PhD : Approche frugale en Sémantique Formelle pour la compréhension des conversation

about the role

Your role is to conduct a PhD thesis on: "Approach in Formal Semantics for Conversation Understanding."

Global Context and Problematic of the Subject
In NLP, there is a representation of the meaning of sentences in the form of a graph (Abstract Meaning Representation, AMR) that allows for the handling of various application tasks using small language models (< 1B, such as FlanT5 with between 256M and 1.2B parameters). For dialogue, this representation is not yet utilized, except for robot commands.

Scientific Objective – Results and Challenges to Overcome
By illustrating the connection with the global context and the problematic, the thesis aims to directly describe what it addresses: "The objective of the thesis is to..."
The objective of the thesis is to adapt generic approaches of formal semantic analysis to conversation understanding. In addition to not requiring specific training for each application, their advantage is their low computational resource requirements.

The main challenges to overcome (scientific or technical) are:

• Understanding the sequence of exchanges (dialogues)
• Managing turn-taking
• Resolving coreferences

Among the recommended approaches to overcome these challenges are:

• Using formal semantics in the form of directed graphs
• Machine Learning / Deep Learning
• Manipulating semantic graphs

Finally, the main expected outcomes (besides publications and manuscript writing) are:

• Trained language models and tools for dialogue
• Evaluation corpus
• Evaluation