Fin-Ally: Pioneering the Development of an Advanced, Commonsense-Embedded Conversational AI for Money Matters
Sarmistha Das, Priya Mathur, Ishani Sharma, Sriparna Saha, Kitsuchart Pasupa, Alka Maurya · 2025
Read Fin-Ally: Pioneering the Development of an Advanced, Commonsense-Embedded Conversational AI for Money Matters by Sarmistha Das, Priya Mathur, Ishani Sharma, Sriparna Saha, Kitsuchart Pasupa, Alka Maurya (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2509-24342.txt.
Introduction
A finance chatbot that jokes about money can lose a user's trust in one line. Sarmistha Das and 5 co-authors tackle that tone problem with Fin-Ally, a conversational system for money matters. It embeds commonsense context into its replies and is tuned with preference optimization to sound professional and human-aligned at once.
The team also built Fin-Vault, a data set of well over a thousand annotated multi-turn finance dialogues. The system goes past account queries to budgeting help, live expense tracking, and automated planning. Tests show commonsense context made guidance more refined and precise in wording. The authors offer the pair as a next-generation base for finance chatbots.
Catalog record
| CREATOR | Sarmistha Das, Priya Mathur, Ishani Sharma, Sriparna Saha, Kitsuchart Pasupa, Alka Maurya |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 43 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2509-24342
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Fin-Ally: Pioneering the Development of an Advanced, Commonsense-Embedded Conversational AI for Money Matters Sarmistha Das, Priya Mathur, Ishani Sharma, Sriparna Saha, Kitsuchart Pasupa, Alka Maurya Abstract The exponential technological breakthrough of the FinTech industry has significantly enhanced user engagement through sophisticated advisory chatbots. However, large-scale fine-tuning of LLMs can occasionally yield unprofessional or flippant remarks, such as “With that money, you’re going to change the world,” which, though factually correct, can be contextually inappropriate and erode user trust. The scarcity of domain-specific datasets has led previous studies to focus on isolated components, such as reasoning-aware frameworks or the enhancement of human-like response generation. To address this research gap, we present Fin-Solution 2.O, an advanced solution that 1) introduces the multi-turn financial conversational dataset, Fin-Vault , and 2) incorporates a unified model, Fin-Ally , which integrates commonsense reasoning, politeness, and human-like conversational dynamics. Fin-Ally is powered by COMET-BART-embedded commonsense context and optimized with a Direct Preference Optimization (DPO) mechanism to generate human-aligned responses. The novel Fin-Vault dataset, consisting of 1,417 annotated multi-turn dialogues, enables Fin-Ally to extend beyond basic account management to provide personalized budgeting, real-time expense tracking, and automated financial planning. Our comprehensive results demonstrate that incorporating commonsense context enables language models to generate more refined, textually precise, and professionally grounded financial guidance, positioning this approach as a next-generation AI solution for the FinTech sector. † † paperid: 5108 † † orcid: 0000-0001-6608-0400 1 Introduction The rapid proliferation of financial chatbots has reshaped customer service and operational efficiency within the financial sector. Market projections indicate a compound annual growth rate of 23.3%, with the global chatbot market expected to reach $15.5 billion by 2028 [ Botpress(2024) ] . This growth is fueled by the increasing need for continuous customer support and cost-effective service delivery. With the expeditious growth of financial LLMs such as FinGPT [ Liu et al.(2023)Liu, Wang, and Zha ] , FinLLAMA [ Konstantinidis et al.(2024)Konstantinidis, Iacovides, Xu, Constantinides, and Mandic ] , InvestLLM [ Yang et al.(2023)Yang, Tang, and Tam ] , and FinROBOT [ Zhou et al.(2024)Zhou, Wang, Wu, and Yang ] , the financial sector has witnessed substantial enhancements in automation. Despite their expanding adoption, existing financial chatbots frequently fall short due to limited commonsense reasoning, which is crucial for accurately managing everyday financial interactions. Early chatbot implementations were often generic and lacked domain-specific contextual understanding, limiting their ability to resolve complex financial queries. For example, when asked about paying utility bills with a credit card, conventional chatbots might provide irrelevant responses such as, “ You can only use your card for shopping ,” underscoring a critical deficiency in contextual comprehension [ Chatbot.com(2024) , LinkedIn(2024) ] . This persistent inadequacy has culminated in notable user dissatisfaction, underscoring the critical need for more sophisticated, context-aware financial assistants. While advanced decoder-only language models such as GPT-3.5 [ Wang et al.(2023)Wang, Cheng, Zhan, Li, Song, and Liu ] and its variants deliver impressive conversational fluency, they often falter in maintaining coherence across domain-specific, multi-turn dialogues. Moreover, large-scale models such as Bloomberggpt [ Wu et al.