Towards Globally Inclusive Multilingual Dialogue Systems for Real-World Applications

With the advent of large language models (LLMs), dialogue systems have become the primary interface for accessing advances in natural language processing (NLP); yet existing research remains largely English-centric, text-based, and benchmark-driven, limiting both global inclusivity and real-world applications. To move beyond this narrow focus, the thesis broadens the scope of multilingual dialogue system research through the creation of new datasets and evaluation methods. It introduces Multi3WOZ, a large-scale, multi-parallel dataset for task-oriented dialogue in Arabic, English, French, and Turkish. It also presents HEALTHDIAL, the first large-scale, speech-first dataset for health communication, spanning diverse language varieties across Arabic, Chinese, English, and Spanish. HEALTHDIAL modernises task-oriented dialogue system design by replacing traditional parsing-based approaches with a retrieval-augmented generation pipeline that more effectively leverages the capabilities of LLMs. In addition, the thesis proposes the first framework for the quantitative measurement of cross-lingual disparities, capturing both those arising during system development and those intrinsic to LLMs. The conventional dataset--model--benchmarking pipeline has driven much of the progress in dialogue system research, but it remains insufficient for informing real-world applications. To bridge this gap, the thesis extends the pipeline in both directions. Upstream, it applies systematic review methodology in combination with global health frameworks to identify user needs, and map the state of NLP for public health in Africa. Downstream, it develops and releases open-source toolkits for multilingual data collection, system development, deployment, and human evaluation, thereby lowering barriers to real-world applications. Beyond its technical contributions, this thesis offers a methodological reflection on how NLP, and dialogue system research in particular, can move beyond benchmarks to generate evidence with real-world relevance. It distils three guiding principles for equitable NLP: research should be evidence-based, grounding decisions in systematic evidence; human-centric, ensuring that development and evaluation reflect the needs and values of the communities served; and context-adaptive, responding to the resources and constraints of diverse linguistic and cultural contexts. Together, these principles outline a framework for developing dialogue systems that are both globally inclusive and socially impactful.