Alessandro Moschitti, Retrieval Augmented Generative Question Answering for Personal Assistants

• Alessandro Moschitti
• April 3, 2024, at 4 PM
• Online (Zoom)
  

Dr. Alessandro Moschitti
(Amazon, USA)

Alessandro Moschitti is a Principal Research Scientist of Amazon Alexa, where he has been leading the science of Alexa information service since 2018. He designed the Alexa Question Answering (QA) system based on unstructured text and more recently the first Generative QA system, which extends the answering skills of Alexa.

He obtained his Ph.D. in CS from the University of Rome in 2003, and then did his postdoc at The University of Texas at Dallas for two years. He was professor of the CS Dept. of the University of Trento, Italy, from 2007 to 2021. He participated to the Jeopardy! Grand Challenge with the IBM Watson Research center (2009 to 2011), and collaborated with them until 2015. He was a Principal Scientist of the Qatar Computing Research Institute (QCRI) for five years (2013-2018). His expertise concerns theoretical and applied machine learning in the areas of NLP, IR and Data Mining.

He is well-known for his work on structural kernels and neural networks for syntactic/semantic inference over text, documented by around 350 scientific articles. He has received four IBM Faculty Awards, one Google Faculty Award, and five best paper awards. He was the General Chair of EACL 2023 and EMNLP 2014, a PC co-Chair of CoNLL 2015, and has had a chair role in more than 70 conferences and workshops. He is currently a senior action/associate editor of ACM Computing Survey and JAIR. He has led ~30 research projects, e.g., a 5-year research project with MIT CSAIL.

Abstract

Recent work has shown that Large Language Models (LLMs) can potentially answer any question with high accuracy, also providing justifications of the provided output. At the same time, other research work has shown that even the most powerful and accurate models, such as ChatGPT 4, generate hallucinations, which often invalidated their answers. Retrieval-augmented LLMs are currently a practical solution that can effectively solve the above-mentioned problem.

However, the quality of grounding is essential in order to improve the model, since noisy context deteriorates the overall performance. In this talk, after we introduce LLMs and Question Answering (QA), and we will present our experience with Generative QA, which uses basic search engines and accurate passage rerankers to augment relatively small language models. We will provide a different but more direct interpretation of retrieval augmented LLMs and contextual grounding. Finally, we will show our latest techniques for Reinforcement Learning from Human Feedback proposed for fine-tuning LLMs that we developed in contemporary with the main stream effort of OpenAI.