Research

Dissertation: Enhancing Translation Systems for Low-resourced Scenarios

Machine Translation (MT)

Robustness

Neural Machine Translation (NMT) systems often struggle when faced with noisy inputs, especially if trained only on clean data. In real-world scenarios, handling noise is crucial—this includes code-switching, transliteration, ungrammatical sentences, misspellings, and dialectal variations. Our research aims to make NMT systems robust to these common challenges, particularly for low-resource languages.

Terminology Integration

As NMT becomes a key tool in professional translation pipelines, it is important for models to respect domain-specific terminologies. We explore methods to incorporate terminological constraints into NMT systems, ensuring that translations remain accurate and consistent with specialized vocabulary.

Evaluation

Evaluating MT quality remains challenging. Existing metrics often require large human-rated datasets or fail to generalize across languages and domains. Our work investigates methods for building reliable, automatic evaluation metrics that can assess translation quality even in low-resource settings.

Low-Resource Focus

Most MT research has historically focused on a small set of well-resourced languages, leaving millions of speakers underserved. Our goal is to shift the focus toward traditionally ignored languages, creating translation systems that serve diverse linguistic communities.

Dictionaries as Resources

Dictionaries are often the easiest or quickest resource available for low-resource languages. We investigate how external resources, combined with limited parallel data, can be leveraged to build effective MT models for these languages.

Speech Translation (ST)

Data Filtering

Much of the data available online is noisy or misaligned, especially for low-resource languages. Poor language identification and alignment issues exacerbate this problem. We focus on developing filtering techniques to identify high-quality data and improve the reliability of mined datasets.

Data Augmentation

Low-resource languages often lack sufficient parallel speech-text datasets. To address this, we explore methods for generating synthetic parallel data, such as leveraging Text-to-Speech technologies, to augment existing datasets and build more robust Speech Translation models.