Core Concepts
A pipeline leveraging Large Language Models (LLMs) for machine translation of slot-annotated spoken language understanding (SLU) training data can effectively extend SLU systems to new languages, outperforming existing state-of-the-art methods.
Abstract
The paper introduces a pipeline that utilizes Large Language Models (LLMs) to extend spoken language understanding (SLU) systems to new languages. The key aspects of the approach are:
The pipeline starts with human-labeled SLU data in English and translates it to multiple target languages (German, Spanish, French, Hindi, Japanese, Portuguese, Turkish, and Chinese) using an LLM-based machine translation model.
The core challenge is the Slot Transfer Task, which involves accurately annotating named entities during the translation process. The authors leverage the EasyProject approach, which uses HTML-like tags to mark named entities, enabling the LLM-based translation model to effectively handle slot transfer.
The translated datasets are then used to train SLU models, which are evaluated on the original MultiATIS++ test sets in the respective languages. This testing phase assesses the quality and fidelity of the translated datasets.
The authors also train a compact, on-device SLU model from scratch (Not-Pretrained Transformer) using the translated datasets, achieving a significant improvement of over 17% relative on the MultiATIS++ dataset compared to the baseline method.
In the cloud-based scenario, the authors' approach outperforms the current state-of-the-art methods, including Fine and Coarse-grained Multi-Task Learning Framework (FC-MTLF) and Global-Local Contrastive Learning Framework (GL-CLEF), on the MultiATIS++ benchmark.
The proposed methodology demonstrates the effectiveness of LLM-based machine translation in addressing the challenges of cross-lingual SLU, providing a scalable and slot-agnostic solution that can be easily deployed in various production scenarios.
Stats
The overall accuracy of the on-device SLU model (Not-Pretrained Transformer + LLM Slot Translator) improved from 5.31% to 22.06% compared to the baseline BiLSTM + GL-CLEF method.
In the cloud scenario, the overall accuracy of the mBERT-based SLU model improved from 53% to 62.18% compared to the state-of-the-art FC-MTLF method.
Quotes
"Our LLM-based MT method represents a significant advancement in overcoming the obstacles faced by conventional MT approaches in the context of cross-lingual SLUs, e.g., [4]."
"Contrary to both FC-MTLF and GL-CLEF, our LLM-based machine translation does not require changes in the production architecture of SLU. Additionally, our pipeline is slot-type independent: it does not require any slot definitions or examples."