FinDialogLens
A hybrid LLM event-extraction system designed for financial chatrooms. It combines fine-tuned classifiers with large-model pipelines to automatically identify and reconstruct RFQ trading events in multi-party dialogues, addressing the problem of missed-trade retrospection in financial sales contexts.
Timeline (last 90 days)
FinDialogLens reframes the problem of missed trades in financial chat rooms as an Event Extraction task over multi-party dialogues.
FinDialogLens employs a hybrid LLM pipeline in which small fine-tuned classifiers serve as scaffolding during inference to assist the large model in completing extraction tasks.
FinDialogLens's fine-tuned classifiers are responsible for detecting RFQ trigger messages and identifying metadata related to quotes and trade outcomes.
FinDialogLens's RFQ-Level Module segments dialogues into independent RFQ windows by event, and the Trade Engine fills in argument roles within each window to reconstruct complete trade events.
When paired with GPT-4o, FinDialogLens achieves 92.1% accuracy on final price extraction.
When paired with GPT-4o, FinDialogLens achieves 94.3% accuracy on trade outcome determination.
A fine-tuned open-source LLM with only 3B parameters can achieve results comparable to the GPT-4o solution given a moderate amount of in-domain data.
FinDialogLens introduces a difficulty-aware router that dynamically allocates RFQs between a low-cost rule engine and a high-performance LLM trade engine.
Difficulty-aware routing reduces LLM calls by 85% on the final price task while closing half of the accuracy gap relative to the pure GPT-4o solution.
At a scale of 70,000 RFQs per day, the difficulty-aware routing design can save more than $300 in costs per day.
All Facts (10)
FinDialogLens reframes the problem of missed trades in financial chat rooms as an Event Extraction task over multi-party dialogues.
50%UnverifiedFinDialogLens employs a hybrid LLM pipeline in which small fine-tuned classifiers serve as scaffolding during inference to assist the large model in completing extraction tasks.
50%UnverifiedFinDialogLens's fine-tuned classifiers are responsible for detecting RFQ trigger messages and identifying metadata related to quotes and trade outcomes.
50%UnverifiedFinDialogLens's RFQ-Level Module segments dialogues into independent RFQ windows by event, and the Trade Engine fills in argument roles within each window to reconstruct complete trade events.
50%UnverifiedWhen paired with GPT-4o, FinDialogLens achieves 92.1% accuracy on final price extraction.
50%UnverifiedWhen paired with GPT-4o, FinDialogLens achieves 94.3% accuracy on trade outcome determination.
50%UnverifiedA fine-tuned open-source LLM with only 3B parameters can achieve results comparable to the GPT-4o solution given a moderate amount of in-domain data.
50%UnverifiedFinDialogLens introduces a difficulty-aware router that dynamically allocates RFQs between a low-cost rule engine and a high-performance LLM trade engine.
50%UnverifiedDifficulty-aware routing reduces LLM calls by 85% on the final price task while closing half of the accuracy gap relative to the pure GPT-4o solution.
50%UnverifiedAt a scale of 70,000 RFQs per day, the difficulty-aware routing design can save more than $300 in costs per day.
50%