Is there a role for linguists in the future of AI?

In the 20th century, most research in AI involved human experts hand-crafting rules to drive “intelligent” decision-making systems. Within NLP, linguists provided grammar rules and lexicons to drive syntactic parsing. Other linguists worked out the rules which govern coherent conversation. The role of the computer scientists was to code programmes to apply these rules efficiently for the analysis or generation of human language.

From around 1980, small groups of researchers started experimenting with an alternative approach, involving machine learning. On the surface, these approaches did away with the need for linguists to hand-craft linguistic knowledge, rather the programmes themselves were tasked to analyse language and produce a functional model of language. This work has culminated in recent years with AI systems such as ChatGPT, which to a large degree reflect human-like intelligent behaviour.

This talk will argue that through the development of LLMs, linguists have played an important role. At one level, linguists have been employed to prepare training data given to LLMS, so that the LLM can replicate human linguistic analysis. At a deeper level, linguists have defined the overall architecture of the LLMs, specifying the various components which work together to allow the overall system to function, such as analytical, generative, conversational and planning components.

The talk will stress that the AI is not a finished product, but will be continuously re-architectured to improve performance, and linguists are needed for this to be done effectively. In particular, AIs currently lack a deep understanding of how human activity is organised in terms of cultures and institutions, and we linguists need to work out how AIs can be trained to function as culturally aware beings.

Mick O'Donnell is a lecturer in English Studies at the Universidad Autónoma de Madrid. He has been working in various areas of NLP since the 1980s, in text generation, parsing, and dialogue management. He is best known for the annotation tools he has developed, including Systemic Coder (1992), RSTTool (1997), and UAM Corpustool (2008). His interests currently concern applying NLP tools to explore second language acquisition.

Join via Zoom.