Publications

Multilingual Neural Machine Translation (MNMT) models enhance translation quality for low-resource languages by exploiting cross-lingual similarities during training—a process known as knowledge transfer. This transfer is particularly effective between languages that share lexical or structural features, often enabled by a common orthography. However, languages with strong phonetic and lexical similarities but distinct writing systems experience limited benefits, as the absence of a shared orthography hinders knowledge transfer. To address this limitation, we propose an approach based on phonetic information that enhances token-level alignment across scripts by leveraging transliterations. We systematically evaluate several phonetic transcription techniques and strategies for incorporating phonetic information into NMT models. Our results show that using a shared encoder to process orthographic and phonetic inputs separately consistently yields the best performance for Khmer, Thai, and Lao in both directions with English, and that our custom Cognate-Aware Transliteration (CAT) method consistently improves translation quality over the baseline.

A majority of language technologies are tailored for a small number of high-resource languages, while relatively many low-resource languages are neglected. One such group, Creole languages, have long been marginalized in academic study, though their speakers could benefit from machine translation (MT). These languages are predominantly used in much of Latin America, Africa and the Caribbean. We present the largest cumulative dataset to date for Creole language MT, including 14.5M unique Creole sentences with parallel translations—11.6M of which we release publicly, and the largest bitexts gathered to date for 41 languages—the first ever for 21. In addition, we provide MT models supporting all 41 Creole languages in 172 translation directions. Given our diverse dataset, we produce a model for Creole language MT exposed to more genre diversity then ever before, which outperforms a genre-specific Creole MT model on its own benchmark for 23 of 34 translation directions.

Lingua is an application that can perform near-real-time interpretation of video recordings and live speeches as well as synchronized automatic video dubbing. It has been developed and is being piloted at the Church of Jesus Christ of Latter-day Saints. The system pipeline includes customized automatic speech recognition (ASR) and machine translation (MT) components. A script may be uploaded in advance to improve the translation accuracy and decrease the lag behind the speaker, while flexibly handling instances when the speaker goes off script. The speed of the text-to-speech (TTS) outputs are dynamically adjusted to match the rate of speech. Lingua is currently capable of interpreting English to 38 other languages.

Language tagging, a method whereby source and target inputs are prefixed with a unique language token, has become the de facto standard for conditioning Multilingual Neural Machine Translation (MNMT) models on specific language directions. This conditioning can manifest effective zero-shot translation abilities in MT models at scale for many languages. Expanding on previous work, we propose a novel method of language tagging for MNMT, injection, in which the embedded representation of a language token is concatenated to the input of every linear layer. We explore a variety of different tagging methods, with and without injection, showing that injection improves zero-shot translation performance with up to a 2+ BLEU score point gain for certain language directions in our dataset.

We propose an approach that improves the performance of VMT (Video-guided Machine Translation) models, which integrate text and video modalities. We experiment with the MAD (Movie Audio Descriptions) dataset, a new dataset which contains transcribed audio descriptions of movies. We find that the MAD dataset is more lexically rich than the VATEX dataset (the current VMT baseline), and we experiment with MAD pretraining to improve performance on the VATEX dataset. We experiment with two different video encoder architectures: a Conformer (Convolution-augmented Transformer) and a Transformer. Additionally, we conduct experiments by masking the source sentences to assess the degree to which the performance of both architectures improves due to pretraining on additional video data. Finally, we conduct an analysis of the transfer learning potential of a video dataset and compare it to pretraining on a text-only dataset. Our findings demonstrate that pretraining with a lexically rich dataset leads to significant improvements in model performance when models use both text and video modalities.

The morphological structure of Semitic languages, such as Arabic, is based on non-concatenative roots and templates. This complex word structure used by humans is obscured to neural models that employ traditional tokenization algorithms, such as byte-pair encoding (BPE) (Sennrich et al., 2016; Gage, 1994). In this work, we present and evaluate Semitic Root Encoding (SRE), a tokenization method that represents both concatenative and non-concatenative structures in Semitic words with sequences of root, template stem, and BPE tokens. We apply the method to neural machine translation (NMT) and find that SRE tokenization yields an average increase of 1.15 BLEU over the baseline. SRE tokenization is also robust against generating combinations of roots with template stems that do not occur in nature. Finally, we compare the performance of SRE to tokenization based on non-linguistic root and template structures and tokenization based on stems, providing evidence that NMT models are capable of leveraging tokens based on non-concatenative Semitic morphology.