AS-SRL: A Chinese Speech-based Semantic Role Labeling Dataset

Description

AS-SRL is the first Chinese speech-based Semantic Role Labeling (SRL) dataset, created by annotating the open-source Mandarin speech corpus AISHELL-1 with semantic role labels following the guidelines of Chinese Proposition Bank 1.0 (CPB1.0). The dataset contains 9,000 speech-text pairs with corresponding SRL annotations, split into training (7,500), development (500), and test (1,000) sets.

This dataset was developed to address the growing need for speech-based language understanding capabilities, particularly for SRL from speech input directly. It enables research on end-to-end approaches for SRL from speech, which can help overcome limitations of traditional pipeline methods (ASR followed by text-based SRL) such as error propagation and loss of useful acoustic features.

Features

• 9,000 authentic Mandarin speech recordings with corresponding transcripts and SRL annotations
• 27,113 predicate-argument tuples across all splits
• 17 semantic role types including core arguments (ARG0-ARG4) and various modifiers (ARGM-*)
• High-quality annotations verified through a rigorous multi-annotator process with 86% inter-annotator agreement

Citation

If you use this dataset, please cite:

@inproceedings{chen-etal-2024-semantic,
title = "Semantic Role Labeling from Chinese Speech via End-to-End Learning",
author = "Chen, Huiyao and Li, Xinxin and Zhang, Meishan and Zhang, Min",
booktitle = "Findings of the Association for Computational Linguistics: ACL 2024",
year = "2024",
pages = "8898--8911"
}