Application of Large Model in Banking Customer Service Center

With the continuous development of financial technology, financial institutions have changed the traditional way and transferred more services from offline to online. In order to improve customer experience and efficiency, financial institutions continue to explore natural language processing, machine learning and other technical means to optimize Customer service, in which large-scale model technology is widely used, has become one of the most popular research directions and is currently the best way to achieve AGI. The earliest attention to large models originated in the NLP field. With the evolution of multi-modal capabilities, the CV field and multi-modal general-purpose large models have gradually become the mainstream of market development.

This report focuses on the research and analysis of the most popular large models at present, focusing on their outstanding performance in the field of natural language processing and the future application prospects of customer service centers. At the same time, this report also discusses the challenges and limitations faced by the current large language model, and puts forward corresponding optimization suggestions, aiming to provide a reference for the large model research of the banking customer service center.

1. Large-scale model development background

(1) The birth stage of the large language model

• In 2017, Google launched the Transformer neural network architecture for processing natural language tasks.

• OpenAI released GPT-1 in 2018.

(2) Large language model exploration stage

• In 2019, OpenAI released the GPT-2 part of the open source model.

• In 2019, Google launched the BERT model.

• In 2020, Baidu launched ERNINE2.0, which can understand semantics.

• In 2021, OpenAI will launch the DALL-E model that can generate images from text.

• FaceBooK launched the CLIP model.

• Huawei released a large model of Pangu.

• OpenAI launches Codex.

(3) Explosion stage of large language model

• In 2022, OpenAI will launch ChatGPT-3.5.

• In 2023, OpenAI will officially release the GPT-4 pre-trained large model to realize image and video. Compared with GPT-3.5, the performance has been significantly improved, and it has reached human level in some professional and academic fields. GPT-4 has a certain multi-modal ability, can receive the input of combination of graphics and text, and output text reply, and the application scope has been further expanded.

• Microsoft released New Bing based on ChatGPT. Microsoft announced that GPT-4 will be connected to the Office family bucket.

• FaceBook publishes LLaMA - 13B.

• Google updates Bard and introduces the Palm 2 model.

• Fudan team released MOSS.

• Ali released the large language model "Tongyi Qianwen", which has functions such as multi-round dialogue, copywriting, logical reasoning, multi-modal understanding, multi-language support, and external enhancement API. At present, products such as DingTalk and Tmall Genie have taken the lead in accessing the Tongyi Thousand Questions test, and products such as AutoNavi Maps, Ele.me, Hema, Youku, and Tao Piao Piao will also be connected to the Tongyi Thousand Questions model in an orderly manner. All Alibaba products will be connected to large models in the future, and will cooperate with OPPO, Geely, Zhiji and other enterprises.

• SenseTime released the "SenseNova" large-scale model system, launching large-scale models and capabilities such as natural language processing, content generation, automatic data annotation, and custom model training, including the large-scale language model "SenseChat" and a series of generative AI application.

• The Pangu model launched by Huawei focuses on AI for Industry, empowering the application of thousands of industries, and is expected to promote the upgrading of artificial intelligence development from "workshop" to "industrialization". Self-developed ModelArts 2.0 AI development platform, computing power chips such as Ascend 910, and AI training servers such as Zhaohan A5900-A series. The Ascend AI industry ecosystem has developed 20+ hardware partners and 1000+ software partners.

• Tencent Hunyuan AI large model. HunYuan cooperates with Tencent's pre-training research and development force to create industry-leading AI pre-training large models and solutions. Tencent's large model can be connected to WeChat, games, short videos, advertisements, TO B-end and other advantageous businesses. Tencent has a large number of partners in SaaS accelerators, WeChat and other businesses.

• Baidu's "Wen Xin Yi Yan" has strong Chinese comprehension ability, and supports the generation of images, audio and video from text, and has multi-modal capabilities. The Baidu AI base can increase the kilocalorie parallel acceleration ratio by more than 90%. At present, 36 large-scale models and 11 large-scale industry models have been released, and the ecology has gathered 5 million developers.

2. The development process and structure of the large model

(1) The development history of the large model

From rule-based to human-based consciousness, large language models are an inevitable outcome of technological progress. The development of natural language processing to large-scale language models can be divided into five stages: rules, statistical machine learning, deep learning, pre-training, and large-scale language models. From 1956 to 1992, the rule-based machine translation system stringed together various functional modules internally. People first obtained knowledge from the data, summarized the rules and taught them to the machine, and then the machine executed the set of rules. This stage is the rule stage; from 1993 to 2012 is the statistical machine learning stage, the machine translation system can be split into a language model and a translation model, this stage has a higher mutation rate than the previous stage, and the transfer of knowledge from humans to machine automatic Learning knowledge from data. At that time, the amount of manually labeled data was about one million; from 2013 to 2018, it entered the deep learning stage, which was less abrupt than the previous stage. From discrete matching to continuous matching, the model changed. The amount of labeled data has increased to tens of millions; the pre-training stage exists from 2018 to 2022. Compared with before, the biggest change is the addition of NLP self-supervised learning, which expands the available data from labeled data to non-labeled data. .

Figure - Large Model Development Stages

In the large model stage, the performance of data labeling, algorithm, and human-machine relationship has been improved by leaps and bounds. From 2023 onwards, it will enter the large-scale model stage, which has a high degree of mutation, and has shifted from special-purpose tasks to general-purpose tasks or presented in the form of a natural language human-machine interface, aiming to make the machine follow the subjective will of humans. In terms of data labeling, large models have evolved from requiring a large amount of labeled data to using massive non-labeled data. More and more data is being used, and human intervention is getting less and less. In the future, there will be more text data and more other data. The shape of the data is used by the model. In terms of algorithms, the expression ability of large models is getting stronger, the scale is getting bigger and bigger, and the ability of self-learning is getting stronger and stronger, and the trend from special to general is obvious.

