Natural Language Processing Applications in Banking: Issues and Trends

Guo Linhai, Head of AI & Blockchain Direction, Innovation Lab, Shanghai Pudong Development Bank Information Technology Department

1. What is Language?Human language, like the sounds of cats and dogs, is a signal that conveys the brain’s cognitive activities to the outside world. Animals perceive the world through their eyes, ears, skin, and other organs, receiving external information. This information is processed by the brain and conveyed through language for communication. Without language, it is hard to imagine how humans could achieve the current level of advancement.

2. What is Natural Language?When first encountering the concept of “natural language processing,” people often wonder why it is called natural language processing instead of just language processing. Natural language typically refers to a language that evolves naturally with culture, such as Chinese, English, French, etc. Since there are naturally evolved languages, there must also be non-evolved, constructed languages (Constructed Language), such as Python and Java. There are also “constructed languages” in human languages; for example, Esperanto, created by Polish-Jewish ophthalmologist Ludwig Lazarus Zamenhof in 1887, has only about 2,000 native speakers.

3. What is Natural Language Processing?With the advent of computers, floppy disks, USB drives, and hard disks became mediums for storing text. People hoped to use computers to efficiently handle a large volume of repetitive tasks, leading to the emergence of natural language processing. The U.S. government has promoted the development of many new technologies; for instance, the military customized the world’s first electronic computer, the Electronic Numerical Integrator and Computer (ENIAC), to calculate ballistic information. Machine translation is no exception; during the Cold War, the military received many Russian intelligence reports, but very few personnel understood Russian, so they hoped to use computers to automatically translate Russian into English, which is the task of machine translation. In summary, natural language processing is the method and process of using computing devices to handle text information, encompassing all aspects of text information processing.

The history of natural language processing development is tortuous, with continuous struggles between linguistic and statistical schools, which have spiraled research upward, driving robust technological advancements. The entire development process has gone through four important stages: the budding phase, rapid development phase, slow development phase, and revival and integration phase.

Natural language processing involves using computers to analyze, understand, and generate natural language, aiming to facilitate human-computer interaction. Looking back in history, NLP technology has gradually spiraled upward amid the “competition” between the statistical and linguistic schools. Of course, due to the inherent complexity, diversity, ambiguity, vagueness, and discreteness of language, understanding language remains a considerable challenge for computers. Currently, as AI-related technologies continue to improve, many scenarios still require a combination of “human” + “intelligent” models for implementation. The phrase “the jewel in the crown” also implies that it is dazzling yet difficult to reach; researchers in the NLP field are still exploring the path toward “strong artificial intelligence.”

Natural language processing technology is a discipline that integrates linguistics, mathematics, and computer science. The NLP technologies used in vertical domain scenarios are diverse and complex, characterized as follows: NLP technology has certain general applicability in some areas, such as general content review, machine translation, and speech recognition; in many application scenarios, the transferability of NLP technology is poor and requires high customization, such as non-standard contract text extraction, conversational robots, document review, and specific domain text error correction; NLP technology still struggles to integrate domain knowledge well, necessitating significant human effort, requiring deep collaboration between business and technical personnel, often focusing heavily on data cleaning and annotation.

The banking scenario emphasizes legality and compliance, as well as data security, which leads to a relatively cautious approach to data analysis and utilization. To some extent, this sacrifices some efficiency to safeguard user privacy. In recent years, there have been many successful cases of NLP technology applications in banks, primarily focusing on customer service, internal operations, and risk exploration: First, in customer service, intelligent customer service and intelligent outbound calls are among the most successful application cases in the internet and financial industries, utilizing AI technologies such as speech recognition, semantic understanding, voice wake-up, and voiceprint recognition, allowing customers to interact using natural language for transactions, inquiries, payments, and investment purchases, significantly reducing the number of customer service positions, achieving substantial cost reduction and efficiency improvement. Second, in internal operations, due to the nature of banking business, a large amount of document processing work is involved in corporate and personal loan business processes, requiring manual information comparison and review. However, manual reviews are affected by high employee turnover and inconsistent review quality, necessitating the combination of OCR, NLP, and other relevant technologies to achieve intelligent recognition, extraction, comparison, and review of various format documents. The application of related capabilities can significantly reduce review risks, improve review efficiency and quality, and reduce customer waiting times. Third, in risk exploration, NLP plays an important role in risk sentiment analysis, address authenticity verification, and post-loan rating report review, capable of avoiding the shortcomings of previous sampling methods and possessing the ability to intelligently process all business text information. Additionally, by integrating capabilities like knowledge graphs, it can structure text, images, and voice information to build multimodal risk graphs, thereby identifying explicit and implicit risk factors within the graphs and enhancing risk monitoring quality.

