential value of robots in the collection field and can provide more effective collection solutions for companies such as Lianxin Group.
Dongting developed a return visit robot for a public security bureau. It was initially used for epidemic return visits and epidemiological investigations, and was later expanded to community return visits. Its main functions include:
Epidemic investigation: During the epidemic, robots can conduct epidemiological investigations of the epidemic, collect and analyze relevant data, so that the Public Security Bureau can better understand the epidemic situation.
Information notification: The robot can automatically send epidemic-related information notifications to residents, including prevention and control measures, vaccination information, etc., to increase public awareness of the epidemic.
Handling complaint and reporting hotlines: In addition to epidemic-related tasks, the robot can also handle daily complaints and reporting hotlines, assisting the Public Security Bureau in handling various problems and needs of community residents.
The versatility of this robot makes it a powerful tool for emergency situations and daily management, which can improve work efficiency, reduce the workload of public security personnel, and better serve community residents.
The intelligent community return visit robot has significantly improved work efficiency, which is specifically reflected in the following aspects:
Information collection: Robots can quickly and accurately collect information about community residents, including health status, epidemic exposure risks, etc. This will help the Public Security Bureau better understand the epidemic situation in the community and take corresponding prevention and control measures.
Knowledge popularization: Robots can provide residents with epidemic-related knowledge, such as protective measures, virus transmission routes, etc., to help improve residents’ health awareness and self-protection capabilities.
Soothing: Anxiety is often high during the pandemic. The presence of robots can provide psychological support and help calm people's emotions and reduce anxiety through communication and information transmission.
The intelligent community return visit robot has achieved a 50% increase in call completion rate, which not only reduces labor costs, but also significantly improves efficiency.
The above is my sharing, thank you all.
▲Scan the QR code in the picture or click " Read the original text " ▲
Direct access to LiveVideoStackCon 2023 Shenzhen Station 10% off ticket purchase channel