Learning Prototypical Functions for Physical Artifacts

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 6941–6951 August 1–6, 2021. ©2021 Association for Computational Linguistics 6941 Learning Prototypical Functions for Physical Artifacts Tianyu Jiang and Ellen Riloff School of Computing University of Utah Salt Lake City, UT 84112 {tianyu, riloff}@cs.utah.edu Abstract Humans create things for a reason. Ancient people created spears for hunting, knives for cutting meat, pots for preparing food, etc. The prototypical function of a physical artifact is a kind of commonsense knowledge that we rely on to understand natural language. For example, if someone says “She borrowed the book” then you would assume that she intends to read the book, or if someone asks “Can I use your knife?” then you would assume that they need to cut something. In this pa- per, we introduce a new NLP task of learning the prototypical uses for human-made physical objects. We use frames from FrameNet to rep- resent a set of common functions for objects, and describe a manually annotated data set of physical objects labeled with their prototypi- cal function. We also present experimental re- sults for this task, including BERT-based mod- els that use language model predictions from masked patterns as well as artifact sense def- initions from WordNet and frame definitions from FrameNet. 1 Introduction Humans are a creative species. New objects are invented by people every day, and most are cre- ated for a reason. Knives were created for cutting, bicycles were created for transportation, and tele- phones were created for communication. Some objects can perform multiple functions (e.g., smart phones) and humans are also creative at finding secondary uses for objects (e.g., heavy objects are often used as makeshift paperweights). But when we mention physical objects in conversation or in writing, people generally infer that the object will be used in the most prototypical way, unless they are told otherwise. The prototypical function of human-made physi- cal artifacts is a kind of commonsense knowledge that often plays a role in natural language under- standing. Consider the following examples of in- ferences that arise from physical artifacts. Example 1 a) He killed the mayor with a gun. b) He killed the mayor with a knife. c) He killed the mayor with a bomb. Example 1 describes a killing with three differ- ent types of instruments. Most readers would as- sume that a) describes a shooting, b) describes a stabbing, and c) describes an explosion. But ex- actly how each instrument was used is implicit. We make different inferences about how they were used based on our knowledge of the objects. Example 2 a) She finished the cigarette. b) She finished the puzzle. c) She finished the movie. Example 2 illustrates how we infer different ac- tions based on the object when the main action is elided (i.e., “finished” means that some action has ended but the action itself is implicit). Most people would assume that the cigarette was …

Content truncated to 500 words