1. Lack of Interpretability in Deep Learning-Based NLP Technologies.In recent years, with continuous improvements in computing power and model capabilities, deep learning-based natural language processing technologies have been widely applied in banking business processes. In some key businesses, banks are very cautious about using deep learning models, such as in risk control and recommendation processes, where deep learning NLP technologies cannot explain their results, leading business personnel to be reluctant to use them widely.

2. Limited Generalization Ability of NLP Technologies in Many Scenarios, Making Scaled Effects Difficult to Achieve.In many banking scenarios, the application of NLP technologies faces challenges in scaling; solutions from scenario A cannot be quickly transferred to scenario B. While the underlying NLP technology can be reused, the technical development closer to the scenario still requires significant human involvement for implementation. By employing personnel division of labor and toolization of underlying capabilities, team development efficiency and scaling capabilities can be enhanced. In recent years, the emergence of large models integrating knowledge graphs, knowledge reasoning, and other technologies has alleviated the cold start issue in scaling processes; however, currently, no method effectively resolves the generalization issue.

3. Insufficient In-Depth Understanding of NLP Technologies by All Parties.Internet companies and fintech firms are also vigorously promoting AI applications. During exchanges with financial enterprises, tech companies often exaggerate the boundaries and application scopes of AI capabilities, leading banking personnel to have unrealistic expectations of AI technologies. These unrealistic expectations result in numerous issues during AI application implementation that require deep involvement from business personnel to resolve, causing a significant shift in their understanding of AI technologies, leading them to believe that AI cannot do anything. This creates two typical perceptions among business personnel: “AI can do anything” and “AI can do nothing.” The belief that “AI can do anything” leads to resource waste, potentially investing significant human and financial resources in the wrong direction, resulting in unsatisfactory outcomes. Conversely, the belief that “AI can do nothing” leads to a disconnect between enterprise development and AI capabilities. Over time, this may result in technological obsolescence and declining service experiences; when attempting to catch up, the gap may be too wide to bridge.

In light of these issues, improvements need to be made from the perspective of NLP technology applications. First, management, business personnel, and technical personnel need to proactively understand the boundaries of NLP technologies, conduct industry research on NLP applications, and provide training courses on NLP to understand the collaborative model between business and NLP technologies. Second, in the formation of NLP technical teams, the positioning of teams must be realistic, avoiding excessive pursuit of cutting-edge developments and not being led by vendors and media.

