Manoj Apte

Tokenization is the first step in training any Large Language Model (LLM), where the text is split into a sequence of tokens as per the model's fixed vocabulary. This tokenization in LLMs is different from the traditional tokenization in NLP where the text is split into a sequence of "natural" words. In LLMs, a natural word may also be broken into multiple tokens due to limited vocabulary size of the LLMs (e.g., Mistral's tokenizer splits "martial" into "mart" and "ial"). In this paper, we hypothesize that such breaking of natural words negatively impacts LLM performance on various NLP tasks. To quantify this effect, we propose a set of penalty functions that compute a tokenization penalty for a given text for a specific LLM, indicating how "bad" the tokenization is. We establish statistical significance of our hypothesis on multiple NLP tasks for a set of different LLMs.

India’s insurance dispute resolution system is burdened by long delays, escalating costs, and growing complaint volumes. With the insurance market projected to reach US$ 222 billion by 2026 and recent data showing that leading public insurers registered over 81,000 complaints in FY2023, traditional processes are no longer viable. This paper proposes a “Dispute Resolution as a Service” model for the insurance claim domain, enabled by AI technologies with service design principles. By rethinking dispute resolution as a service, the solution addresses not only the technical requirements for speed and cost reduction but also the softer aspects such as transparency, user-centeredness, fairness, resilience, and trust, that technology alone cannot deliver. The proposed service is designed to streamline claim processes, mitigate implicit bias, and customize procedures based on domain expertise. This integrated approach promises to save time, reduce judicial burdens, and enhance stakeholder satisfaction, underscoring the necessity of aligning service design in dispute resolution.

Disputes between two parties occur in almost all domains such as taxation, insurance, banking, healthcare, etc. The disputes are generally resolved in a specific forum (e.g., consumer court) where facts are presented, points of disagreement are discussed, arguments as well as specific demands of the parties are heard, and finally a human judge resolves the dispute by often favouring one of the two parties. In this paper, we explore the use of large language models (LLMs) as assistants for the human judge to resolve such disputes, as part of our DRAssist system. We focus on disputes from two specific domains -- automobile insurance and domain name disputes. DRAssist identifies certain key structural elements (e.g., facts, aspects or disagreement, arguments) of the disputes and summarizes the unstructured dispute descriptions to produce a structured summary for each dispute. We then explore multiple prompting strategies with multiple LLMs for their ability to assist in resolving the disputes in these domains. In DRAssist, these LLMs are prompted to produce the resolution output at three different levels -- (i) identifying an overall stronger party in a dispute, (ii) decide whether each specific demand of each contesting party can be accepted or not, (iii) evaluate whether each argument by each contesting party is strong or weak. We evaluate the performance of LLMs on all these tasks by comparing them with relevant baselines using suitable evaluation metrics.

Nitin Ramrakhiyani, Delton Myalil, Sachin Pawar, Manoj Apte, Rajan M A, Divyesh Saglani, Imtiyazuddin Shaik. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 3: Industry Track). 2025.

For different factors/reasons, ranging from inherent characteristics and features providing decentralization, enhanced privacy, ease of transactions, etc., to implied external hardships in enforcing regulations, contradictions in data sharing policies, etc., cryptocurrencies have been severely abused for carrying out numerous malicious and illicit activities including money laundering, darknet transactions, scams, terrorism financing, arm trades. However, money laundering is a key crime to be mitigated to also suspend the movement of funds from other illicit activities. Billions of dollars are annually being laundered. It is getting extremely difficult to identify money laundering in crypto transactions owing to many layering strategies available today, and rapidly evolving tactics, and patterns the launderers use to obfuscate the illicit funds. Many detection methods have been proposed ranging from naive approaches involving complete manual investigation to machine learning models. However, there are very limited datasets available for effectively training machine learning models. Also, the existing datasets are static and class-imbalanced, posing challenges for scalability and suitability to specific scenarios, due to lack of customization to varying requirements. This has been a persistent challenge in literature. In this paper, we propose behavior embedded entity-specific money laundering-like transaction simulation that helps in generating various transaction types and models the transactions embedding the behavior of several entities observed in this space. The paper discusses the design and architecture of the simulator, a custom dataset we generated using the simulator, and the performance of models trained on this synthetic data in detecting real addresses involved in money laundering.

Financial audits establish trust in the governance and processes in an organization, but they are time-consuming and knowledge intensive. To increase the effectiveness of financial audit, we address the task of generating audit suggestions that can help auditors to focus their investigations. Specifically, we present NLP techniques to extract hidden knowledge from a corpus of past financial audit reports of many companies, and use it for generating audit suggestions. The extracted knowledge consists of a set of automatically identified sentences containing adverse remarks, the financial variables mentioned in each sentence and automatically assigned XBRL categories for them, since XBRL is a standardized taxonomy in the financial domain. In the absence of suitable labeled data, we adopted a weak supervision approach. We designed a set of high precision linguistic rules to identify adverse remark sentences, created automatically labeled training data using them, and trained BERT-based and other classifiers to identify such sentences. We next presented novel techniques (which are either unsupervised or zero-shot) to assign zero, one, or more XBRL categories to any given adverse remark sentence. We evaluated the proposed approaches, on a large corpus of real-life financial statements and audit reports, against competent baselines. Given a company’s financial statements (already identified as suspicious), and given a subset of financial variables in them that contribute to suspiciousness, we match these with the extracted knowledge base and identify aligned adverse remarks that help the auditor in focusing on specific directions for further investigations.

Sushodhan Vaishampayan, Nitin Ramrakhiyani, Sachin Pawar, Aditi Pawde, Manoj Apte, Girish Palshikar. Proceedings of the Sixth Workshop on Financial Technology and Natural Language Processing. 2023.