Brian Davis Assistant Professor in Computing and SFI Funded Investigator at the ADAPT Centre. Prior to joining DCU in 2019, he was a Lecturer at the Department of Computer Science, Maynooth University. From June 2014- August 2017, he was a Research Fellow, Adjunct Lecturer and Research Unit Leader at the INSIGHT Center for Data Analytics, NUI Galway (NUIG), where he led the Knowledge Discovery Unit. In addition, he was Principal Investigator of two SFI co-funded Targeted Projects (Elsevier and DataLive, respectively), Furthermore he was Coordinator of a 3 year Horizon 2020 Innovation Action - SSIX - Social Sentiment Financial Indexes (Grant No 645425). His research is oriented towards bridging the gap between computational linguistics/language processing research and the implementation of practical applications with potential real-world use. Examples include i) pipelined neural architectures for building Data2Text Natural Language Generation (NLG) systems ii) bias detection and removal with the context of resume text analysis in the hiring and recruitment domains iii) and language processing applied to online safety and child protection from short noisy texts in collaboration with the DCU Anti-Bullying Centre (ABC) and Iv) and recently Irish Language Technology.

Description:
Large Language Models (LLMs) have demonstrated impressive results in zero and/or few shot settings for many NLP tasks and recent advances with respect to Retrieval Augmented Generation(RAG) offer improved accuracy. The drawbacks of LLMs surrounding bias, accuracy/faithfulness and explainability have been discussed in the literature. Moreover their performance gains vs computational/environmental cost should give us cause to reflect and what NLP approach to take for the requirements at-hand when building practical real world applications. I will draw on example scenarios from ongoing applied and industry research NLP projects in my research group such as bias detection in a HR context, comparing pipeline neural NLP architectures with end to end systems. Other examples include low resource scenarios - such as the challenges involved in working with language data derived from vulnerable populations such as children for building NLP systems for online safety. Throughout my talk, I will advocate a pragmatic approach for building NLP systems, reflecting on the cost benefit of a particular solution, noting that certain solutions may be better than others in certain circumstances. I will reflect on when and how to discern this and most importantly what key skills and knowledge are required.