Inducing criteria for lexicalization parts of speech using the Cyc KB

1 Introduction

In semantic lexicons, a lexical mapping describes the re­ lationship between a concept and phrases used to refer to it. This mapping includes syntactic information for deriv­ ing variant phrases, including the part of spee.ch of the ?ea~­ word. Selecting the part of speech for the lexical mappmg IS required so that proper inflectional variations can be re.cog­ nized and generated for the term. Although often a straight­ forward task, there are special cases that can pose problems, especially when fine-grained speech part categories are used.

To reduce the need for linguistic expertise in producing these lexicalizations for a large knowledge base like Cyc [Lenat, 1995], linguistic criteria can be inferred fr~m de~i­ sions that have been made by lexical knowledge engmeers m /exicalizing preexisting terms.

The Cyc knowledge base, containing 120,000 concepts and over one million axioms, 1 divides roughly into three layers. The upper ontology formalizes fundamental distinctions (e.g., tangibility versus intangibility). The lower ontology collects specific facts, often related to concrete applications, and the middle ontology encodes commonsense knowledge about the world. The KB also includes a broad-coverage English lexi­ con mapping words and phrases to terms throughout the KB.

Natural language lexicons are integrated directly into the Cyc KB [Burns and Davis, 1999]. Binary predicates, as in (nameString HEBCompany "HEB"), map names to terms. A

1 These figures and the results discussed later are based on Cyc KB v576 and system vl.2577. See www.cyc.com for detailed doc­ umentation on the KB and [O'Hara eta!., 2003] for more technical details related to this work.

denotational assertion maps a phrase, specified via a lexical concept with optional string modifiers, into a concept, usu­ ally a collection. The part of speech is specified by Cyc's SpeechPart constants. The simplest type of denotational map­ ping uses the denotation predicate. For example, (denotation Device-Word CountNoun 0 Physica!Device) indicates that sense 0 of the count noun 'device' refers to Physica!Device (via the wordforms "device" and "devices"). Three additional predi­ cates account for phrasal mappings: compoundString, head­ Media/String, and multiWordString are used for phrases with the headword at the beginning, the middle, and the end, respectively. . . .

These denotational assertions, excludmg lex1cal mappmgs for technical , informal and slang terms, form the training data for a lexicalization speech part classifier.

2 Inference of lexicalization part of speech

Our method of inferring the part of speech for lexicalizations is to apply decision tree learning over existing lexical map­ pings. For each target denotaturn term, the correspondmg def­ initional information (e.g., isa), asserted or inferable via tran­ sitivity, is extracted from the ontology. For simplicity, th~se definitional types are referred to as ancestor terms. Associa­ tions between the lexicalization parts of speech and common ancestor terms underlie the lexicalization speech part clas­ sifier and its special case, the mass-count classifier ( distin­ guishing, e.g., "much sand" from "many books"). To reduce the size of the training feature vector, only the most frequent 256 atomic terms from the thousands of possible ancestor terms are selected, after excluding certain internal bookkeeping constants. .

Given a training instance, such as a denotatiOn from a lex­ erne into a specific Cyc concept using a particular SpeechPart (e.g., MassNoun or a CountNoun), the feature specification is derived by determining all the ancestor terms of the de­ notatum term and converting this into a vector of occurrence indicators, one indicator per reference term. Then the head­ word is checked for the occurrence of a set of commonly used suffixes. If found, the suffix itself is added to the vector (in a position set aside for suffixes of the same length). The part of speech serves as the classification variable.

We use decision trees for this classification. Part of the motivation is that the result is readily interpretable and can be incorporated directly by knowledge-based applications. A

simple fragment from the resulting decision tree shows how ontological features interact with morphological ones:

if (isa Abstractlnformationa!Thing) then

if (suffix = "-er") then

if (not isa SomethingExisting) then AgentiveNoun

if (isa SomethingExisting) then Mass Noun

if (suffix= "-ed") then Mass Noun

if (suffix= "-a!") then Adjective

if (suffix= "-or") then AgentiveNoun

Table 1 shows the results of 10-fold cross validation for the mass-count classification. This was produced using the 148 algorithm in the Weka machine learning package [Witten and Frank, 1999], which is an implementation of Quilan 's C4.5 [Quinlan, 1993] decision tree learner. This shows that the sys­ tem achieves an accuracy of 90.5%, an improvement of 22.3 percentage points over a baseline of always selecting the most frequent case. The OpenCyc version of the classifier also per­ forms well. This suggests that sufficient data is already avail­ able in OpenCyc (available online at www.opencyc.org) to allow for good approximations for such classifications.

The mass/count noun distinction can be viewed as a special case of speech part classification. Running the same classi­ fier using the full set of speech part classes yields the results shown in Table 2. Here the overall result is not as high, but there is a similar improvement over the baseline.

3 Discussion

Contextual part of speech tagging [Brill, 1995] has received substantial attention in the literature, but there has been rela­ tively little written on automatically determining default lex­ icalization parts of speech. Woods [Woods, 2000] describes an approach to this problem using manually-constructed rules incorporating syntactic, morphological, and semantic tests

(via an ontology). Our rules are induced from the knowledge base, which alleviates the need for rule maintenance as well as rule construction.

This paper shows that an accurate decision procedure (90.5%) for determining the mass-count distinction in lexi­ calizations can be induced from the lexical mappings in the Cyc KB. The performance (74.6%) in the general case is also promising, given that it is a much harder task with over 30 part of speech categories to choose from. The features in­ corporate semantic information, in particular Cyc's ontolog­ ical types, in addition to syntactic information (e.g., head­ word morphology). Although the main approach incorporates Cyc's conceptual distinctions, it can be extended to non-Cyc applications via the WordNet mapping [O'Hara et al., 2003].

This work is just a small initial step in applying machine learning techniques to the massive amount of data in Cyc. The recent release of OpenCyc enables wider investigation and exploitation of the information in the Cyc knowledge base for intelligent applications.

Acknowledgements

The lexicon work at Cycorp has been supported by many staff mem­ bers, and, in part by grants from NIST, DARPA, and ARDA. The first author is currently supported by a GAANN fellowship from the Department of Education. The work utilized NMSU resources ob­ tained through Mil Grants EIA-981 0732 and EIA-0220590.