Pro-HAN: A Heterogeneous Graph Attention Network for Profile-Based Spoken Language Understanding

TL;DR

Pro-HAN addresses the ambiguity in profile-based SLU by jointly reasoning over multiple profile sources—Knowledge Graph, User Profile, and Context Awareness—using a heterogeneous graph attention network with edge types that capture intra-profile structure, cross-profile interactions, and utterance-driven relevance. The model encodes UP/CA triplets, KG triplets, and the utterance, builds a graph with utterance, triplet, and global nodes, and performs L-layer heterogeneous graph attention to obtain a unified representation for intent and slot prediction. Empirical results on the ProSLU dataset show Pro-HAN achieving state-of-the-art performance, with about 8% improvements across F1, intent accuracy, and overall accuracy, and ablations confirm the critical roles of intra-Pro, inter-Pro, and utterance-Pro connections as well as heterogeneity. The work demonstrates that adaptive, cross-source reasoning over multi-source profiles can significantly mitigate utterance ambiguity, with practical implications for real-world task-oriented dialogue systems, and provides publicly available code for replication.

Abstract

Recently, Profile-based Spoken Language Understanding (SLU) has gained increasing attention, which aims to incorporate various types of supplementary profile information (i.e., Knowledge Graph, User Profile, Context Awareness) to eliminate the prevalent ambiguities in user utterances. However, existing approaches can only separately model different profile information, without considering their interrelationships or excluding irrelevant and conflicting information within them. To address the above issues, we introduce a Heterogeneous Graph Attention Network to perform reasoning across multiple Profile information, called Pro-HAN. Specifically, we design three types of edges, denoted as intra-Pro, inter-Pro, and utterance-Pro, to capture interrelationships among multiple Pros. We establish a new state-of-the-art on the ProSLU dataset, with an improvement of approximately 8% across all three metrics. Further analysis experiments also confirm the effectiveness of our method in modeling multi-source profile information.

Figures (1)

• Figure 1: The illustration of Pro-HAN. The initial representations of graph nodes are introduced in Section \ref{['sec:encoder']}.