Data intelligence industry innovation service media
——Focus on digital intelligence and change business
Recently, Hang Seng Electronics and its subsidiary Hang Seng Juyuan officially released new digital financial products based on big language model technology: the financial intelligent assistant Photon and the newly upgraded intelligent investment research platform WarrenQ. In addition, LightGPT, a large model of Hang Seng Electronics' financial industry, also made its debut.
Liu Shufeng, chairman of Hundsun Electronics, said that the large model is the latest breakthrough in information technology, refreshing people's understanding of machine intelligence, and at the same time refreshing the industry's application of traditional AI models.
At the beginning of this year, the generative AI represented by ChatGPT set off a global upsurge, and the large-scale model technology is redefining all walks of life. Among them, the financial industry is the pioneer of digitalization and intelligence, and it is also regarded as the latest implementation of large-scale model technology. good field.
In the era of large-scale models, large-scale models with general-purpose capabilities have become infrastructure, which will have a profound impact on the level of intelligence and digitization of the financial industry.
As a technology company serving the financial industry, Hundsun Electronics is combining its own technical capabilities and in-depth understanding of financial business to continue to create large models of the financial industry and new digital intelligence products based on large models, providing new momentum for the application of large models in the financial industry.
Technological advancement drives industry change
Every technological advancement will promote major changes in society, and the large model is a new breakthrough in the field of information technology.
Judging from the development of information technology for decades, we can clearly see the superimposed form of the "three waves". From the earliest mainframe and PC informatization to the Internet, network, and mobile Internet, the entire earth has become a "village". .
The birth of large models is the third wave of this round of information technology. The impact from 5G, cloud computing, including traditional AI has made the outside world feel the wave of digitalization brought by AI, but large models have pushed this wave to a new level. the height of.
In fact, the development of China's financial technology is accompanied by the superposition of "three waves". In the vertical field of financial supervision, the disruption brought by the wave of technological progress may not be so strong, but every progress will also give birth to nascent business.
Liu Shufeng gave an example. In the information age, China Merchants Bank established itself with technology and grasped the power brought by information technology to grow. The development of the Internet has given birth to financial service products in new scenarios such as Alipay and Yu'E Bao, as well as Internet-native financial service institutions such as Ant Financial, Oriental Fortune, and Zhongan Insurance.
Therefore, large-scale model technology is inevitable for the update of the traditional era, which means higher technical threshold and larger scale. In the large-scale model era, "data + algorithm + computing power" constitutes the basic elements of the new paradigm, and these The basic elements will enter the general and vertical fields and continue to expand.
Based on the basic large model, Hang Seng Electronics has seen two possibilities for the connection with the vertical scene: one is to connect horizontally, and the other is to combine vertically. Lianheng uses industry plug-ins as a plug-in method to add different types of knowledge modules to the big language model, including common sense knowledge, domain knowledge, event knowledge, etc., to adapt and integrate with the big language model, so as to improve its performance in complex tasks. Performance. Another form is the "joint vertical" mode that Hang Seng Electronics is exploring.
Liu Shufeng said that in the financial field, the "Lianheng" model dominated by large-scale model suppliers will encounter many problems, such as the ownership of data property rights. When it is difficult to achieve in-depth application of the "connected horizontal model", it is necessary to establish an "industry large model" to undertake the "vertical" needs of the vertical field.
He also mentioned that large-scale industry models also face inherent challenges, such as computing power collaboration, internal and external data collaboration, scenario collaboration, inter-agency collaboration, and so on. "Especially data coordination, which is a very specific and very difficult problem to solve, especially the financial industry is facing a lot of compliance restrictions."
To a certain extent, the commercial application of the large model focuses on the financial vertical field, and still has limitations in terms of domain knowledge timeliness, data security and privacy protection, and specific application support. Under the joint vertical model, the cooperation between the upstream and downstream of the large-scale industry can be fully utilized to improve the level of AI applications in the financial field while solving data security and privacy protection issues, providing stronger support for financial digital intelligence.
Digital intelligence in the financial industry is changing from quantitative to qualitative
"Building a good financial large model depends on high-quality data, excellent basic large models, professional large-scale model capabilities, and sufficient computing power." Bai Shuo, dean of Hang Seng Research Institute and chief scientist of Hang Seng Electronics, said.
Beginning in 2014, Hundsun Electronics officially launched AI research work, created capabilities such as NLP, OCR, CV, and knowledge graphs, and empowered AI technology capabilities to intelligent customer service, intelligent investment research, intelligent operations, intelligent marketing, and intelligent investment. Gu and data and risk-related business systems.
Up to now, Hundsun Electronics has released 20+ artificial intelligence products. The service organizations include banks, securities, funds, futures and other financial institutions. It has more than 500 customer cases and has achieved the progress of AI products from usable to easy to use.
WarrenQ is a professional integrated investment research tool platform launched by Hang Seng Juyuan for investment research and investment scenarios. At this press conference, WarrenQ launched two AI tool products - WarrenQ-Chat and ChatMiner.
WarrenQ-Chat is a chat product in the financial vertical field. It uses large-scale model superposition search and Juyuan financial database to easily obtain financial market conditions, information and data through dialogue instructions, and each generated dialogue supports original text traceability to ensure that the news The source can be traced, and financial professional reports can also be generated, easily realizing "voice control".
