The critical analysis and understanding of financial texts largely depend on techniques like intent classification and entity extraction. The former technique aims to discern a text's underlying intention or purpose, whereas the latter focuses on identifying and extracting specific data from the text. These approaches play a pivotal role in streamlining and automating the analysis and processing of financial documents, making them more understandable to machine learning systems. This dissertation offers an in-depth exploration of these techniques as applied in the financial sphere, discussing their challenges and potential. The primary objective is to give a complete picture of the current cutting-edge methodologies in this field while also providing useful insights for both researchers and practitioners intrigued by the application of these techniques in the analysis of financial documents. We utilized a comprehensive approach, Multiple Novel Intent Detection (MNID), which identifies and classifies new intents in financial reports and documents. The MNID framework effectively manages new, unexplored intents with the help of clustering, annotation, and retraining methods. Our assessment of the MNID framework using financial natural language understanding datasets (Finer-139) demonstrated performance superior to other established benchmarks.
The MNID framework, a novel approach for detecting and categorizing new intents in conversational agents was employed in this study. By combining clustering, annotation, and retraining, the MNID framework adeptly manages new intents that were previously unencountered. When evaluated on a Natural Language Understanding (NLU) dataset, Finer-139, this framework exhibited superior performance when compared to competitive baselines (instances where data points for each new class are provided).