Research map / Label-efficient and robust learning / Recognition and long-tail
Explainable Earnings Call Representation Learning (Student Abstract)
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a structure-aware, information-bottleneck key-sentence extractor for earnings call transcripts, trained self-supervised to predict Q&A content, and evaluated on risk forecasting and information retrieval.
Problem
Earnings call transcripts are long and complex, making manual analysis slow and error-prone, and black-box transcript representations are opaque and unreliable for finance tasks.
Approach
A key sentence extractor selects a compressed subset of transcript sentences, trained with an infoNCE objective that links selected Presentation sentences to the Q&A section plus a variational information bottleneck penalty for conciseness, so the chosen sentences serve as explanations.
Key results
- On risk forecasting, the proposed model reaches MSE(3d) 0.7371 vs 0.7401 for BERT and 0.7868 for MR-QA, at a compression ratio α=0.6.
- On information retrieval, precision is 0.9400 vs 0.9300 for PMI and 0.8380 for TextRank, at α=0.15.
- Mean Rank for the proposed method (2.4020) is not best; PMI attains 2.1020 MR and higher MRR (0.9595 vs 0.9602 is ours, marginally higher).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the proposed transcript representation instead of BERT | Risk forecasting · 3d | MSE: 0.7401 -> 0.7371 | Helped |
| Use the proposed transcript representation instead of MR-QA | Risk forecasting · 15d | MSE: 0.2561 -> 0.2633 | Hurt |
| Use the proposed transcript representation instead of MR-QA | Risk forecasting · 60d | MSE: 0.1792 -> 0.1789 | Helped |
| Use the proposed transcript representation instead of BERT | Risk forecasting · 3d | MAE: 0.6724 -> 0.6711 | Helped |
| Use the proposed transcript representation instead of MR-QA | Risk forecasting · 15d | MAE: 0.3888 -> 0.3983 | Hurt |
| Use the proposed transcript representation instead of MR-QA | Risk forecasting · 60d | MAE: 0.3170 -> 0.3177 | Hurt |
| Use the proposed transcript representation instead of PMI | Information retrieval · Information retrieval | Precision: 0.9300 -> 0.9400 | Helped |
| Use the proposed transcript representation instead of PMI | Information retrieval · Information retrieval | MR: 2.1020 -> 2.4020 | Hurt |
| Use the proposed transcript representation instead of PMI | Information retrieval · Information retrieval | MRR: 0.9595 -> 0.9602 | Helped |
Surprising findings
- The model matches or slightly beats full-language-model baselines while keeping only 60% (risk) or 15% (retrieval) of sentences.
Limitations
- Results are reported on a single dataset split from U.S. firms with limited baselines.
- Only two downstream tasks are evaluated, and the paper is a short student abstract with little detail.
Experimental setup
Datasets: Earnings call transcripts from U.S. firms (SeekingAlpha, Thomson Reuters StreetEvents)
Benchmarks: Earnings call transcript risk forecasting (MSE, MAE at 3d/15d/60d) Earnings call transcript information retrieval (Precision, MR, MRR)
Metrics: MSE MAE Precision Mean Rank Mean Reciprocal Rank
Baselines: BERT SimCSE Profet MR-QA LexRank TextRank PMI
Builds on: information bottleneck infoNCE SimCSE
Closest papers by meaning
- Unraveling the Influence of Training Data and Internal Structures in Large Language Models for Enhanced Explainability (Student Abstract) (AAAI 2025 · similarity 0.81)
- The SEC Filings Dataset: Reconstructing U.S. Corporate and Financial Disclosures into Layout-Faithful and Token-Efficient Pretraining Data (NeurIPS 2026 · similarity 0.81)
- Explainable Neural Networks with Guarantee: A Sparse Estimation Approach (AAAI 2025 · similarity 0.80)
- Linking Industry Sectors and Financial Statements: A Hybrid Approach for Company Classification (AAAI 2025 · similarity 0.80)
- SubjECTive-QA: Measuring Subjectivity in Earnings Call Transcripts' QA Through Six-Dimensional Feature Analysis (NeurIPS 2024 · similarity 0.80)
- Predictive Concept Decoders: Training Scalable End-to-End Interpretability Assistants (NeurIPS 2026 · similarity 0.80)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.