CroPrompt: Cross-task Interactive Prompting for Zero-shot Spoken Language Understanding

TL;DR

CroPrompt tackles zero-shot spoken language understanding by explicitly enabling cross-task interaction between intent detection and slot filling through a two-stage prompting pipeline. It further mitigates error propagation with a multi-task self-consistency mechanism that ensembles reasoning across routes for both sentence-level intents and token-level slot spans. Empirical results on SNIPS across multiple LLMs show CroPrompt consistently surpasses vanilla prompting and benefits substantially from self-consistency, with GPT-4 achieving strong sentence accuracy and slot F1. The approach also demonstrates token-cost reductions and better generalization to downstream dialogue tasks under SFT, suggesting cross-task prompting as a viable path for data-efficient SLU. Overall, CroPrompt establishes a principled framework for cross-task information exchange in prompt-based SLU with robust performance improvements.

Abstract

Slot filling and intent detection are two highly correlated tasks in spoken language understanding (SLU). Recent SLU research attempts to explore zero-shot prompting techniques in large language models to alleviate the data scarcity problem. Nevertheless, the existing prompting work ignores the cross-task interaction information for SLU, which leads to sub-optimal performance. To solve this problem, we present the pioneering work of Cross-task Interactive Prompting (CroPrompt) for SLU, which enables the model to interactively leverage the information exchange across the correlated tasks in SLU. Additionally, we further introduce a multi-task self-consistency mechanism to mitigate the error propagation caused by the intent information injection. We conduct extensive experiments on the standard SLU benchmark and the results reveal that CroPrompt consistently outperforms the existing prompting approaches. In addition, the multi-task self-consistency mechanism can effectively ease the error propagation issue, thereby enhancing the performance. We hope this work can inspire more research on cross-task prompting for SLU.

Figures (6)

• Figure 1: (a) Vanilla SLU Prompting directly utilizes a single conversation turn for prompting intent detection and slot filling without any interaction information while (b) CroPrompt considers explicit interaction across multiple tasks by directly incorporating the result of intent detection for slot filling.
• Figure 2: The workflow of CroPrompt, which first utilizes the preliminary task solution prompting for intent detection and then the follow-up task solution prompting is introduced for slot filling conditioned with the predicted intent results by task information interaction.
• Figure 3: Multi-Task Self-Consistency Prompting for Intent Detection and Slot Filling.
• Figure 4: Results of CroPrompt with different multi-task self-consistency methods. By different temperature utilizes the outputs of CroPrompt with different temperatures. By different prompt utilizes the outputs of three different prompt methods.
• Figure 5: Results of CroPrompt for Join SC and DAR on Mastodon.