The inaugural YNLG workshop served as a platform for community building and knowledge transfer among early-career researchers. This document synthesizes the core insights from the workshop and presents a comprehensive summary of advice from the Program Committee on navigating research methodology, publishing, and mental health. Furthermore, it highlights key takeaways from a panel of experts on mentorship, the evolving landscape of academia versus industry, and the future challenges of evaluation in NLG.

This document provides an overview of my research on data generation in low-resource settings. It also highlights several open questions and discussion points for the 1st Workshop for Young Researchers in Natural Language Generation.

Neural Language Models have become central to Natural Language Generation (NLG) research and can produce fluent and coherent text. However, when models generate text from complex, structured or long-form data such as tables or event logs, they often hallucinate and introduce factual errors. These hallucinations limit the practical deployment of large language models (LLMs) in applications where factual accuracy is critical. In my research, factual accuracy refers to the faithfulness of the generated text to the given input data. My PhD focuses on reducing hallucinations and improving factual accuracy in data-to-text generation, which I address through two core approaches: (i) analysing how input quality and structure improve factual accuracy, and (ii) developing a manual error annotation protocol and extending it into an LLM-as-Judge framework. This work aims to assess when automatic evaluation can complement human annotation and enable larger-scale evaluation.

My research focuses on improving textual inference in large language models (LLMs) for natural language generation, particularly in data-to-text generation. While LLMs are increasingly used to generate reports and insights from data, they often produce factually inaccurate or shallow outputs, limiting their usefulness. I work on integrating LLMs with symbolic operations through code generation for deeper and more faithful inferences. As generation tasks are often under-specified, both models and humans rely on implicit presuppositions, and mismatches can lead to errors or misinterpretation. I investigate how such presuppositions affect generation outputs and evaluation, how human presuppositions shape the perceived interestingness of the insights, and how they can be leveraged to improve insight generation.

Large Language Models (LLMs) excel in many NLP tasks but remain biased toward high-resource languages. This position paper discusses the author’s current findings on efficient strategies for low-resource settings: (i) modular training, where only non-embedding parameters are tuned after learning language-specific tokenizers and embeddings, and (ii) artificial language initialization, which leverages structurally biased synthetic languages for faster, parameter-efficient pretraining. The paper also shares plans for future research and topics that the author would like to discuss during the round-table.