Health coaching (HC) aims to promote sustainable behavior change through goal-oriented dialogue, but research in this area is limited by the scarcity of authentic, transcript-based corpora. Existing datasets are small, English-only, and Western-centric, overlooking cultural and linguistic factors that shape real-world HC interactions. We introduce CoachLah, the first Singlish–English parallel corpus of HC conversations collected from a randomized controlled trial in Singapore. The dataset comprises 36,852 utterances transcribed from almost 160 hours of recorded HC sessions with 51 clients and 4 professional health coaches. Each dialogue is speaker-labeled, transcribed in Singlish, and aligned with high-quality English translations to preserve linguistic and cultural nuances. All sessions include HC summaries written by health coaches after each HC session, from which behavioral goals were manually annotated. To demonstrate the dataset’s utility, we benchmark two downstream tasks: (i) Singlish-to-English translation using fine-tuned open-weight models (e.g., Gemma-2-9B-it) with Low-Rank Adaptation, and (ii) behavioral goal extraction from unstructured HC summaries using span-based modeling (e.g., DeBERTa-v3-base). Together, these contributions establish the first culturally grounded benchmark for low-resource, goal-oriented dialogue research in HC. Both the code and the dataset are available at: https://github.com/IvaBojic/CoachLah.

Singlish, a creole rooted in English and influenced by Singapore’s multilingual and multicultural environment, poses significant challenges for those proficient in standard English due to its unique and often complex lexical and syntactic structures. Despite significant advancements in language translation for both high- and low-resource languages, translating Singlish to English remains largely underexplored. This gap is primarily due to the lack of dedicated datasets for language detection and Singlish-to-English translation, as well as the absence of robust models capable of addressing the unique linguistic challenges posed by Singlish. In this work, we curate a word-level language detection dataset, a Singlish-to-English translation dataset, and propose a Language Detection-driven Masked Language Modelling approach for translating Singlish into English. We evaluate the performance of existing models and the proposed approach on two Singlish-to-English translation datasets, including our proposed SEAT dataset. The results demonstrate that the proposed LD-MLMTrans approach outperforms the baseline model and exhibits high proficiency in Singlish-to-English translation.

Automatic extraction of behavioral goals from health coaching notes is essential for scalable monitoring of coaching programs, yet training data is scarce and exhibits substantial domain shift across programs. We collect and annotate 157 notes from a coaching program and show that models trained on the only existing public corpus, SMARTSpan (173 notes), suffer a drop of up to 30 points in exact-match F1 when transferred to our data. To address this, we propose a factorized synthetic data generation pipeline that decomposes note variation into three largely independent axes, health coach documentation structure, patient goal content, and patient persona, extracts empirical priors from a small in-domain seed set, and samples from them to produce diverse synthetic notes with embedded goal-span labels validated via cycle-consistency filtering. In low-resource experiments with only 57 in-domain training notes, our approach outperforms rephrasing and backtranslation baselines on both exact-match and partial-match F1. Ablation analysis demonstrates that augmentation must target the in-domain distribution to be effective, and a human evaluation confirms that synthetic notes are structurally faithful, with detection driven by surface artifacts rather than content or organizational flaws.All code and generated data will be published at GitHub repository: https://github.com/Michael-Tanzer/cl4health-factorized-augmentation.