Margot Flowers

EDISON: An engineering design invention system operating naively

The goal of the EDISON project is to design a program capable of creating novel mechani-cal devices, by using knowledge of naive physical relationships, qualitative reasoning, plan-ning, and discovery/invention heuristics applied to abstract devices organized and indexed in an episodic memory. The EDISON program operates in two modes: brainstorming mode and problem-solving mode. In problem-solving mode, a goal specification is given as input and EDISON attempts to achieve the goal through plan selection and sub-goal satisfaction. A goal specification can be to alter or improve a device. Devices are represented symboli-cally, and are reasoned upon by EDISON without performing numerical computations. In brainstorming mode, EDISON starts with a device recalled from memory, and attempts to create novel devices through processes of mutation, generalization and analogical reason-ing. The devices EDISON manipulates consist of simple, everyday mechanisms, such as mousetraps, nail clippers, can openers and doors. A goal of the EDISON project is to gain computational insight into the processes of naive physical reasoning [Hayes 78] and inven-tion [Lenat 76] which people exhibit. To do so, we must address a number of issues, includ-ing: (a) how devices are represented and manipulated without detailed mathematical rea-soning, (b) how devices are organized, indexed, and retrieved from a personal, episodic memory of devices and experiences of device use, (c) how new devices are discovered or in-vented during problem solving and/or brainstorming, and (d) how the resulting inventions are assessed for their novelty and/or ingenuity.

Argument representation for editorial text

This paper presents a theory of argument representation for computer comprehension of editorial text. The theory has been implemented in a prototype system called OpEd, which reads politico-economic editorials and answers questions about their contents. The theory characterizes five major classes of knowledge structures: beliefs, attack relationships, support relationships, argument units (AUs), and meta-argument units (meta-AUs). During editorial comprehension, OpEd represents explicitly the beliefs of the editorial writer and his/her implicit opponents. Beliefs involve predictions about goals, plans, events, and states. Three types of predications are distinguished: evaluations about plans, causal relationships, and beliefs about beliefs. Beliefs in an editorial are involved in support relationships or attack relationships. An attack is a relationship between two beliefs whose contents involve mutually-exclusive planning situations or opposite effects of a plan on interrelated goals. A support is a relationship that consists of a belief, the justification for the belief, and a warrant (itself a belief) that connects a belief to its justifications. Belief justifications are based on refinements of plan evaluations, refinements of plan-goal relationships, analogies, and examples. Beliefs, attack relationships and support relationships are organized by AUs, which encode language-free and domain-free knowledge about argument structure and content. With the aid of domain-specific knowledge, AUs can be instantiated to model arguments in which an arguer refutes his/her opponents position that a given plan should/should not be used. Arguments about arguments are represented as meta-AUs, which specify argument errors that result from either inconsistencies between actions and beliefs, or support strategies involving plausibilities, circularities, self-contradictions, or shifts on the burden of proof. Meta-AUs are represented in OpEd as attacks on warrants and are used to model discussions about the nature of valid reasoning.

Argument comprehension and retrieval for editorial text

OpEd is a prototype editorial comprehension and question-answering system in the domain of politico-economic protection. Editorial comprehension in OpEd involves the application of: domain-specific knowledge, abstract knowledge of argumentation, and strategies for mapping input editorial text into conceptual structures. OpEds model of domain-specific knowledge includes four major elements: authority triangles and social acts, to represent explicitly all the information associated with conflicts in international trade, including beliefs, goals, and conflict-resolution methods: goals and plans, to represent political and economic actions in terms of desired economic states; a trade graph, to represent causal relationships among the economic quantities associated with products and consumers; and reasoning scripts, to represent common chains of cause-effect relationships in editorials. OpEds central, organizing constructs of argument knowledge are termed argument units (AUs). AUs consist of configurations of support and attack relationships among beliefs, where the content of each belief refers to goal/plan situations. The knowledge of argumentation encoded by AUs is highly abstract and independent of any particular domain. As a result, argument comprehension in OpEd requires instantiating AUs with the aid of politico-economic knowledge. Associated with each knowledge construct in OpEd are one or more processing strategies. These strategies are invoked to recognize knowledge constructs that are not explicitly stated in the text, but which must be inferred to understand the argument, planning and causal structure of the text. The result of processing strategies is the construction of an argument graph, organized in terms of beliefs, belief relationships, and AUs. Question answering in OpEd requires retrieving appropriate beliefs, belief relationships, or AUs from the editorials argument graph. Intial entry to the argument graph is provided by indexing structures associated with argument participants, plans and goals. To answer belief-related questions, OpEd analyses the contents of each question into one of six conceptual question categories. Each conceptual question category leads to the selection of a specific search and retrieval strategy.