Inducing criteria for lexicalization parts of speech using the Cyc KB

Bjørn Aldag

Abstract

We present an approach for learning part-of-speech distinctions by induction over the lexicon of the Cyc knowledge base. This produces good results (74.6%) using a decision tree that incorporates both semantic features and syntactic features. Accurate results (90.5%) are achieved for the special case of deciding whether lexical mappings should use count noun or mass noun headwords. Comparable results are also obtained using OpenCyc, the publicly available version of Cyc.

1 Introduction

In semantic lexicons, a lexical mapping describes the relationship between a concept and phrases used to refer to it. This mapping includes syntactic information for deriving variant phrases, including the part of speech of the headword. Selecting the part of speech for the lexical mapping is required so that proper inflectional variations can be recognized and generated for the term. Although often a straightforward 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 from decisions that have been made by lexical knowledge engineers in lexicalizing 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 lexicon 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 denotational assertion maps a phrase, specified via a lexical concept with optional string modifiers, into a concept, usually a collection. The part of speech is specified by Cyc’s SpeechPart constants. The simplest type of denotational mapping uses the denotation predicate. For example, (denotation Device-Word CountNoun 0 PhysicalDevice) indicates that sense 0 of the count noun ‘device’ refers to PhysicalDevice (via the wordforms “device” and “devices”). Three additional predicates account for phrasal mappings: compoundString, head-MedialString, and multiWordString are used for phrases with the headword at the beginning, the middle, and the end, respectively. These denotational assertions, excluding lexical mappings 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 mappings. For each target denotatum term, the corresponding defnitional information (e.g., isa), asserted or inferable via transitivity, is extracted from the ontology. For simplicity, these definitional types are referred to as ancestor terms. Associations between the lexicalization parts of speech and common ancestor terms underlie the lexicalization speech part classifier and its special case, the mass-count classifier (distinguishing, 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 lexeme 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 denotatum term and converting this into a vector of occurrence indicators, one indicator per reference term. Then the headword 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 classification over Cyc lexical mappings, using Cyc reference terms and headword suffixes as features. Instances is size of the training data. Classes indicates number of choices. Baseline selects most frequent case. A simple fragment from the resulting decision tree shows how
ontological features interact with morphological ones:

if (isa AbstractInformationalThing) then
if (suffix = “-er”) then
if (not isa SomethingExisting) then AgentiveNoun
if (isa SomethingExisting) then MassNoun
if (suffix = “-ed”) then MassNoun
if (suffix = “-al”) 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 J48
algorithm in theWeka 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 system 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 performs well. This suggests that sufficient data is already available 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 classifier
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 relatively little written on automatically determining default lexicalization 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 lexicalizations 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 incorporate semantic information, in particular Cyc’s ontologicaltypes, in addition to syntactic information (e.g., headword morphology). Although the main approach incorporates Cyc’s conceptual distinctions, it can be extended to non-Cyc applications via the WordNet mapping [O’Retard 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 members, 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 obtained through MII Grants EIA-9810732 and EIA-0220590.