Source-linked AI summary
A Novel Bi-directional Interrelated Model for Joint Intent Detection and Slot Filling
Haihong E, Peiqing Niu, Zhongfu Chen, Meina Song
TL;DR
Existing SLU joint models do not establish bi-directional interrelated connections between intent detection and slot filling. The paper proposes an SF-ID network with iterative information exchange between the two tasks, and reports higher sentence-level semantic frame accuracy on ATIS and Snips than the state-of-the-art model.
Problem
Existing joint SLU models do not establish bi-directional direct connections that let intent detection and slot filling use each other’s information.
Method
The paper proposes an SF-ID network with SF and ID subnets that exchange information through reinforce vectors and an iterative mechanism.
Results
The model outperforms baselines on ATIS and Snips and improves sentence-level semantic frame accuracy relative to the state-of-the-art model.
Takeaways & Limitations
Bi-directional interrelated connections between slots and intent provide guidance for future SLU work.
Abstract
from arXiv · showhide
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.
1 Introduction
SLU extracts sentence intent and word-level semantic constituents through intent detection and slot filling. Existing pipeline systems separate these tasks, while joint models seek to address error propagation and establish stronger task interactions.
- SLU Tasks: SLU identifies an utterance’s intent and assigns slot labels to its words.These tasks capture sentence-level and word-level semantics, respectively.
- Prior Approaches: Traditional approaches treat intent detection as classification and slot filling as sequence labeling.Examples include SVM and RNN methods for intent detection, and CRF and LSTM methods for slot filling.
- Motivation: Pipeline error propagation motivates joint models for intent detection and slot filling.Joint modeling is presented as the emerging direction for addressing the limitations of separate pipelines.
- Research Gap: Existing joint models lack fully bi-directional direct connections between intent and slot information.Earlier approaches either link the tasks implicitly or pass information in only one direction.
- Contributions: The proposed SF-ID network connects an SF subnet and an ID subnet, with an iteration mechanism designed to strengthen their interaction.The contributions include experiments on two benchmark datasets demonstrating the model’s effectiveness and superiority.
2 Proposed Approaches
The model integrates contextual information with attention and connects slot filling and intent detection through an SF-ID network. Its two subnetworks exchange reinforce vectors iteratively, with customizable execution order and CRF-based sequence decoding.
- 2.1 Integration of Context: Attention derives slot and intent context vectors from BLSTM hidden states to capture dependencies across utterances.Slot context is computed for each slot, while intent context produces one sentence-level intent representation.
- 2.2 SF-ID Network: The SF-ID network contains SF and ID subnets whose order can be configured as SF-First or ID-First.The earlier subnet sends an active effect to the later subnet through a medium vector.
- 2.2.1 SF-First Mode: In SF-First mode, the SF subnet combines intent and slot context to produce a slot reinforce vector for the ID subnet.This transfers slot information into intent detection through the inter-subnet connection.
- 2.2.1 SF-First Mode: The iteration mechanism feeds the intent reinforce vector back into the SF subnet, repeatedly updating both subnetworks’ reinforce vectors.The interaction can repeat, and the vectors serve as changing links between the two subnets.
- Final Prediction: After iteration, reinforce vectors participate in final predictions for intent and slot labels alongside BLSTM hidden states.The slot-filling subnet can additionally use a CRF layer to jointly decode the best label chain.
- 2.2.2 ID-First Mode: ID-First mode performs intent detection before slot filling and changes the first-iteration ID-subnet calculation.Subsequent iteration details remain nearly the same apart from subnet order.
3 Experiment
Experiments on ATIS and Snips evaluate task-specific and overall performance, subnet interactions, model modes, iteration count, and the CRF layer. The SF-ID network outperforms baselines, while three iterations maximize sentence accuracy in SF-First mode.
- Experimental Setup: The experiments use ATIS and Snips, with slot filling evaluated by F1-score, intent detection by accuracy, and joint performance by sentence accuracy.Sentence accuracy counts sentences whose slots and intent are both correctly predicted.
- Model Performance: Our model outperforms baselines in slot filling, intent detection, and sentence accuracy on ATIS and Snips.The reported relative sentence-accuracy improvements are around 3.79% on ATIS and 5.42% on Snips.
- Subnet Analysis: Both the SF-only and ID-only subnet models outperform the BLSTM model, indicating that each subnet contributes to performance improvement.The analysis separately evaluates the effects of the two subnets.
- Interaction and Modes: The interacting SF-ID network performs better than independent SF and ID subnets, while ID-First favors slot filling and SF-First favors intent detection.The difference between the two modes is described as minor overall.
- Iteration Mechanism: Sentence accuracy in SF-First mode increases with iteration count and reaches its maximum at three iterations on both ATIS and Snips.After three iterations, sentence accuracy stabilizes with a minor drop.
- CRF Layer: The CRF layer improves general model performance, with a larger improvement for slot filling than for intent detection.The paper attributes this difference to CRF’s focus on sequence labeling.
4 Conclusion
The paper concludes that its SF-ID network establishes bi-directional interrelated connections between intent detection and slot filling, strengthened by an iteration mechanism. The model outperforms baselines on two public datasets, supporting further SLU work using this mechanism.
- Conclusion: The proposed SF-ID network provides bi-directional interrelated connections between intent detection and slot filling.The two tasks are intended to promote each other mutually.
- Conclusion: An iteration mechanism enhances the interrelated connections between intent and slots.
- Conclusion: The bi-directional interrelated model outperforms baselines on two public datasets.