DEVELOPMENT OF AN AI-POWERED REMITTANCE CHATBOT

TEAM BERT

HDSC Spring ’24 Cohort

Fatimah Adwan, Oyewale Daniel Oyewola, Olanrewaju Stephen Amudipe, Rafiu Adekoya Badekale, Babatunde Akanfe, Prosperity Oguama, Praise Adetayo, Jeremiah Adetunji, Abimbola Sekinat Abimbola, Chukwuebuka Obi, Uwemishie Emmanuel, Eke Mong Eke

Table of Contents

INTRODUCTION

LITERATURE REVIEW

METHODOLOGY

RESULT

CONCLUSION

REFERENCE

INTRODUCTION

In an increasingly interconnected global economy, remittances play a crucial role in supporting millions of families and driving economic development in numerous countries. These cross-border financial transfers, typically sent by migrant workers to their home countries, have become a lifeline for many communities, particularly in developing nations. However, the complexity of remittance processes, varying costs, and ever-changing regulations can pose significant challenges for both senders and recipients.

As digital technologies continue to reshape the financial landscape, there is a growing need for innovative solutions that can simplify and streamline remittance-related information and services.

This research paper presents the development and implementation of a Remittance Chatbot, an artificial intelligence-driven virtual assistant designed to address these challenges by providing accessible, accurate, and timely information on remittance patterns, costs, and processes.

The Remittance Chatbot leverages on the understanding of human natural language and machine learning techniques to understand and respond to user queries in a conversational manner. By integrating a comprehensive knowledge base of remittance data, regulatory information, and domain-specific expertise, the chatbot aims to empower users with the information they need to make informed decisions about their international money transfers.

This study explores the design, development, and evaluation of the Remittance Chatbot, focusing on its potential to enhance financial literacy, improve access to remittance services, and ultimately contribute to more efficient and cost-effective international money transfers. Through this research, we seek to demonstrate how AI-powered conversational interfaces can address real- world challenges in the remittance sector and contribute to broader goals of financial inclusion and economic empowerment.

The following sections will delve into the background of remittances and chatbot technologies, review relevant literature, detail our methodology, present our findings, and discuss the implications and potential future directions for this technology in the context of global financial services.

LITERATURE REVIEW

Remittances had a positive and significant effect on economic growth in developing countries from 2001 to 2013. There exists a nonlinear relationship between these two which are influenced by financial development and investment levels [3]. In Nigeria, there are mixed results on the effects of migration and remittances on health and education but evidence suggests that remittance- receiving households tend to prioritize education spending, leading to improved school enrolment and educational outcomes for children [1].

Azam and Gubert 2005 discussed how migration and remittances in Africa are collective decisions made by families to diversify risks and build social networks. However, the remaining family members may exert less effort due to the expectation of compensation from migrants, impacting technical efficiency in agriculture.[2] Remittance inflows positively impact financial development with lower elasticity values in developing countries like China, India, El Salvador, and the Philippines.[4] However, the significance of foreign direct investment and institutional setups need to be emphasized to enhance financial development across countries.

Financial development indicators correlate positively with economic growth. While Panel tests reveal cross-sectional dependence and integration of series, causality tests show remittances drive financial development, with exceptions in some indicators [4].

Haoyuan Wang [5] in his paper discusses various models used in chatbot development, focusing on Natural Language Processing (NLP), Machine Learning (ML), and the shift towards Neural Networks in building chatbots, indicating a trend toward more sophisticated chatbot technologies. The implications of this shift on chatbot performance, scalability, and maintenance in practical settings are yet to be fully understood. The ethical considerations surrounding chatbot usage in the financial industry, such as data privacy, security, and potential biases in chatbot interactions, are crucial aspects to consider in the deployment of chatbots. Overall, the paper underscores the importance of leveraging cutting-edge technologies like NLP, ML, and Neural Networks to develop efficient and effective chatbots that can revolutionize customer service delivery in the service industry.

Ionuț-Alexandru Cîmpeanu, et al [6] also researched on how Artificial Intelligence (AI) is utilized in banking through chatbots to automate tasks and enhance efficiency, focusing on five applications in the international banking system. The paper emphasizes the benefits of AI in banking and predicts a future dominated by such IT applications. Overall, the paper demonstrates the significant impact of AI chatbots in transforming traditional banking practices, paving the way for a more technologically advanced and customer-centric banking experience.