ChatMiner is a financial document miner, built on the basis of large models and vector databases, it can quickly interpret specified documents according to user dialogue instructions, provide accurate retrieval and positioning, extract key information, and effectively integrate and refine information Or expand, intelligently process massive text data.
Bai Xue, deputy general manager and product director of Hang Seng Juyuan, mentioned that there are many scenarios and functions in WarrenQ, including omnipotent readers and writers, citations and calculations, and calculation version models.
For example, ChatMiner can collect a large collection of news events. If you are more interested in one of them, you can trace the news and check the content of the research report in the original text. If you see the middle paragraph and want to keep it, you can click and hold it, and drag it into the note with one click.
From the perspective of a product manager, Bai Xue said that after trying many products this year, she truly feels that the big model is changing software products, changing software interaction, and changing the software industry. What the future of "big model + data + software" will look like is a topic that will be widely discussed in the industry.
In their view, the bottom layer is the data layer, which is different from the generated data layer. The traditional data layer is to help the organization build a digital intelligence platform. Many of the bottom layers are public data plus the organization's own database. Now the organization The database will add financial basic corpus, product corpus, and public large-scale model corpus to form the bottom layer of digital intelligence products in the new era of large-scale models.
At the model layer, the pre-training of financial data can be done by combining the prepared data and corpus, and fine-tuning of supervision can also be done at the same time. After the adjustment, a large model of the financial version can be obtained. In terms of productization in the financial field, it is necessary to continue to train plug-ins.
How to build a more professional financial model?
At the press conference, LightGPT, a large model of the financial industry created by Hundsun Electronics, also made its debut. As a professional large-scale model for the financial industry, LightGPT is more professional, more compliant, and lighter than general-purpose large-scale models.
Bai Shuo, dean of Hang Seng Research Institute and chief scientist of Hang Seng Electronics, said that LightGPT has more professional financial corpus accumulation processing and more efficient and stable large-scale model training methods, using financial field data of more than 400 billion tokens and languages of more than 40 billion tokens Strengthen the data and use it as the secondary pre-training corpus of the large model to support the fine-tuning of more than 80+ financial-specific task instructions, so that LightGPT has the ability to accurately understand the financial field. LightGPT will complete a new round of financial capability upgrades at the end of September and officially open the trial interface.
In Bai Shuo's view, there are three main reasons why the existing large models are difficult to implement in specific financial fields:
First, the quality of the model itself. The financial industry has very high requirements for the content of the answers and the quality of the services. It is difficult to achieve general large-scale model training based on public data, and there is a clear gap between the effect and the needs of the financial industry.
Second, in terms of compliance supervision, the supervision of the financial industry has clear restrictions on data flow and identity (such as whether to operate with a license), which is difficult to meet with the existing general model.
Third, in terms of computing power cost, combined with the requirements of the above two aspects, many financial institutions clearly require private deployment when using large models, at least within the credible range of the industry. Under such a deployment method, if the parameter volume is not large enough , the quality may be relatively poor. If the parameter volume is large enough, the computing power cost of deployment will be relatively high just from the inference point of view.
Therefore, Hundsun Electronics firmly chooses to stand on the shoulders of giants, fully absorb the high-quality results of the existing basic large-scale models, and continue to polish the financial large-scale models needed by the industry.
In this regard, Bai Shuo summarized four key factors for building a good financial model.
First, high-quality data. Finance has high professional requirements, and the data must reflect the specialization of the industry, which is the key to distance from the general model. The data sources of LightGPT include Juyuan’s text data and structured data of more than 400 billion tokens in the past 20 years, financial textbooks, financial encyclopedias, government reports, regulations and other data of more than 40 billion tokens, as well as some fine-tuned data sets.
Second, an excellent basic model. Starting from an excellent basic large model, there is room for improvement when facing the professional field. Hundsun Electronics actively explores cooperation with large financial models at home and abroad, so that LightGPT has a higher starting point.
Third, professional large-scale model capabilities. Train the professional capabilities of large models, including algorithms, talents, and external cooperation. In terms of algorithms, it is necessary to learn and adopt more advanced algorithms, including parameter freezing, fine-tuning of domain-related instructions, and algorithms related to reinforcement learning. In terms of talent reserves, Hundsun Electronics has a national post-doctoral scientific research workstation. In terms of external cooperation, Hundsun has reached cooperation with universities such as Fudan University, Zhejiang University, Chinese Academy of Sciences, and University of Science and Technology of China.
Fourth, sufficient computing power. On the one hand, it comes from the cooperation between Hang Seng and Internet vendors and cloud vendors in terms of computing power. On the other hand, Hang Seng has also invested some of its own computing power to support the research and development of LightGPT.
As a leading financial technology company in China, Hundsun Electronics is gradually showing its ability to build for the financial industry. Facing the further deepening of the market, Liu Shufeng said that it is expected that by 2030, the entire financial industry will complete the digital intelligence upgrade. As a pioneer among them, Hundsun Electronics will enter a new era of digital intelligence together with the financial industry.
Text: Zhou Yao / Data Ape
