Intent Classification Using Pre-trained Language Agnostic Embeddings For Low Resource Languages

Abstract

Building Spoken Language Understanding (SLU) systems that do not rely on language specific Automatic Speech Recognition (ASR) is an important yet less explored problem in language processing. In this paper, we present a comparative study aimed at employing a pre-trained acoustic model to perform SLU in low resource scenarios. Specifically, we use three different embeddings extracted using Allosaurus, a pre-trained universal phone decoder: (1) Phone (2) Panphone, and (3) Allo embeddings. These embeddings are then used in identifying the spoken intent. We perform experiments across three different languages: English, Sinhala, and Tamil each with different data sizes to simulate high, medium, and low resource scenarios. Our system improves on the state-of-the-art (SOTA) intent classification accuracy by approximately 2.11% for Sinhala and 7.00% for Tamil and achieves competitive results on English. Furthermore, we present a quantitative analysis of how the performance scales with the number of training examples used per intent.

On Building Spoken Language Understanding Systems for Low Resourced Languages

“…This is why the pipeline that uses phonetic transcriptions outperforms Wav2Vec based embeddings. (Yadav et al, 2021) show that this is true even when Allosaurus embeddings are compared to phonetic transcriptions generated by Allosaurus. As the amount of available data decreases, intent classification systems built using phonetic transcriptions begin to outperform systems based on Allosaurus embeddings, thus showing that projecting input speech into phonetic transcriptions is the most exhaustive way to use the scarce amount of labelled data in the compounded low-resourced settings.…”

“…Allosaurus is trained to recognize and transcribe input speech into a series of phones contained in the utterance, providing superior representations of input audio which can also be used for languages linguistically distant from high resourced languages like English. (Yadav et al, 2021) show that using embeddings generated from Allosaurus to encode speech content outperforms previous state-of-the-art methods for Sinhala and Tamil by large margins, while maintaining high performance on high resourced languages like English (99.08% classification accuracy for a 31-class intent classification problem). But the performance drops as the dataset size decreases and is not optimal for the task-specific low resourced settings that we are dealing with in this paper.…”

“…A series of recent works (Gupta et al, 2020bYadav et al, 2021) replace the ASR module in the SLU pipeline by a universal phone recognition system called Allosaurus (Li et al, 2020). Allosaurus is a universal phonetic transcription system that creates language and speaker independent representations of input speech.…”

“…One way to tackle this setting is to use an ASR system built for a higher resourced language and use the transcriptions generated to perform downstream tasks as used in (Buddhika et al, 2018;Karunanayake et al, 2019b,a). It was later shown in Yadav et al, 2021) that using language and speaker independent systems trained on many languages to ex-tract speech features works much better than using ASR systems built for a different language, as a different language usually contains a different set of phonemes with a different phone to phoneme set mapping. When this setting is compounded by task-specific low-resourced-ness, we are at an extremely low resourced setting where each data point becomes valuable.…”