Advances in Natural Language Processing (NLP) and Generative AI’s ability to understand texts, images, sounds, and associated contexts for conversation have inspired intelligent applications such as AI-powered chatbots. AI chatbot development depends substantially on fine-tuning pre- trained Large Language Models (LLMs) to understand input and generate conversational human- like output accordingly. LLAMA, known as Language Model Adaptation is an open-source LLM developed by Meta. LLAMA is a pre-trained large text data up to 70B parameters that understand and generate human-like text. A simple technique for building an AI-chatbot with LLAMA involves integrating a chain that combines user prompts from an interface to the LLAMA model.

Streamlit library can be used as user interface to accept input as a prompt. LangChain then takes this input and passes it through a defined prompt template. The output from the prompt template is fed to the LLAMA language model. The LLAMA model processes this input and generates a conversational response based on its training. [7]

METHODOLOGY

Our research methodology for developing the Remittance Chatbot followed a systematic approach, incorporating various techniques and models to achieve the most effective solution. The process involved several key stages:

1. Data Collection and Preprocessing:

We collected a diverse set of remittance-related data, including transactional data, regulatory information, FAQ repositories, and domain-specific knowledge bases. This data was sourced from reputable financial institutions, government agencies, and international organizations. The collected data underwent rigorous preprocessing, including cleaning, structuring, and transformation into a format suitable for training the chatbot model.

2. Knowledge Base Construction:

Using the preprocessed data, we constructed a comprehensive knowledge base that encapsulated relevant information required to answer remittance-related queries effectively. This knowledge base served as the primary source of information for the chatbot model during training and inference.

3. Model Exploration and Implementation:

We explored various natural language processing (NLP) and machine learning approaches to develop an effective chatbot model. Our exploration included:

a) Model Finetuning with PDF Data:

• We utilized PDFs on remittance collected from authoritative sources.
• Implemented the “bert-large-uncased-whole-word-masking-finetuned-squad” model from the Langchain library.
• Developed using Streamlit for the user interface.
• Created embeddings and vector stores to enhance query understanding and response accuracy.

b) OLLAMA Models:

• Explored models including Gemma, Gemma2, Mistral, and Llama3.
• Implemented similar training processes as the PDF-based approach.

c) Hugging Face Models:

• Tested GPT-2, GPT-Medium, and GPT-2 Large models.
• Followed similar training procedures as previous approaches.

d) OpenAI GPT-3.5-Turbo:

• Implemented the GPT-3.5-Turbo model, which showed promising results in terms of intelligence and performance on CPU-based systems.

4. Model Evaluation and Selection:

Each model underwent rigorous evaluation, considering factors such as response accuracy, relevance, processing speed, and resource requirements. We employed techniques like cross- validation and holdout testing to assess model performance. Based on these evaluations, we identified GPT-3.5-Turbo as the most suitable model for our purposes, despite the challenge of API access costs.

5. Iterative Training and Refinement:

The selected model underwent iterative training and refinement cycles. We continuously monitored its performance, fine-tuned parameters, and incorporated feedback from simulated user interactions to improve accuracy and relevance over time.

6. Integration and Deployment:

• Local Deployment: We successfully implemented a local deployment using Streamlit, a popular Python library for creating web applications. This allowed us to create a functional user interface for the chatbot, demonstrating its capabilities in a controlled environment.
• User Interface: The Streamlit-based interface provides users with an intuitive platform to input their remittance-related queries and receive responses from our trained model.
• Testing: We conducted initial testing of the chatbot’s functionality and user experience through this local deployment.

We suggest that future work should focus on:

• Implementing cloud deployment for global accessibility.
• Enhancing security measures for user data protection.
• Conducting more extensive real-world testing.
• Developing APIs for integration with existing remittance systems.
• Collecting and incorporating user feedback for continuous improvement.

In spite of the limitations encountered in the course of this research I’ll, our local deployment with Streamlit represents a significant milestone, proving the concept’s viability and setting a strong foundation for future development and broader implementation of the Remittance Chatbot.

