[Paper Review] A Novel Bi-directional Interrelated Model for Joint Intent Detection and Slot Filling

E Haihong, Peiqing Niu+2 more|arXiv (Cornell University)|Jun 30, 2019

Topic Modeling14 references95 citations

TL;DR

Introduces the SF-ID network that bi-directionally interrelates slot filling and intent detection, with an iteration mechanism to enhance mutual reinforcement, achieving improvements on ATIS and Snips datasets.

ABSTRACT

A spoken language understanding (SLU) system includes two main tasks, slot filling (SF) and intent detection (ID). The joint model for the two tasks is becoming a tendency in SLU. But the bi-directional interrelated connections between the intent and slots are not established in the existing joint models. In this paper, we propose a novel bi-directional interrelated model for joint intent detection and slot filling. We introduce an SF-ID network to establish direct connections for the two tasks to help them promote each other mutually. Besides, we design an entirely new iteration mechanism inside the SF-ID network to enhance the bi-directional interrelated connections. The experimental results show that the relative improvement in the sentence-level semantic frame accuracy of our model is 3.79% and 5.42% on ATIS and Snips datasets, respectively, compared to the state-of-the-art model.

Motivation & Objective

• Motivate joint SLU for slot filling and intent detection by exploiting bi-directional connections between tasks.
• Propose an SF-ID network to directly connect slot filling and intent detection.
• Introduce an iteration mechanism to strengthen inter-task interactions.
• Evaluate on ATIS and Snips to show improvements in slot filling, intent detection, and sentence-level accuracy.

Proposed method

• SF-ID network with two subnets: a slot filling (SF) subnet and an intent detection (ID) subnet.
• An attention-based integration of context for both slots and intents to form c_slot and c_inte vectors.
• Slot reinforcement vector r_slot derived from a correlation factor f that links c_slot and c_inte.
• Intent reinforcement vector r_inte generated from slot information via the ID subnet.
• An iteration mechanism where r_inte and r_slot are refined through repeated SF and ID subnet interactions.
• Option to operate in SF-First or ID-First modes and optional CRF layer on SF outputs.

Experimental results

Research questions

• RQ1 Can a bi-directional interrelated mechanism between slot filling and intent detection improve SLU performance?
• RQ2 Does an explicit SF-ID network with an iterative interaction outperform prior joint models on standard SLU benchmarks?
• RQ3 What is the impact of SF-First vs ID-First ordering and the CRF layer on overall performance?

Key findings

• SF-ID network improves sentence-level semantic frame accuracy by 3.79% on ATIS and 5.42% on Snips compared to the state-of-the-art model.
• Both SF and ID subnets contribute to performance; interaction between subnets yields better results than separate or non-interacting configurations.
• ID-First mode improves slot filling performance, while SF-First mode enhances intent detection performance.
• Iteration mechanism with three rounds yields peak sentence accuracy on both ATIS and Snips in SF-First mode.
• CRF layer generally helps, especially for slot filling, by enabling better sequence decoding.