Abstract
Frame-semantic parsing (FSP) aims to extract frame-semantic structures from text, involving three subtasks: target identification (TI), frame identification (FI), and frame-semantic role labeling (FSRL). Most previous studies focus on only one or two of these subtasks, and limited research exists on constructing a unified framework to solve the entire FSP task systematically. Furthermore, efficiently utilizing lexical unit knowledge and improving span representations remain open questions. This work presents a novel unified framework for FSP based on marker attention. Special markers are strategically inserted into the input sequence to incorporate lexical unit knowledge and frame information, enhancing span representations through attention mechanisms without additional complex modules. Experiments on two public benchmarks demonstrate the effectiveness of the proposed framework.
Introduction
Frame-semantic Parsing (FSP) is a core task in natural language processing that extracts semantic structures based on FrameNet theory. It identifies semantic scenarios (frames) and their participants (arguments), answering the fundamental question of "who did what to whom." The process is traditionally divided into three subtasks: Target Identification (TI), Frame Identification (FI), and Frame-semantic Role Labeling (FSRL).
Despite its importance, unified FSP models face challenges such as "span missing," where overlapping targets and arguments are not correctly identified. Moreover, representing complex, long-range arguments effectively is a persistent hurdle. This paper introduces a two-stage framework that leverages a novel "Marker Attention" mechanism to address these limitations.
Methodology
The proposed framework decomposes the FSP task into two structurally similar stages, allowing for more specialized handling of target and argument identification.
Two-Stage Architecture
- Joint TI and FI (TFIMA): This stage identifies potential targets and assigns them frames. Since both are span-level classification problems, they are handled jointly. The model uses paired levitated markers (
<c>and</c>) to wrap candidate target spans. - FSRL (FSRLMA): This stage is treated as a relation classification problem between identified targets and potential arguments. It utilizes single levitated markers to represent argument spans. To enhance efficiency, the model employs graph-based pruning using dependency graphs to eliminate grammatically incoherent argument spans, reducing the search space by 50%.
Marker Attention and Saliency Vectors
The core innovation is the Marker Attention mechanism. Instead of building complex neural architectures, the authors modify the Transformer's attention matrix. Markers inserted into the input are permitted to attend to their corresponding text spans, while original tokens are restricted from attending to the markers. This ensures that markers capture rich, span-specific representations without disrupting the natural linguistic flow of the sentence.
To guide the model, Saliency Vectors are introduced. These vectors act as a knowledge-guided mechanism by flagging potential targets based on FrameNet's lexical units. By encoding this external knowledge directly into the input representation, the model learns to prioritize relevant parts of the sentence during the attention process.
Experiments and Results
The framework was evaluated on the FrameNet 1.5 and 1.7 benchmarks, as well as the MASC corpus for out-of-domain testing.
Performance Benchmarks
The model achieved new state-of-the-art (SOTA) results across all subtasks. On the FrameNet 1.7 dataset, it reached F1 scores of 79.19 for TI, 71.88 for FI, and 55.88 for FSRL. Notably, the marker-based approach demonstrated significant robustness in out-of-domain scenarios, showing a recall improvement of up to 22.8% on the MASC dataset compared to baseline models.
Comparison with Large Language Models
A key finding is the performance gap between specialized FSP models and general-purpose Large Language Models (LLMs). In head-to-head FSRL tests, the FSRLMA model (75.00 F1) significantly outperformed LLaMA-2-13B-chat (26.56 F1). This suggests that abstract semantic role labeling remains a major weakness for general LLMs, highlighting the continued necessity of specialized semantic parsing frameworks.
Conclusion and Future Work
This work demonstrates that a unified framework based on marker attention can effectively solve the complexities of Frame-semantic Parsing. By separating the task into two stages and utilizing knowledge-guided attention, the model achieves superior performance and robustness. Future research will explore the integration of Constructional Grammar knowledge and the potential of instruct fine-tuning to enhance the semantic capabilities of LLMs in the context of FrameNet.
Summary updated