Language, Interaction and Computation Laboratory (CLIC)
Looking for datasets? Click here!
What we do
The CLIC is an interdisciplinary group of researchers interested in studying interactions between natural and artificial intelligence, with a strong focus on language. Our current research spans over the following areas:
Theoretical linguistics and its relation to human cognition (the LiCo group): we study the role of language in various cognitive abilities, developing theoretical and computational models of the structure of human language, of how it is learned and represented in the brain, and which of its properties may be due to biological constraints. We address these questions using interdisciplinary methods and tools that include corpora research, models with neural networks and neuroimaging techniques.
Computational models of multimodality (the LaVi group): we aim at understanding multimodal communication, in which intelligent agents can converse using information received through text, images or sounds. Our research tries to understand the role of these different modalities in learning certain reasoning skills and the language-reasoning interplay. Both the linguistic / cognitive aspects and the possible technological applications of this type of model are considered.
Human-aware Machine Intelligence (the HAMI group): we study how to ensure artificial agents are aligned with their users and can fruitfully collaborate with them. This involves developing AIs that can learn from data, reason according to symbolic knowledge, interactively request and provide information, understand the instructions they are given, and explain their reasoning to human stakeholders. Our research builds on recent ideas and techniques from fields such as explainabile AI, neuro-symbolic AI, representation learning, and causality.
DEep Neural ArtificiaL Intelligence (the DENALI group): The text describes the research laboratory's focus on computer vision, particularly in the areas of language and vision, zero-shot classification, temporal action localization, and visual instruction tuning. The laboratory aims to develop advanced visual intelligence systems by bridging the gap between natural language processing and computer vision, recognizing and categorizing unfamiliar objects, accurately detecting and locating actions in video streams, and developing intelligent systems that can learn from human instructions and perform complex tasks accurately.
Meaning and Computation Lab (the MCL group) Broadly speaking, we study computational processes involved in human language with particular attention to semantics. Much of our current work has focused on the meaning of quantification in natural language, logical reasoning, and explaining cross-linguistic similarities between languages. In our research, we have paid particular attention to how various complexity measures can help us understand the difficulty of core cognitive abilities, such as language learning, comprehension, or reasoning. Our approach is characterized by a mixture of formal (logic, computational modeling, simulations) and empirical (neurobehavioral experiments, corpus linguistics) methods. It draws inspiration from methods and concepts in theoretical computer science, e.g., formal language theory, computational complexity theory, logic, machine learning, and information theory. The underlying goal is to understand human cognition better from various theoretical angles and contribute to the mathematically and computationally rigorous theories of language and cognition.
Code, Datasets, and Pretrained Models
Code and resources from our groups:
Resources from the COMPOSES project:
- LAMBADA
- SICK (Sentences Involving Compositional Knowledge)
- Distributional Memory (DM)
- "Don't count, predict" vectors
DISSECT toolkit for creating distributional semantics spaces:
Archive
;){kind=link}