LifeNet: A Propositional Model of Ordinary Human Activity

We describe LifeNet, a new common sense knowledge base that captures a first-person model of human experience in terms of a propositional representation. LifeNet represents knowledge as an undirected graphical model relating 80,000 egocentric propositions with 415,000 temporal and atemporal links between these propositions. We explain how we built LifeNet by extracting its propositions and links from the Open Mind Common Sense corpus of common sense assertions, present a method for reasoning with the resulting knowledge base, evaluate the knowledge in LifeNet and the quality of inference, and describe a knowledge acquisition system that lets people interact with LifeNet to extend it further. INTRODUCTION We are interested in building ‘common sense’ models of the structure and flow of human life. Today’s computer systems lack such models—they know almost nothing about the kinds of activities people engage in, the actions we are capable of and their likely effects, the kinds of places we spend our time and the things that are found there, the types of events we enjoy and types we loathe, and so forth. By finding ways to give computers the ability to represent and reason about ordinary life, we believe they can be made more helpful participants in the human world. An adequate common sense model should include knowledge about a wide range of objects, states, events, and situations. For example, a common sense model of human life should enable the following kinds of predictions: • When someone is thirsty, it is likely that they will soon be drinking a liquid beverage. • When someone is at an airport, it is likely they possess a plane ticket. • When someone is typing at a computer, it is possible that they are composing an e-mail. • When someone is crying, it is likely that they feel sad or are in pain. • After someone wakes up, they are likely to get out of bed. Most previous efforts to encode common sense knowledge have made use of relational representations such as frames or predicate logics. However, while such representations have proven expressive enough to describe a wide range of common sense knowledge (see Davis [1] for many examples of how types of common sense knowledge can be formulated in first-order logic, or the Cyc upper level ontology [2]), it has been challenging finding methods of default reasoning that can both make use of such powerful representations and also scale to the number of assertions that are needed to encompass a reasonably broad range of common sense knowledge. In addition, as a knowledge base grows, it is increasingly likely that individual pieces of knowledge will suffer from bugs of various kinds; it seems necessary that we find methods of common sense reasoning that are tolerant to some errors and uncertainties in the knowledge base. However, in recent years there has been much progress in finding ways to reason in uncertain domains using less expressive propositional representations, for example, with Bayesian networks and other types of graphical models. Could such methods be applied to the common sense reasoning problem? Is it possible to take an approach to common sense reasoning that begins not with an ontology of predicates and individuals, but rather with a large set of propositions linked by their conditional or joint probabilities? Propositional representations are less expressive than relational ones, and so it may take a great many propositional rules to express the same constraint as a single relational rule, but such costs in expressivity often come with potential gains in tractability, and in the case of common sense domains, this trade-off seems to be rather poorly understood. The potential benefits of a proposition representation go beyond just matters of efficiency. From the perspective of knowledge acquisition, interfaces for browsing and entering propositional knowledge are potentially much easier to use because they do not require that the user learn to read and write some complex syntax. From the perspective of applying common sense reasoning within applications, propositional representations have such a simple semantics that they are likely quite easy to interface to. Thus, while propositional representations may be less expressive and require a larger ontology of propositions than relational representations for the same domain, they are in many ways easier to build, understand and use. In this paper we explore such questions by describing LifeNet, a new common sense knowledge base that captures a first-person model of human experience in terms of Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. K-CAP’03, October 23–25, 2003, Sanibel Island, Florida, USA. Copyright 2003 ACM 1-58113-583-1/03/0010...$5.00. a propositional representation. LifeNet represents knowledge as a graphical model relating 80,000 egocentric propositions with 415,000 temporal and atemporal links between these propositions, e.g. • I-put-my-foot-on-the-brake-pedal → I-stop-a-car • I-pour-detergent-into-wash → I-clean-clothes • I-put-quarter-in-washing-machine → I-clean-clothes • I-am-at-a-zoo → I-see-a-monkey • I-put-on-a-seat-belt → I-drive-a-car • I-put-a-key-in-the-ignition → I-drive-a-car We explain how we built LifeNet by extracting its propositions and links from the Open Mind Common Sense corpus of common sense assertions supplied by thousands of members of the general public, present a method for reasoning with the resulting knowledge base, evaluate the knowledge in LifeNet and the quality of inference, and describe a knowledge acquisition system that lets people interact with LifeNet to extend it further. LIFENET LifeNet is a large-scale temporal graphical model expressed in terms of ‘egocentric’ propositions, e.g. propositions of the form: • I-am-at-a-restaurant • I-am-eating-a-sandwich • It-is-3-pm • It-is-raining-outside • I-feel-frightened • I-am-drinking-coffee Each of these propositions is a statement that a person could say was true or not true of their situation, perhaps with some probability. In LifeNet these propositions are arranged into two columns representing the state at two consecutive moments in time, and these propositions are linked by joint probability tables representing both the probability that one proposition follows another, and also the probability of two propositions being true at the same time. A small sample of LifeNet is shown in Figure 1 below: