Neural conversation models with expert knowledge
This project will have an external supervisor from Kindly, in addition to an internal supervisor from LTG.
Neural conversation models are increasingly popular to develop various types of interactive systems such as chatbots. However, they are often based on large neural architectures with millions of parameters, and require therefore large amounts of training data. Getting hold of sufficient amounts of training data is not always easy given the time and effort involved in creating such datasets.
That said, even when we don't have much training data for our application, we might still have access to so-called expert knowledge (for instance, in the form of heuristic rules). Can we exploit such “expert knowledge” to build neural conversational models? There is indeed a growing body of work in NLP on how to inject prior knowledge into neural models using weak supervision methods. Weak supervision (Ratner et al. 2017) is a relatively recent AI paradigm that allows machine learning models to be learned from a combination of noisy supervision signals instead of large amounts of manually labelled data (see for instance this paper).
The master thesis will explore how these methods can be applied to the development of neural conversation models and evaluate their empirical performance with a few experiments with human users interacting with a chatbot (the exact chatbot domain remains to be determined), and will be carried out in close cooperation with Kindly.
Kindly is a language technology company based in Oslo. Kindly was founded in 2016 and has since grown to a team of 40+ employees, several of whom have graduated from the Language Technology Group (LTG) at UiO with master’s and doctoral degrees. Natural language processing and machine learning are at the core of the Kindly chatbot platform, which powers the conversational agents for some of the leading enterprises in the Nordics, such as Norwegian Air Shuttle, Elkjøp, Kahoot!, Thon Hotels, and Finn. This master’s project will be carried out in close cooperation with Kindly.