Interactive Natural Language Explanations of Cyc Inferences

This paper describes the inference explanation capabilities of Cyc, a logical reasoning system that includes a huge “commonsense” knowledge base and an inference engine that supports both question answering and hypothesis generation. Cyc allows the user to compose queries by means of English templates, and tries to find answers via deductive reasoning. If deduction is fruitless Cyc resorts to abduction, filling in missing pieces of logical arguments with plausible conjectures to obtain provisional answers. Cyc presents its answers and chains of reasoning to the user in English, provides drilldown to external source references whenever possible, and reasons about its own proofs to determine optimal ways of presenting them to the user. When a chain of reasoning relies on conjectures introduced via abduction, the user can interact with the inference explanation to confirm or deny the abduced supports. These capabilities are grounded in the integration of Cyc’s natural language components with the knowledge base and inference engine, and in Cyc’s capacity to maintain an explicit in-memory record of the facts, rules, and calculations used to produce successful proofs during inference. Introduction: The Cyc System The user of a rule-based system will find, ideally, that the system responds to queries by returning informative and possibly surprising answers. Depending on the task at hand, the user might want to see how the system has reasoned to obtain its results. Indeed, the more important the task and informative or surprising the results, the more we might expect the user to want to assess the soundness of the reasoning that produced the results. The utility of a rule-based system for high-impact, moneyor life-saving, mixed initiative applications, therefore, is often proportionate to the transparency of its explanations. The challenges to be surmounted in order to clearly present complex proofs grow as the sheer number of relevant facts and rules in the knowledge base (KB) increases, and as the variety of inference patterns used to generate proofs multiplies. These challenges are especially extreme for Cyc, a knowledge-based reasoning system that includes a huge KB, an inference engine, Cyc’s Semantic Knowledge Source Integration (SKSI) facility, and several natural lanCopyright © 2005, American Association for Artificial Intelligence (www.aaai.org). All rights reserved. guage (NL) components. The knowledge stored in the Cyc KB is represented in a formal language, CycL, which subsumes and extends the predicate calculus of first-order logic. The KB contains 2.7M assertions (facts and rules) that interrelate 300K concepts, encoding both “commonsense” knowledge (consensus reality) and specialized domain knowledge. Cyc’s millions of assertions are distributed over thousands of explicitly represented and semantically significant logical contexts, or microtheories. Cyc’s inference engine can produce a single inference result (answer set) by combining information from a myriad of contexts, and by employing an armory of defeasible reasoning procedures, both sound (resolution, class subsumption, argumentation) and unsound (abduction). The inference engine consists of a “harness” that invokes any of 800 pattern-specific reasoning modules, and falls back on general resolution-based theorem proving as a last resort. Each module implements an efficient inference procedure for a common type of problem, such as a technique for calculating the transitive closure of a binary predicate. One significant suite of modules implements abductive inference, and thereby enables Cyc to introduce new conjectures (hypothetical statements) during inference. The SKSI facility allows Cyc to access the structured contents of external data sources of information, such as relational databases and web pages. Once the structure and meaning of an external knowledge source have been described by assertions in the KB, the inference engine can automatically create specialized inference modules for the source and can treat it, thenceforth, as a virtual extension of the KB. An explicit in-memory record of completed proofs underpins Cyc’s capacity to explain its own reasoning. Cyc can reason about the features of it own proofs when the user asks to view them, determining which statements are the most salient, which statements are too trivial to show, and which statements are most appropriate for the different levels of detail supported in the display. Cyc’s NL components include modules for parsing English to CycL, template-based generation of English from CycL, and discourse management, along with a generalpurpose lexicon and several smaller, domain-specific lexi1 For more information about the development of SKSI, see [Masters and Güngördü 2003]. For more information about Cyc, past and present, see [Lenat and Guha 1990], [Lenat 1995], and the whitepapers and other material available on Cycorp’s web site: http://www.cyc.com . Figure 1: Cyc’s Query Library (QL) interface displaying a query and answers. cons. Nearly all of Cyc’s grammatical and lexical knowledge is represented in CycL and stored in the KB, thus making it available for use by the inference engine. Cyc supports several text-based and graphical user interfaces. Most notable for the focus of this paper is the Query Library (QL) interface, which allows the user to construct queries, submit them to the inference engine, view the resulting answers (Figure 1), and obtain a detailed display of the proofs that support each answer (Figures 2 and 3). The QL employs Cyc’s natural language generation capabilities to display queries, answers, and proofs (also referred to as justifications) in English, rather than in CycL. An Inference Example To illustrate the complexity of a typical Cyc justification, here we consider an example of interest to intelligence analysts that draws on Cyc’s knowledge of recent events in the Middle East. Let us suppose that the date is February 16, 2005, and that an analyst (user) has constructed the following English query with the QL interface: [EQ1] Who had a motive for the assassination of Rafik Hariri? The corresponding CycL version of this query, and the form in which it must be posed to the inference engine, is as follows: [Q1] (agentHasMotiveForAction ?WHO TerroristAttack-February-14-2005-Beirut) This query asks the inference engine to find terms that can be substituted for the variable ?WHO to yield a closed logical sentence. When the analyst poses the query [Q1], the inference engine tries to find answers by using heuristic search, drawing on data and rules stored in the KB, and possibly in other sources accessible through SKSI. After a few seconds, the inference engine obtains the following substitution terms (CycL values) for ?WHO, each of which constitutes an answer to [Q1]: UnitedStatesOfAmerica, Syria, Israel, and AlQaida. The rest of this section describes part of the chain of reasoning that produces one of the answers, Syria. First, the inference engine selects rule [R1], since the predicate in the consequent of [R1] matches the predicate in [Q1]: