Jurassic is (almost) All You Need: Few-Shot Meaning-to-Text Generation for Open-Domain Dialogue
This work addresses the problem of improving response quality and coverage for open-domain dialogue systems like Alexa Prize's Athena, offering a novel approach with potential real-world deployment, though it is incremental in applying few-shot learning to this specific context.
The paper tackled the challenge of generating truthful, high-quality responses in open-domain dialogue systems by experimenting with few-shot prompt-based learning using GPT-Neo and Jurassic-1 models across multiple domains. Results showed that with 10-shot prompting, Athena-Jurassic significantly outperformed in coherence and semantic accuracy, achieving only 4% untrue hallucinations in video games, while Athena-GPT-Neo dropped to 0.41 semantic accuracy and 12% hallucinations in 2-shot cross-domain settings.
One challenge with open-domain dialogue systems is the need to produce truthful, high-quality responses on any topic. We aim to improve the quality and coverage of Athena, an Alexa Prize dialogue system. We experiment with few-shot prompt-based learning, comparing GPT-Neo to Jurassic-1, for the movies, music, TV, sports, and video game domains, both within and cross-domain, with different prompt set sizes (2, 3, 10), formats, and meaning representations consisting of either sets of WikiData KG triples, or dialogue acts. Our evaluation uses BLEURT and human metrics, and shows that with 10-shot prompting, Athena-Jurassic's performance is significantly better for coherence and semantic accuracy. Experiments with 2-shot cross-domain prompts results in a huge performance drop for Athena-GPT-Neo, whose semantic accuracy falls to 0.41, and whose untrue hallucination rate increases to 12%. Experiments with dialogue acts for video games show that with 10-shot prompting, both models learn to control dialogue acts, but Athena-Jurassic has significantly higher coherence, and only 4% untrue hallucinations. Our results suggest that Athena-Jurassic produces high enough quality outputs to be useful in live systems with real users. To our knowledge, these are the first results demonstrating that few-shot semantic prompt-based learning can create NLGs that generalize to new domains, and produce high-quality, semantically-controlled, conversational responses directly from meaning representations.