4. RESULT

Our research and development efforts in creating a Remittance Chatbot yielded several significant outcomes:

1. Model Performance:

• After exploring various models, we found that the OpenAI GPT-3.5-Turbo model demonstrated the best performance in terms of understanding and responding to remittance-related queries.
• The model showed high accuracy in interpreting user intents and providing relevant information from our constructed knowledge base.

2. Local Deployment:

• We successfully deployed the chatbot locally using Streamlit, creating a functional and user- friendly interface.
• Initial testing of this deployment showed promising results in terms of user interaction and query response accuracy.

3. Knowledge Base Effectiveness:

• Our curated knowledge base, constructed from diverse remittance-related data sources, proved effective in providing comprehensive and accurate information.
• The chatbot demonstrated the ability to handle a wide range of queries, from basic remittance concepts to specific regulatory information.

4. Challenges Identified:

• CPU-based training led to slower response times in some models.
• Some models (e.g., certain OLLAMA and Hugging Face models) showed limitations in maintaining context or providing consistently relevant answers.
• API access and associated costs presented a challenge for sustained use of the GPT-3.5-Turbo model.

5. User Experience:

• Preliminary feedback from test users indicated a positive response to the chatbot’s ease of use and the relevance of its responses.

5. CONCLUSION

The development of our Remittance Chatbot represents a significant step towards leveraging AI and natural language processing to address challenges in accessing remittance information. Our research demonstrates the potential of conversational AI in simplifying complex financial processes and improving financial literacy.

Key conclusions from our study include:

1. Viability of AI-powered chatbots in the remittance sector: Our project proves the concept that a well-trained AI model can effectively handle diverse remittance-related queries, potentially reducing barriers to information access.
2. Importance of model selection: The superior performance of GPT-3.5-Turbo highlights the importance of choosing the right model for domain-specific applications.
3. Value of comprehensive knowledge bases: Our approach of creating a detailed, domain-specific knowledge base significantly enhanced the chatbot’s ability to provide accurate and relevant information.
4. Potential for scalability: While our deployment was limited to a local environment, the successful implementation using Streamlit indicates the potential for broader deployment and integration.
5. Areas for future development: Our research identified several areas for improvement and expansion, including global deployment, enhanced integration with financial systems, and continuous model refinement based on user interactions.

In conclusion, while our Remittance Chatbot is still in its early stages, it demonstrates significant promise in revolutionizing how individuals access and understand remittance-related information. Future work should focus on overcoming the identified challenges, particularly in terms of deployment scalability and sustainable model access. With further development, such AI-powered tools could play a crucial role in enhancing financial inclusion and empowering individuals in the global remittance ecosystem.

REFERENCES

1. Okodua, H., Ewetan, O. O., & Urhie, E. (2015). Remittance expenditure patterns and human development outcomes in Nigeria. Developing Country Studies, 5(2), 70–80.
2. Azam, J. P., & Gubert, F. (2005). Migrant remittances and economic development in Africa: A review of evidence.
3. Eggoh, J., Bangake, C., & Semedo, G. (2019). Do remittances spur economic growth? Evidence from developing countries. The Journal of International Trade & Economic Development, 28(4), 391–418.
4. Bhattacharya, M., Inekwe, J., & Paramati, S. R. (2018). Remittances and financial development: empirical evidence from heterogeneous panel of countries. Applied Economics, 50(38), 4099- 4112.
5. Haoyuan, Wang. (2023). Chatbot in the Service Industry: Challenges and Perspectives. Highlights in Science Engineering and Technology, doi: 10.54097/hset.v57i.10025
6. Ionuț-Alexandru, Cîmpeanu., Denis-Alexandru, Dragomir., Razvan, Daniel, Zota. (2023). Banking Chatbots: How Artificial Intelligence Helps the Banks. Proceedings of the … International Conference on Business Excellence, 17:1716–1727. doi: 10.2478/ picbe-2023–0153
7. Rittika Jindal, Apr 12, 2024, “Gen AI –Part 3: Building a Chatbot with LLAMA and Streamlit: A Beginner’s Guide”, accessed 9 July 2024, <https://rittikajindal.medium.com/gen-ai-part-3- building-a-chatbot-with-llama-and-streamlit-a-beginnersguide-78642be5b9c6>