Natural Language Processing and Generation for Customer Service Applications
The Context
Customer-fronting companies harvest voluminous amounts of data from their interactions and transactions with customers. A significant portion of these data are in the form of natural language conversations, and originate from various sources, such as email conversations, call-centre voice records and transcripts, and discussions on message boards or forums.
The information contained within such data is valuable for several business applications, such as chatbots for automating the process of replying to customers or for suggesting potential answer templates to customer service agents.
The Challenge
However, processing textual data and automatically generating good quality answers or templates poses several challenges, such as detecting the topic being dealt with or the intent of the question. These challenges are further compounded in domain-specific settings, such as the insurance sector, characterized by a specialized terminology, and in languages other than English like French for instance, for which existing natural language processing (NLP) resources are scarce.
Key Question & Goals
Based on the aforementioned elements, the key question here can be formulated as follows:
“How to find the best compromise between the size of a large language model and its performance for deployment in a low-resource environment?”
The main goals of this research project will be to:
Investigate which large language model is best suited to our tasks, for example Llama 2 or Flacon.
Investigate methods for reducing the size of large language models: LoRa QLora Pruning Knowledge distillation Reducing dimensions
Investigate the performance metrics of the model, including the carbon footprint.
Investigate hallucination mitigation methods.
About Partenamut
1.2 million customers make Partenamut the largest health insurer in Belgium. It is a socially engaged non-profit organization with local roots. Indeed, it is a non-profit and reinvests its financial gains in its team members, its service offering and society. Currently Partenamut supports over 20 start-ups’ active within healthcare. It simplifies healthcare by advising its members at every step in their health care and by offering quality services which meet their needs.
