In contrast to classical cognitive science which studied brains in isolation, ecological approaches focused on the role of the body and environment in shaping cognition. Similarly, in this thesis we adopt an ecological approach to grounded natural language understanding (NLU) research. Grounded language understanding studies language understanding systems situated in the context of events, actions and precepts in naturalistic/simulated virtual environments. Where classic research tends to focus on designing new models and optimization methods while treating environments as given, we explore the potential of environment design for improving data collection and model development. We developed novel training and annotation approaches for procedural text understanding based on text-based game environments. We also drew upon embodied cognitive linguistics literature to propose a roadmap for grounded NLP research, and to inform the development of a new benchmark for measuring the progress of LLMs on challenging commonsense reasoning tasks. We leveraged the richer supervision provided by text-based game environments to develop Breakpoint Transformers, a novel approach to modeling intermediate semantic information in long narrative or procedural texts. Finally, we integrated theories on the role of environments in collective human intelligence to propose a design for AI-augmented "social thinking environments" for knowledge workers like scientists.
The paper discusses advancements in Natural Language Understanding (NLU) through the development of systems grounded in structured training environments, enabling a deeper understanding of language.
Ecological semantics is introduced as a method to enhance grounded language learning to broader domains by considering environments as crucial to semantic representations.
It suggests that by understanding and manipulating environments via language, models can achieve a more general, grounded understanding and use of natural language.
Despite challenges such as the need for extensive hard-coding and the integration of complex representations, ecological semantics provides a promising pathway towards more human-like language understanding and interaction.
In the rapidly evolving field of Natural Language Understanding (NLU), there's an ongoing quest for models that not only process but deeply understand language in a way that mirrors human comprehension. Recent trends have seen a significant leap in the capabilities of LLMs, like BERT and GPT-3, to grasp and generate human-like text across a variety of tasks. However, despite their impressive performances, these models often lack a deeper understanding of the texts they process, struggling with tasks that require more nuanced comprehension, generalization, and grounding in real-world knowledge.
A promising direction to bridge this gap is the development of systems that are not only informed by distributional semantics (the statistical relationships between words) but are also grounded in more structured, richer training environments. This approach, known as situated or grounded language learning, has shown potential in narrower domains but is often limited by the scale and pre-defined nature of its environments.
A paper, “Ecological Semantics: Programming Environments for Situated Language Understanding,” proposes an innovative approach to extend grounded language learning to broader domains by enabling language models not just to act within environments but also to understand and manipulate these environments through language. Inspired by theories from contemporary cognitive science, this proposal suggests treating environments as essential components in semantic representations.
The key insight here is the recognition of the environment's critical role in cognitive processes - an aspect grounded in the concept of "affordances" (what actions are possible in a given situation). By harnessing the language of affordances, we can construct "mental worlds" that specify possible actions, thus facilitating a deeper understanding that extends beyond statistical patterns.
The proposed ecological semantic framework offers a theoretical and practical pathway towards implementing systems capable of more general and grounded understanding. Importantly, this framework emphasizes the role of models as participants in creating and configuring environments, moving beyond their existing role as mere actors. This shift could enable models to understand and use modal language, facilitating the dynamic construction of relevant representations as needed.
One practical demonstration of this approach is the use of Interactive Fiction (IF) programming languages to create dynamic environments that can support a variety of NLU tasks. These environments allow for the cost-effective simulation of complex situations and the dynamic generation of actionable knowledge at scale, addressing common challenges in grounding and common-sense reasoning.
While this ecological approach opens new doors for advancing NLU, it also presents significant challenges. Creating and managing rich, actionable external knowledge requires extensive hard-coding and harnesses AI in the coding of its common-sense. Furthermore, symbolic knowledge, while valuable, may need to be supplemented with more complex geometric and multi-modal representations to capture the full breadth of human common-sense.
Despite these challenges, the ecological semantics approach offers a promising direction for research and development in NLU. By grounding models in richly structured and programmable environments, we can move closer to achieving systems that not only process language with human-like proficiency but also deeply understand and interact with the world in a meaningful way.
In conclusion, the future of NLU lies not just in improving the models but equally in innovating the environments they learn from. Ecological semantics provides a blueprint for such innovation, paving the way for richer, more grounded, and ultimately more human-like language understanding.