Back AMORE Mini-workshop on Referential Information in Deep Learning Models
AMORE Mini-workshop on Referential Information in Deep Learning Models
When: Monday 11 April, 15:00h - 17:00h
Zoom link: https://upf-edu.zoom.us/j/92979825202
Schedule (see talk abstracts below):
Raquel Fernàndez - Efficient Language Production Strategies in Visually Grounded Dialogue
Speakers are thought to use efficient information transmission strategies for effective communication. For example, the Uniform Information Density principle states that speakers plan their utterances to reduce fluctuations in the density of the information transmitted. Previous work analysing this principle in dialogue has failed to take into account how the information content of utterances varies as a function of the available discourse context, and has completely ignored the role of extralinguistic information. In this talk, I will present work that tests whether and within which contextual units this principle holds in visually grounded task-oriented dialogues. We analyse production strategies in these dialogues combining information-theoretic measures with probability estimates and visuo-linguistic alignment scores obtained from transformer-based language models and multimodal models. Our findings show that efficient strategies are at play in dialogue when we zoom in on topically and referentially coherent contextual units. Beside providing empirical insights on human production strategies, our studies can inform the development of more human-like natural language generation models.
Anna Rogers - Challenges in Defining and Testing Machine Verbal Reasoning Skills
Natural language understanding is a quickly growing field, both in terms of modeling and data work. However, there is little agreement on what specific reasoning "skills" the models are supposed to learn, which is why many resource descriptions and error analysis sections rely on ad-hoc categories. This talk presents a taxonomy of verbal reasoning skills for future resource and model analysis work, based on the findings of a large-scale survey of current resources for reading comprehension and question answering. I will also discuss the key challenges in model evaluation and collecting data for training/testing specific "skills", as well as some proposed solutions.