(2023)Wu, Irsoy, Lu, Dabravolski, Dredze, Gehrmann, Kambadur, Rosenberg, and Mann ] and GPT-4o [ Hurst et al.(2024)Hurst, Lerer, Goucher, Perelman, Ramesh, Clark, Ostrow, Welihinda, Hayes, Radford, et al. ] , despite their enhanced capabilities, remain proprietary, rendering their architectures inaccessible for fine-grained task-specific adaptation, which poses significant challenges for domain customisation. Compounding this, the sheer resource intensity of these models frequently results in a degradation of response quality over extended interactions; for example, a user query may elicit an unprofessional reply such as, ”Okay! Let me spoon-feed you,” which can erode user trust and diminish engagement. The demand for intuitive chatbots is underscored by the fact that 80% of financial institutions regard AI-driven conversational agents as strategic assets for enhancing customer engagement [ of Code(2024) ] , while 43% of U.S. banking clients prefer chatbots for resolving financial queries, reflecting growing reliance on automated financial services. Figure 1: An advanced commonsense-aware financial chatbot capable of coherent multi-turn conversations. Where commonsense context enriches situational needs, reasons, wants, and effects Motivation: Existing research addressing this task has predominantly concentrated on isolated components—such as dataset benchmarking [ Yuan et al.(2024)Yuan, Wang, Zhu, Yuan, Zhou, Zhu, and Wei , Xie et al.(2024)Xie, Han, Chen, Xiang, Zhang, He, Xiao, Li, Dai, Feng, et al. ] , reasoning enhancement [ Srivastava et al.(2024)Srivastava, Malik, Gupta, Ganu, and Roth ] , or improved alignment [ Ouyang et al.(2024)Ouyang, Yun, and Zheng ] —while overlooking critical dimensions such as contextual comprehension and situational awareness, both of which are pivotal for navigating complex, real-world financial interactions. Conventional chatbots often generate contextually inappropriate responses; for example, when asked, “ What should I do if my debit card is stolen? ” a typical bot might respond with, “ Visit the nearest ATM to check your balance ” reflecting a severe lack of commonsense reasoning. To address these limitations, we present an ultimate solution, Fin-Solution2.O, that 1) introduces the multi-turn financial conversational dataset Fin-Vault , and 2) incorporates a unified model, Fin-Ally , which integrates commonsense reasoning, politeness, and human-like conversational dynamics. Through Fin-Ally , we aim to study the incorporation of each component’s importance for dynamically interpreting user queries and providing contextually appropriate responses, such as recommending immediate card blocking and reporting in case of theft as described in Figure 1 . The Fin-Vault dataset is meticulously annotated with 1,417 multi-turn conversations, enriched with politeness categories to enhance Fin-Ally ’s user-friendliness. To ensure responses are closely aligned with human preferences, the model leverages Direct Preference Optimization (DPO), enabling the generation of refined, contextually appropriate, and human-aligned conversational outputs. Research Objectives: 1. To explore the integration of commonsense reasoning for generating coherent, context-aware multi-turn dialogues within the financial domain, effectively addressing routine monetary concerns. 2. To rigorously assess the extent to which DPO enhances alignment with human conversational preferences, contributing to more natural and contextually appropriate responses. Contributions: Our key contributions are as follows: 1. Initially, we introduce Fin-Solution 2.O , an integrated solution designed to address everyday financial tasks, comprising a unique, domain-specific dataset paired with a user-centric conversational AI assistant. 2. Subsequently, we present Fin-Vault , a pioneering financial dataset consisting of 1,417 meticulously annotated multi-turn dialogues spanning key financial sectors, including banking, credit and debit card management, insurance, and market investments. 3. Further, we introduce Fin-Ally 1 1 1 Fin-Ally: https://github.com/sarmistha-D/Fin-Ally . , the first commonsense-aware financial chatbot augmented with DPO, engineered to deliver contextually enriched and human-aligned conversations, establishing a new benchmark in AI-driven financial dialogue systems. 4. Moreover, we conduct comprehensive empirical evaluations under both zero-shot and fine-tuned settings, benchmarking multiple state-of-the-art LLMs to determine optimal architectures for task-specific financial chatbot deployment. Our findings provide an in-depth analysis of LLM performance, adaptability, and robustness within specialized financial dialogue contexts. 2 Related Works The progression of financial chatbots mirrors the broader evolution of conversational AI, advancing from basic rule-based systems to sophisticated AI-powered agents. Early models lacked contextual understanding, but breakthroughs in machine learning and NLP have enabled personalised, context-aware interactions [ Unknown(2022) ] . The latest architectures harness LLMS and Retrieval-Augmented Generation (RAG) frameworks, enabling real-time data access and hyper-personalised financial advisory services [ IJERT(2024) ] . Recent innovations spotlight domain-specific