Why search engines and chatbots are becoming more alike
Search engines are getting better at answering our questions. And chatbots are increasingly likely to search the internet for relevant sources. ‘Search engines and chatbots will become more closely entwined’, says Professor Suzan Verberne.
Type a question in Google and you’ll often immediately see an answer, even before clicking on a result. Search engines are becoming more interactive and more like chatbots, says Verberne. For their part, some chatbots are developing features of search engines. For example, if a chatbot is asked a question, it performs an internet search in the background and refers to sources in its answers.
Hallucinations and bias
It is ‘essential’ that search engine technology are added to chatbots, Verberne will argue in her inaugural lecture. A Professor of Natural Language Processing at the Leiden Institute of Advanced Computer Science (LIACS), her main focus is on finding information in large text collections. ‘We have all seen how powerful chatbots like ChatGPT are. But we are also getting a much clearer picture of the risks, such as hallucinations and bias.’ Hallucinations are when chatbots fabricate responses that sound true and bias refers to the biases that chatbots reflect and sometimes even magnify from the texts they are trained on.
Verberne believes that part of the solution to both problems lies in the implementation of search engine technology. ‘Chatbots should not only generate plausible answers but also retrieve correct information from the right sources.’ Verberne is working on methods where search engines retrieve relevant documents ‘at the back end’ of the language model. The language model then formulates an answer based on the search results.
Verberne and her PhD candidates and students work with open-source language models. ‘These are models that have been trained for general use that you can adapt to a particular system you are working on. You can adapt large models to a specific system but you need powerful computers to do so. We are lucky enough to have such computers at LIACS and through SURF, the national collaborative organisation for ICT.’ Task-oriented training helps the language models improve at specific tasks. This makes them a valuable tool for researchers.
Search engine for archaeologists
Verberne works with researchers from other disciplines on AI systems that can be used to research archaeological finds, legal texts or medical issues, for example. ‘Archaeologists sometimes have to find the relevant information about excavations among tens of thousands of long texts. Together with a PhD candidate from the Faculty of Archaeology, we have developed a search engine specifically for that purpose.’
The search engine uses a language model to recognise relevant search terms in archaeological reports. ‘To accomplish that, we got student assistants to indicate the relevant terms in some of these reports, for example for sites, certain artefacts or periods of time. We then trained the language model with this data. The search engine was then able to find relevant terms in the rest of the documents.’
Moments of frustration
Since the launch of ChatGPT, language models have become familiar to a wide audience and AI chatbots have received a lot of media attention. For Verberne it took ‘some getting used to’ to have her research field thrust into the limelight. ‘I did have some moments of frustration when reading the newspapers. It sometimes felt that journalists had consulted the wrong experts, and some stories were published without pause for thought.’ She can now see that journalists understand the technology better and are aware of the risks and dangers.
‘No one saw this hype coming’
Verberne had not expected a language model to receive so much attention. ‘No one could have anticipated that it would become such a global hype. ChatGPT was actually a minor innovation compared to the models that already existed. Interesting that its release caused such a furore.’
Developments are so rapid that it is difficult for researchers to keep up. ‘New language models are being released all the time. That’s why I think we should research aspects that are not specific to a certain model. How can we evaluate output properly? How can we resolve bias in training data? How can we make the models applicable to more languages and specific domains? These aspects remain relevant, regardless of which model you are working with.’
Text: Tom Janssen