In September 2022 I began brainstorming topics for my research project. As a problem solver, it was only natural that my research project would begin with a problem. The issue I faced was grounded in semantic search. When utilising online digital libraries for research during my MSc, I found it extremely difficult to retrieve relevant publications with reference to my search query, which in fact hindered the research task itself.
How can relevant terms and results be retrieved from a knowledge base, if the knowledge base itself is inaccurate?
At the time, LLMs such as OpenAI GPT and Google’s BERT, BARD, and Gemini, were yet to be released into the public domain as chatbots. Many semantic search systems employed simple keyword-based retrieval techniques to match terms, exact or fuzzy, from databases – the age-old ‘highlighted keyword on a search results page’. Some solutions had even started to employ LLMs and investigate the introduction of knowledge bases to better structure semantics. However, the deeper I delved, the more I found it was not necessarily the retrieval task itself which led to poor results, but the foundational task of building the knowledge bases using extraction tasks. Another age-old adage – ‘garbage in, garbage out’. How can relevant terms and results be retrieved from a knowledge base, if the knowledge base itself is inaccurate?
This finding led to my research topic: Joint Relation Extraction in Scientific Literature with Transformer Architecture. In layman’s terms: the task of extracting complex compound nouns from scientific text, (‘three- dimensional objects’ and ‘affine-invariant image patches’) and the sequential task of deriving the relationships between those nouns (‘affine-invariant image patches’ coded as a ‘feature-of’ ‘three-dimensional objects’) using Google’s then Transformer-based LLM, BERT, trained specifically on scientific language for language extraction tasks. The output of the first task feeds directly into the second.
At the time there was a great deal of research from the scientific community grounded in the assumption that LLMs were unable to derive global context and shared knowledge, due to their architectural approaches. LLMs merely generate a representation of words, and parts of words (tokenisation), and their similarity. So, how on earth can they be expected to derive shared knowledge, such as who’s next at the bar in a queue of people of different genders who aren’t stood directly in a line.
However, this research found that in fact it was the error propagation between the two tasks which had the biggest impact on the accuracy of the total output. If the first task scores 90% accuracy, the input to the second task is the output of the first task, and the second task scores 90% accuracy, then the overall accuracy is in fact 81%. This issue is called chained error propagation.
Upon completion of my Research Project and MSc Computer Science I received a distinction, which is one of my proudest achievements to date. I was fortunate enough to work alongside my supervisor, Majid Latifi, to publish the work in MDPI journal.
The question I find myself asking is: We’re all building agents, but are we really tracking the accuracy of their outputs?
Throughout the period of 18 months whilst I worked alongside my supervisor on this Research Project, interest in AI exploded. In November 2022 OpenAI released GPT 3.5 alongside the ChatGPT interface, which has since become embedded in modern society. However, what’s most interesting about the topic is what has since followed. This research was novel in its conclusion that chained error propagation from Entity Extraction to Relation Extraction is responsible for low F1 scores (accuracy) for this task. Yet, fast forward to current times, LLMs are employed in agentic contexts. Orchestrator agents manage task-specific sub-agents, feeding outputs as inputs into other task-specific sub-agents. The question I find myself asking is: We’re all building agents, but are we really tracking the accuracy of their outputs?
Leave a Reply