Building Dialogue Understanding Models for Low-resource Language Indonesian from Scratch

TL;DR

A Bi-Confidence-Frequency Cross-Lingual transfer framework (BiCF) is proposed, which consists of “BiCF Mixing”, “Latent Space Refinement” and “Joint Decoder’, respectively, to overcome the lack of low-resource language dialogue data and performs reliably and cost-efficiently on different scales of manually annotated Indonesian data.

Abstract

Making use of off-the-shelf resources of resource-rich languages to transfer knowledge for low-resource languages raises much attention recently. The requirements of enabling the model to reach the reliable performance lack well guided, such as the scale of required annotated data or the effective framework. To investigate the first question, we empirically investigate the cost-effectiveness of several methods to train the intent classification and slot-filling models for Indonesia (ID) from scratch by utilizing the English data. Confronting the second challenge, we propose a Bi-Confidence-Frequency Cross-Lingual transfer framework (BiCF), composed by ``BiCF Mixing'', ``Latent Space Refinement'' and ``Joint Decoder'', respectively, to tackle the obstacle of lacking low-resource language dialogue data. Extensive experiments demonstrate our framework performs reliably and cost-efficiently on different scales of manually annotated Indonesian data. We release a large-scale fine-labeled dialogue dataset (ID-WOZ) and ID-BERT of Indonesian for further research.

Figures (8)

• Figure 1: The investigation of utilizing off-the-shelf resources and models for generating low-resource language dialogue understanding models from scratch.
• Figure 2: Illustration of the proposed framework (BiCF), which consists of BiCF Mixing, Latent Space Refinement, and Joint Decoder. The frequency-word and confidence-word set in the first stage are derived from English dataset and confidence-translated parallel sentences, respectively. By fusion and mixing, the mixed data is generated. The cross-lingual space refinement module will generate a target-specific embedding model to represent Indonesian better. The final stage is to decode and output intent and slots jointly.
• Figure 3: The comparison of different methods on four domains.
• Figure 4: The comparison of different methods on four domains.
• Figure 5: An example of editing template interface.