In conversation with AI: building better language models

A new paper by researchers from the University of Adelaide, Auckland University of Technology (AUT), and University of Cambridge presents a system for teaching language modeling (LM) to humans, called Gopher. The Goopher algorithm uses machine learning techniques to teach itself to answer questions posed using natural language prompts. The system can effectively handle tasks such as question answering, summarization, and entity recognition, with the potential to improve human performance in these areas. In Gopher, the researchers create a graph of knowledge graphs for natural language questions and entities, representing them through structured graphs that use knowledge-based information exchange (KBIE) techniques. The KBIE architecture is based on a graph neural network (GNN) model and supports multiple tasks. To evaluate Goopher’s performance in question answering, the researchers conducted experiments with four different datasets: a large collection of news articles, a corpus of textual entailment questions from Yago, a dataset for named entity recognition (NER), and a dataset for summarization. The results showed that Gopher can perform well on these tasks, outperforming human performance. In question answering, Gopher was able to answer questions with an average accuracy of 82% and 76% on news and corpus datasets, respectively. In NER, it achieved an average accuracy of 90%, while in summarization, it achieved a maximum accuracy of 96%. The Goopher system is also capable of learning to ask natural language questions, making it particularly useful for teaching LM to humans. The system’s ability to learn can potentially improve the accuracy of human question answering and NER, as well as other types of machine learning tasks. Overall, the Gopher system offers a new approach for teaching LM that allows for effective training with natural language prompts and questions. It provides a new way for humans to improve their skills in language modeling, which could help advance artificial intelligence research and improve human-machine cooperation.