Speech translation has been traditionally tackled under a cascade approach, chaining speech recognition and machine translation components to translate from an audio source in a given language into text or speech in a target language. Leveraging on deep learning approaches to natural language processing, recent studies have explored the potential of direct end-to-end neural modelling to perform the speech translation task. Though several benefits may come from end-to-end modelling, such as a reduction in latency and error propagation, the comparative merits of each approach still deserve detailed evaluations and analyses. In this work, we compared state-of-the-art cascade and direct approaches on the under-resourced Basque–Spanish language pair, which features challenging phenomena such as marked differences in morphology and word order. This case study thus complements other studies in the field, which mostly revolve around the English language. We describe and analysed in detail the mintzai-ST corpus, prepared from the sessions of the Basque Parliament, and evaluated the strengths and limitations of cascade and direct speech translation models trained on this corpus, with variants exploiting additional data as well. Our results indicated that, despite significant progress with end-to-end models, which may outperform alternatives in some cases in terms of automated metrics, a cascade approach proved optimal overall in our experiments and manual evaluations.

Comparable corpora can benefit the development of Neural Machine Translation models, in particular for under-resourced languages. We present a case study centred on the exploitation of a large comparable corpus for Basque-Spanish, created from independently-produced news by the Basque public broadcastereitb, where we evaluate the impact of different techniques to exploit the original data, in order to complement parallel datasets for this language pair in both translation directions. Two efficient methods for parallel sentence mining are explored, which identified a common core of approximately half of the total number of aligned sentences, each one uniquely identifying valid parallel sentences not captured by the other method. Filtering the data via identification of length-difference outliers proved highly effective to improve the models, as was the use of tags to discriminate between comparable and parallel data in the training corpora. The use of backtranslated data is also evaluated in this work, with results indicating that alignment-based datasets remain the most beneficial, although complementary backtranslations should also be included to fully exploit the available comparable data. Overall, the results in this work demonstrate that this type of data needs to be carefully analysed prior to its use as training data for Neural Machine Translation, since issues such as information imbalance between source and target data can lead to unoptimal results for a given translation pair.

Jesús Calleja, Thierry Etchegoyhen, David Ponce. Findings of the Association for Computational Linguistics: EMNLP 2024. 2024.

David Ponce, Thierry Etchegoyhen, Jesús Calleja, Harritxu Gete. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.