Stanford CS224N NLP with Deep Learning | 2023 | Lecture 11 - Natural Language Generation

Natural language generation (NLG) encompasses systems that produce fluent, coherent text, ranging from non-open-ended tasks like machine translation to highly open-ended creative story generation. Effective generation relies on autoregressive models trained via maximum likelihood, though this approach often suffers from exposure bias and repetition. Decoding strategies, including top-k and top-p sampling, mitigate these issues by introducing stochasticity and preventing the selection of degenerate sequences. Training advancements, particularly reinforcement learning from human feedback (RLHF), align models with human preferences, while evaluation remains a significant challenge. Traditional lexical overlap metrics like BLEU often fail to capture semantic quality, necessitating more robust, model-based metrics and rigorous human evaluation. Ethical considerations, including toxicity mitigation and factual accuracy, remain critical as large-scale models become increasingly integrated into real-world applications.