4. The Gap Between Business and NLP Technologies.The business side mainly falls into four categories: First, the business side has no system construction at all and lacks precedents for collaboration with NLP technologies; second, the business side has some system construction experience but no precedents for NLP technology collaboration; third, the business side has some system construction experience and limited NLP technology collaboration experience; fourth, the business side has some system construction experience and rich NLP technology collaboration experience. In a large enterprise, different departments have varying degrees of understanding of AI and NLP technologies, with most concentrated in the first two levels, leading to typical issues in the integration of business and NLP technologies: First, unclear requirements. Due to the traditional business processes not necessitating informatization, digitalization, or intelligence, some departments lack precedents and experience in collaborating with technology, resulting in problems of “intelligence for the sake of intelligence,” and when forming requirement specifications, they cannot clearly articulate their key needs, potentially leading to resource waste with low investment-output ratios. Second, the collaboration model has not kept pace with the times. Business personnel continue to use traditional development requirement proposal models, adopting a “client-vendor” approach to collaboration, proposing requirements and waiting for implementation. In the AI era, the collaboration model between business personnel and NLP technical personnel has fundamentally changed, especially in the intelligent application of NLP, which requires high customization, necessitating deep collaboration between both parties. Business personnel need to discuss the logic of business organization with technical personnel, and technical personnel must “code” the corresponding logic to achieve intelligence, meaning business personnel must teach the “machine” to achieve intelligence. Third, there is excessive focus on algorithm models, detached from business logic. Some NLP technical personnel overly focus on the technology itself, lacking the willingness to understand the entire business process and business logic, resulting in “slow and difficult implementation” issues in NLP technology applications, and may also lead to “incompatibility” phenomena where service outcomes fall short of expectations. Additionally, less experienced NLP developers often overly emphasize cutting-edge algorithm capabilities, lacking objective judgment, leading to significant efforts being placed on technological innovation rather than application innovation targeting specific scenarios. The typical issues mentioned above require continuous promotion of the value orientation of combining “NLP technologies” with “banking business” and engaging in exchange activities such as “NLP technology application case analyses” to enhance mutual understanding between business and technology, thus bridging the gap between business and NLP technologies.

5. Talent Issues.Currently, the banking industry struggles to recruit top technical talent. First, regarding fresh graduate recruitment: In the competition with internet companies, most top talent has flowed toward the internet sector. The banking industry is at a disadvantage in the recruitment of AI talent and needs to adjust its institutional mechanisms and supporting policies to significantly improve talent recruitment issues. Second, in terms of social recruitment: Some candidates with excellent educational backgrounds and cross-disciplinary, cross-industry experience in NLP technology can be hired, but due to differing backgrounds, there are significant differences in understanding NLP technologies, requiring substantial time and effort to unify the overall team development direction.

1. Large Model Compression and Distillation Technologies Will Become Important Capabilities for NLP Technologies.In natural language processing technology, as the number of model parameters continues to increase, it consumes a significant amount of computational resources, leading to performance issues and high resource consumption during application. Additionally, many financial scenarios require near-real-time transactions and edge computing, posing challenges for the implementation of real-time and near-real-time banking services. The emergence of large model compression and distillation technologies can make models more lightweight, allowing large models to be applied in more financial scenarios, improving online operational performance and facilitating deployment.

2. Further Improvement in NLP Model Interpretability.Currently, the effectiveness of deep learning-based NLP models often comes at the expense of model interpretability. Deep learning models are becoming increasingly complex, making their training and inference processes difficult to explain. Providing predictions and explanations to business personnel and efficiently responding to business model fine-tuning requests will become increasingly challenging. Currently, research on interpretability is a hot topic in academia, and the banking industry needs to continuously follow up on relevant research outcomes.

3. Strengthening Independent Research and Development.To enhance independent R&D capabilities and gradually reduce reliance on vendors, the focus should initially be on independent development of localized functions, primarily through self-research and joint development with vendors, thus enhancing self-development capabilities. Meanwhile, in smaller scenarios, exploring full-chain independent development models is necessary. As team size continues to grow, the formation of full-chain, core scenario, and core capability independent development will be achieved.

4. Promoting Scalable Applications of NLP Technologies.The high degree of customization in NLP scenarios makes scaling applications challenging. Efforts should be made to promote the scalable application process of NLP technologies from three aspects: “platform construction exploration,” “scalable R&D teams,” and “standardized R&D processes.” NLP-related technologies are gradually evolving from auxiliary tool types to platform-oriented developments, forming a knowledge computation service system that further deeply embeds into business processes, creating innovative business models. The banking industry needs to accumulate capabilities, forming a solid foundation for NLP technology capabilities while enhancing team building and multi-party collaboration to create complementary advantages and explore new models for scalable NLP applications.

5. Deep Integration of Business and Technology.Grasping the essence of “technology serving business” and aiming to “empower business through technology,” strengthening mutual understanding, promoting the formation of joint teams between business and technology, is essential for future deep integration of business and technology teams. Meanwhile, both sides should be results-oriented, jointly promoting the rapid implementation of NLP technologies.

(Column Editor: Han Weimi)