(2) Large-scale model technology route

Large-scale model technology routes have their own emphases, and MaaS has become an industry trend. There are three technical routes for the development of large-scale language model research: Bert mode, GPT mode, and hybrid mode. Most of them in China adopt a mixed model, and most mainstream large-scale language models follow the GPT technical route. Until the end of 2022, ChatGPT will be produced on the basis of GPT-3.5. After 2019, the Bert route basically has no iconic new model updates, while the GPT technical route tends to prosper. In the process of evolving from Bert to GPT, the model is getting bigger and bigger, and the performance achieved is more and more general.

The future development of large models will tend to be generalized and specialized in parallel, and platformized and simplified. At the same time, the MaaS model will become a new form of AI application and develop rapidly, reconstructing the commercial structure ecology of the AI ​​industry, and stimulating new industrial chain divisions and business models. In the future, the large model will be deeply applied to the life of users and the production mode of the enterprise, unleashing creativity and productivity, activating creative thinking, reshaping the working mode, helping the organizational change and operating efficiency of the enterprise, and empowering the industrial transformation.

(3) Large-scale model technical architecture

The artificial intelligence large language model represented by the technology behind ChatGPT is giving birth to a new wave of artificial intelligence, setting off a global competition for artificial intelligence large language model technology, technology giants are accelerating their deployment, and the field of generative AI is surging. The technical architecture of the AI ​​large model is divided into five major sections: the basic layer, the technical layer, the capability layer, the application layer, and the user layer. The basic layer involves hardware infrastructure and three core elements of data, computing power, and algorithm model. The technical layer mainly involves model building. At present, the Transformer architecture occupies a dominant position in the field of AI large models, such as BERT and GPT series. AI large models include NLP large models, CV large models, multimodal large models, etc. The capability layer has the capabilities of text, audio, image, video, code, strategy, and multi-modal generation, which are applied in multiple fields to provide customers with products and services. The architecture diagram is shown below.

Figure-AI Large Model Technical Architecture

3. Application of the large model in the customer service center

The development and application of large models in the customer service center requires three elements: data resources, algorithms and models, funds and resources. At present, the application of large models in customer service centers faces challenges such as large computing power requirements, high training and inference costs, poor data quality, weak cross-scenario adaptation, high knowledge base construction costs, and privacy and security issues. There are mainly the following problems in the application of the customer service center:

1. data problem

The online service data of the customer service industry is insufficient, and the diversity of data cannot be guaranteed; the difficulty of labeling industry data limits the quantity and quality of data accumulation. The knowledge base is generalized, and the number and quality of the entries are not high.

1. algorithm problem

For the new scene of customer service online service, there are few corpus and lack of knowledge base in the early stage. Model capabilities require accumulated experience in projects, and industry models need to be systematically improved. The large model tests the full-stack large model training and R&D capabilities, such as data management experience, computing power infrastructure privatization construction capabilities and engineering operation capabilities, underlying system optimization and algorithm design capabilities, etc.

1. logical reasoning questions

Complex, rigorous, flexible logical reasoning and self-learning ability are still the core challenges faced by most large language models. The emerging ability of the large language models known so far determines the basic performance of large language models in terms of logical reasoning. At present, most large language models can make simple judgments on human emotions. Creating content based on understanding and emotional needs is To meet the needs of the customer service industry, understanding human emotions based on logical reasoning is a higher way of thinking for intelligent customer service. At present, most financial consultants can only provide some basic product introductions and recommendations, lack of comprehensive, in-depth, flexible, and effective analysis of large-scale, diverse, and ever-changing financial market data, and the efficiency of investment research is not high.

1. timeliness issue

The ChatGPT-based model is usually trained based on historical data. It does not have the ability to acquire and process new data in real time, and it is difficult to update the knowledge reserve in the model in real time. For the latest information or questions of customers with strong real-time characteristics, the model may output inaccurate or wrong information, but to make the training data include the latest customer service information will consume a lot of time and cost for training, and the update speed will be much slower on search engines.

4. Prospects for the future development of large models

As the number of customers continues to increase, customers' expectations for customer service centers will continue to increase. How to deliver powerful customer service support in bank customer service centers has become particularly important. Large-scale deep learning and transfer learning are required in certain scenarios. It is used to improve the level of AI assistants, and integrates the existing natural language processing, computer vision, intelligent voice, knowledge graph and other AI core technical capabilities of the customer service center to create a large AI language model capability system for banking customer service centers. And improve the system related to the safe application of generative AI. In the face of technological ethics risks, effective content review and supervision mechanisms should be established to prevent the generation and dissemination of bad and illegal content. Strengthen technical supervision and review of large language model applications. Explore specific risk prevention measures and means for the practical application of large language models.

Improve the ability of customer service robots to understand intent. The combination of the large model and special data in the customer service center can improve the intention understanding ability of the customer service robot, and reduce the initial access cost based on the intention analysis of the customer service industry model. Using the knowledge graph of large models, natural language processing technology and algorithmic models, complex questions can be transformed into simple and easy-to-understand instructions to provide more accurate answers.

Improve video/virtual human interaction capabilities. With the integration of generative AI and large language models, the video/virtual human production cycle will be greatly shortened and the creation process will be simplified. At the same time, in view of the deepening of the large model in the logical understanding of the user's language, the recognition, perception and analysis and decision-making capabilities of the virtual human in the application of customer service scenarios will be significantly improved, the interaction ability during communication will be improved, and the individual needs of users will be more accurately met.