This paper describes the process of data processing and training of an automatic speech recognition (ASR) system for Cook Islands Māori (CIM), an Indigenous language spoken by approximately 22,000 people in the South Pacific. We transcribed four hours of speech from adults and elderly speakers of the language and prepared two experiments. First, we trained three ASR systems: one statistical, Kaldi; and two based on Deep Learning, DeepSpeech and XLSR-Wav2Vec2. Wav2Vec2 tied with Kaldi for lowest character error rate (CER=6±1) and was slightly behind in word error rate (WER=23±2 versus WER=18±2 for Kaldi). This provides evidence that Deep Learning ASR systems are reaching the performance of statistical methods on small datasets, and that they can work effectively with extremely low-resource Indigenous languages like CIM. In the second experiment we used Wav2Vec2 to train models with held-out speakers. While the performance decreased (CER=15±7, WER=46±16), the system still showed considerable learning. We intend to use ASR to accelerate the documentation of CIM, using newly transcribed texts to improve the ASR and also generate teaching and language revitalization materials. The trained model is available under a license based on the Kaitiakitanga License, which provides for non-commercial use while retaining control of the model by the Indigenous community.
Pretrained multilingual models are able to perform cross-lingual transfer in a zero-shot setting, even for languages unseen during pretraining. However, prior work evaluating performance on unseen languages has largely been limited to low-level, syntactic tasks, and it remains unclear if zero-shot learning of high-level, semantic tasks is possible for unseen languages. To explore this question, we present AmericasNLI, an extension of XNLI (Conneau et al., 2018) to 10 Indigenous languages of the Americas. We conduct experiments with XLM-R, testing multiple zero-shot and translation-based approaches. Additionally, we explore model adaptation via continued pretraining and provide an analysis of the dataset by considering hypothesis-only models. We find that XLM-R’s zero-shot performance is poor for all 10 languages, with an average performance of 38.48%. Continued pretraining offers improvements, with an average accuracy of 43.85%. Surprisingly, training on poorly translated data by far outperforms all other methods with an accuracy of 49.12%.
Word embeddings are critical for numerous NLP tasks but their evaluation in actual under-resourced settings needs further examination. This paper presents a case study in Bribri, a Chibchan language from Costa Rica. Four experiments were adapted from English: Word similarities, WordSim353 correlations, odd-one-out tasks and analogies. Here we discuss their adaptation to an under-resourced Indigenous language and we use them to measure semantic and morphological learning. We trained 96 word2vec models with different hyperparameter combinations. The best models for this under-resourced scenario were Skip-grams with an intermediate size (100 dimensions) and large window sizes (10). These had an average correlation of r=0.28 with WordSim353, a 76% accuracy in semantic odd-one-out and 70% accuracy in structural/morphological odd-one-out. The performance was lower for the analogies: The best models could find the appropriate semantic target amongst the first 25 results approximately 60% of the times, but could only find the morphological/structural target 11% of the times. Future research needs to further explore the patterns of morphological/structural learning, to examine the behavior of deep learning embeddings, and to establish a human baseline. This project seeks to improve Bribri NLP and ultimately help in its maintenance and revitalization.
Linguistic tone is transcribed for input into ASR systems in numerous ways. This paper shows a systematic test of several transcription styles, using as an example the Chibchan language Bribri, an extremely low-resource language from Costa Rica. The most successful models separate the tone from the vowel, so that the ASR algorithms learn tone patterns independently. These models showed improvements ranging from 4% to 25% in character error rate (CER), and between 3% and 23% in word error rate (WER). This is true for both traditional GMM/HMM and end-to-end CTC algorithms. This paper also presents the first attempt to train ASR models for Bribri. The best performing models had a CER of 33% and a WER of 50%. Despite the disadvantage of using hand-engineered representations, these models were trained on only 68 minutes of data, and therefore show the potential of ASR to generate further training materials and aid in the documentation and revitalization of the language.
This paper presents the results of the 2021 Shared Task on Open Machine Translation for Indigenous Languages of the Americas. The shared task featured two independent tracks, and participants submitted machine translation systems for up to 10 indigenous languages. Overall, 8 teams participated with a total of 214 submissions. We provided training sets consisting of data collected from various sources, as well as manually translated sentences for the development and test sets. An official baseline trained on this data was also provided. Team submissions featured a variety of architectures, including both statistical and neural models, and for the majority of languages, many teams were able to considerably improve over the baseline. The best performing systems achieved 12.97 ChrF higher than baseline, when averaged across languages.
This paper presents a neural machine translation model and dataset for the Chibchan language Bribri, with an average performance of BLEU 16.9±1.7. This was trained on an extremely small dataset (5923 Bribri-Spanish pairs), providing evidence for the applicability of NMT in extremely low-resource environments. We discuss the challenges entailed in managing training input from languages without standard orthographies, we provide evidence of successful learning of Bribri grammar, and also examine the translations of structures that are infrequent in major Indo-European languages, such as positional verbs, ergative markers, numerical classifiers and complex demonstrative systems. In addition to this, we perform an experiment of augmenting the dataset through iterative back-translation (Sennrich et al., 2016a; Hoang et al., 2018) by using Spanish sentences to create synthetic Bribri sentences. This improves the score by an average of 1.0 BLEU, but only when the new Spanish sentences belong to the same domain as the other Spanish examples. This contributes to the small but growing body of research on Chibchan NLP.
This paper presents three ongoing projects for NLP in Cook Islands Maori: Untrained Forced Alignment (approx. 9% error when detecting the center of words), speech-to-text (37% WER in the best trained models) and POS tagging (92% accuracy for the best performing model). Included as part of these projects are new resources filling in a gap in Australasian languages, including gold standard POS-tagged written corpora, transcribed speech corpora, time-aligned corpora down to the level of phonemes. These are part of efforts to accelerate the documentation of Cook Islands Maori and to increase its vitality amongst its users.