Data-to-Text: A Constrained Approach to Diverse Candidate Text Generation and Ranking

Data-to-text (D2T) generation is a crucial NLP task, where transformer-based language models excel at understanding complex relationships and contextual dependencies in structured input data. However, the decoding process in text generation poses a challenge, as the generated text quality is sensitive to the decoding method, often trading off between diversity and accuracy. We propose a novel multi-step approach that leverages constrained randomness through top-p and top-k sampling to generate multiple candidate outputs. These candidates are then evaluated using learning-based models and ranked to select the top-ranked output. Our extensive experiments on benchmark datasets demonstrate the effectiveness of this approach in producing diverse and accurate text, with significant implications for advancing human-like dialogue systems.