NuAegis: Earnings Call Transcripts Summarization, Sentiment Analysis and Conversational AI Integration

Project Overview

This project establishes the use of Natural Language Processing techniques to analyze Earnings Call Transcripts (ECTs). It performs tasks such as summarization, sentiment-based summarization and interactive Q&A using a chatbot. The final solution is implemented as an Android application, NuAegis, aimed at providing users with concise insights on financial reports.

Table of Contents

1. Objectives
2. Dataset
3. Methodology
   • Data Collection
   • Summarization
   • Sentiment-based Summarization
   • Conversational AI (Chatbot)
4. Results
5. User Interface
6. Installation
7. Usage
8. References

Objectives

• Provide users with a summarized version of extensive earnings call transcripts.
• Generate sentiment-based summaries, offering both positive and negative viewpoints.
• Develop a chatbot for interactive financial Q&A which allows personalized inquiries.

Dataset

This project uses ECTs from the top 50 companies, scraped from AlphaSpread (www.alphaspread.com) and organized by company, year and quarter. Each transcript provides information about key financial metrics enabling analysis across various time frames.

Methodology

Data Collection

• ECTs were scraped using BeautifulSoup and Requests libraries, subsequently cleaned and stored as text files organized by respective company, year and quarter.
• Standard preprocessing techniques (tokenization, lowercasing) were applied to the text.

Summarization

• The BART transformer model was used for summarization, handling large transcripts by splitting them into manageable chunks.
• Generated summaries capture critical financial metrics, strategies and market sentiments for each earnings call.

Sentiment-based Summarization

1. Sentiment Analysis: Using a BiLSTM model with GloVe 100-d embeddings and multi-head attention mechanism, the model achieved 90% accuracy on the Financial Phrasebank dataset.
2. Concise Summarization: Utilizing the T5 model fine-tuned on the ECTSum dataset, the model generates sentiment-based summaries, classified as positive or negative with a context window of three sentences.

Conversational AI (Chatbot)

• The chatbot module utilizes FinBERT embeddings for retrieving contextually relevant chunks of data based on user queries.
• Google Gemini model generates responses to user inquiries, enhanced with prompt engineering techniques for relevance and accuracy.

Results

• Summarization: ROUGE-1 F1 Score of 0.2090 and ROUGE-L F1 Score of 0.1994.
• Sentiment Analysis: Test accuracy of 90% with BiLSTM and multi-head attention.
• Concise Summarization: ROUGE-1 score of 0.356, ROUGE-2 score of 0.237, ROUGE-L score of 0.326, with a high BERTScore F1 of 0.879.

User Interface

NuAegis - A Financial Advisor app summarizes ECTs calls, analyses the sentiment of the transcripts and presents them also in a summary and has a chatbot to discuss details about a particular company's ECT or compare companies performance.