The Mask One at a Time Framework for Detecting the Relationship Between Financial Entities

EasyChair Preprint 10463

Abstract

In the financial domain, understanding the relationship between two entities helps in understanding financial texts. In this paper, we introduce the Mask One At a Time (MOAT) framework for detecting the relationship between financial entities. Subsequently, we benchmark its performance with the existing state-of-the-art discriminative and generative Large Language Models (LLMs). We use the SEC-BERT embeddings along with the one-hot encoded vectors of the types of entities and their relation group as features. We benchmark MOAT with three such open-source LLMs, namely, Falcon, Dolly and MPT under zero-shot and few shot settings. The results prove that MOAT outperforms these LLMs.

Keyphrases: Financial texts, Relation Extraction, large language models

@booklet{EasyChair:10463,
  author    = {Sohom Ghosh and Sachin Umrao and Chung-Chi Chen and Sudip Kumar Naskar},
  title     = {The Mask One at a Time Framework for Detecting the Relationship Between Financial Entities},
  howpublished = {EasyChair Preprint 10463},
  year      = {EasyChair, 2023}}