David Peng

Simeng Han, Hailey Schoelkopf, Yilun Zhao et al. · salesforce

Large language models (LLMs) have achieved remarkable performance on a variety of natural language understanding tasks. However, existing benchmarks are inadequate in measuring the complex logical reasoning capabilities of a model. We present FOLIO, a human-annotated, logically complex and diverse dataset for reasoning in natural language (NL), equipped with first-order logic (FOL) annotations. FOLIO consists of 1,430 examples (unique conclusions), each paired with one of 487 sets of premises used to deductively reason for the validity of each conclusion. The logical correctness of the premises and conclusions is ensured by their FOL annotations, which are automatically verified by an FOL inference engine. In addition to the main NL reasoning task, NL-FOL pairs in FOLIO constitute a new NL-FOL translation dataset. Our experiments on FOLIO systematically evaluate the FOL reasoning ability of supervised fine-tuning on medium-sized language models. For both NL reasoning and NL-FOL translation, we benchmark multiple state-of-the-art language models. Our results show that a subset of FOLIO presents a challenge for one of the most capable {Large Language Model (LLM)} publicly available, GPT-4.

Roxana Girju, David Peng

In this study, we explore how language captures the meaning of words, in particular meaning related to sensory experiences learned from statistical distributions across texts. We focus on the most frequent perception verbs of English analyzed from an and Agentive vs. Experiential distinction across the five basic sensory modalities: Visual (to look vs. to see), Auditory (to listen vs. to hear), Tactile (to touch vs. to feel), Olfactory (to smell), and Gustatory (to taste). In this study we report on a data-driven approach based on distributional-semantic word embeddings and clustering models to identify and uncover the descriptor sensory spaces of the perception verbs. In the analysis, we identified differences and similarities of the generated descriptors based on qualitative and quantitative differences of the perceptual experience they denote. For instance, our results show that while the perceptual spaces of the experiential verbs like to see, to hear show a more detached, logical way of knowing and learning, their agentive counterparts (to look, listen) provide a more intentional as well as more intimate and intuitive way of discovering and interacting with the world around us. We believe that such an approach has a high potential to expand our understanding and the applicability of such sensory spaces to different fields of social and cultural analysis. Research on the semantic organization of sensory spaces for various applications might benefit from an the Agentive/Experiential account to address the complexity of multiple senses wired with each other in still unexplored ways.