SpeechEE@XLLM25: End-to-End Structured Event Extraction from Speech

Event extraction from text is a complex task that involves the identification of event triggers and their supporting arguments.When applied to speech, this task becomes even more challenging due to the continuous nature of audio signals and the need for robust Automatic Speech Recognition (ASR).This paper proposes an approach that integrates ASR with event extraction by utilizing the Whisper model for speech recognition and a Text2Event2 Transformer for extracting events from English audio samples.The Whisper model is used to generate transcripts from audio, which are then fed into the Text2Event2 Transformer to identify event triggers and their arguments.This approach combines two difficult tasks into one, streamlining the process of extracting structured event information directly from audio.Our approach leverages a robust ASR system (Whisper) followed by a parameter-efficient transformer (Text2Event2 fine-tuned via LoRA) to extract structured events from raw speech.Unlike prior work trained on gold textual input, our pipeline is trained end-to-end on noisy ASR outputs.Despite significant resource constraints and data noise, our system ranked first in the ACL 2025 XLLM Shared Task II.