The Knowledge Required to Interpret Noun Compounds

1 Introduction

Knowledge acquisition involves building knowledge bases (KBs) from the information provided by standard sources of expertise, such as people and texts. In addition to extract­ ing relevant information, knowledge acquisition involves re­ expressing the information in a formal language suitable for machines to reason with. In general, this is difficult because the information is initially expressed in natural languages, and these expressions are notoriously imprecise and incom­ plete. However, the goal of our project is to improve knowl­ edge acquisition methods by automating the translation of some kinds of expressions from natural languages to formal ones. We call this task loose-speak interpretation.

Several kinds of expressions are good candidates for loose­ speak interpretation by knowledge acquisition systems. For example, noun compounds omit information that can often be inferred, e.g. concrete floor is "a floor made of concrete", while gymnasium floor is "the floor region of a gymnasium". Another candidate is metonymic expressions. These expres­ sions contain incompatible terms and must be expanded to make meaningful phrases. For example, the statement Joe read Shakespeare, means "Joe read text written by Shake­ speare".

This paper focuses on the first kind of loose-speak: in­ terpreting noun compounds in the context of knowledge ac­ quisition. A noun compound is a sequence of nouns com­ posed of a head noun and one (or more) modifiers. The head noun determines the type of the whole compound (with few exceptions), and the modifiers specialize the type from the head noun. Although we limit our study to only pairs of nouns, our results can be applied to longer noun compounds by bracketing them into pairs of nouns (with few exceptions), and then interpreting each pair [Liberman and Sproat, 1992; Pustejovsky and Bergler, 1993; Barker, 1998].

The computational linguistics community has studied noun compound interpretation extensively [Leonard, 1984; Down­ ing, 1977; Levi, 1979; Finin, 1986; Fabre, 1996; Lauer and Dras, 1994; Barker, 1998; Vanderwende, 1994]. In these studies, the task is to select a single semantic category for each pair of nouns. The selection is usually made from a small list of semantic categories, for example category part­ of is selected for noun compounds like human lung, category material for marble statue, and category object-of for troop movement.

Our task is more general. Rather than selecting a single semantic category, our task is to find a sequence of semantic relations that links two nouns in a compound. Semantic rela­ tions are a list of about 50 thematic roles such as agent, ob­ ject, has-part, location, ... , . For example, given animal virus, a traditional interpretation may classify this as a location cat­ egory (animal virus is a virus in an animal). A loose-speak interpretation may be composed of a combination of semantic relations, such as: "an animal virus is a virus that is the agent of an invade, such that the object of the invade is the cell part of an animal".

Furthermore, computational linguists approach the noun compound interpretation task armed with lots of examples, but little or no knowledge about the constituent nouns. Typ­ ical solutions are based on statistical patterns discovered in the corpus of examples. In contrast, we approach the task in the context of the constituent nouns knowledge - their tax­ onomic classification, at least - but few examples of noun compounds, let alone a corpus.

2 Interpreting noun compounds during knowledge acquisition

During knowledge acquisition, the domain expert (or, more generally, the knowledge source) may provide a noun com­ pound in any dialogue that expects a noun. Our knowl­ edge acquisition system successfully interprets the noun com­ pound if it finds a correct sequence of semantic relations be­ tween the head noun and its modifier and builds a correct for­ mal representation of the noun compound.

If noun compound interpretation requires a priori, detailed knowledge of the head noun and its modifier, then the cost of acquiring this knowledge may overshadow the benefit of interpreting the compound. If, on the other hand, noun com­ pounds can be successfully interpreted without much knowl­ edge about the specific constituent nouns, then the problem is avoided, and a knowledge acquisition system might interpret one concept (the noun compound) while related concepts (the constituent nouns) are only skeletal. Knowledge bases tend to grow in this uneven way - following the lead of the knowl­ edge sources providing expertise - and a knowledge acquisi­ tion system should support it.

The purpose of this study is to determine what sort of knowledge is required to interpret noun compounds, and how this knowledge might be obtained. Before delving into the details of the study, it's important to understand what we are not attempting to do.

We are not presenting a novel algorithm for noun com­ pound interpretation. Our algorithm is quite simple and is derived from previous research. Also, we are not introducing a new type of knowledge representation or a novel technique of automated reasoning. Finally, we are not using a new, com­ prehensive knowledge base. We built a couple of them rather quickly and we're using another- not built for this task - "off the shelf."

In summary, what we are doing is evaluating the knowl­ edge requirements of a standard search algorithm applied to a variety of typical knowledge bases through a series of abla­ tion studies.

3 Experiments

The challenge in measuring an algorithm's sensitivity to knowledge base content is that the results may vary across domains and across knowledge bases. We attempt to neu­ tralize these factors by replicating our study in three domains with quite different knowledge bases.

The noun compound interpretation task can be viewed as follows: given a knowledge base encoded as a conceptual graph, and a pair of nouns corresponding to two nodes in the graph, find a path of semantic relations between them. The algorithm we used is a breadth-first search algorithm on a knowledge base. The algorithm is given a noun compound of the form< C1, C2 >where C1 and C2 are the KB con­ cepts that are mapped from the constituents of the given noun compound. Each of the first two steps conducts a breadth­ first search of the knowledge base along all semantic relation arcs. The first search starts from C1 and looks for C2 or any superclass or subclass of it. The second search starts from C2 and looks for C1 or any superclass or subclass of it. Step 3 combines the results, sorted by path length.

To avoid getting results that are skewed to a particular do­ main or knowledge representation, we used a variety of quite different data sets. The first consists of 224 noun compounds from a college-level cell biology text [Alberts et a!. , 1998]. The second consists of 294 noun compounds from a small engine repair manual. The third data set consists of224 com­ pounds from a Sun Sparcstation manual. The nouns used in these data sets are mapped to the corresponding concepts in knowledge bases on these topics.

Despite these commonalities, the K.Bs differ significantly. First they differ in terms of how they were built. The knowl­ edge base for the biology text was built using the generic Component Library [Barker et a!. , 200 I] to answer end-of­ the-chapter style questions, as one of the challenge problems for DARPA's Rapid Knowledge Formation project [Clark et a!., 200 I]. The knowledge bases for the other two data sets (the small engine repair manual and the Sparcstation manual) were built "on top of" the knowledge in WordNet [Fellbaum, 1998]. We augmented WordNet with the upper ontology of the generic Component Library plus about ten concepts that are important to each of the two domains whose partonomies are not complete in WordNet. Through this process, we en­ coded 416 concepts in about 50 man-hours. The advantage of using WordNet as the foundation for these knowledge bases is two-fold: i.t includes most of the terms used in the data sets, linked with both taxonomic and partonomic relations, and it is widely available and well used. The K.Bs also di ffer in content. Other than the shared upper ontology of the generic component library, they have few concepts in common.

The importance of each level of the ontology is measured through a series of ablations. When a level is ablated, the concepts on that level and all their axioms are deleted from the knowledge base. The superclasses of the subclasses of these concepts are changed to the superclasses of the concepts being deleted. As a special case, when the Oth level (the root level concept) is deleted, it is replaced by a generic concept of "Thing". Because the root level concept is vacuous, deleting it has no affect.

4 Results

Ablating levels of the ontology shows that they differ in their importance for the noun compound interpretation task. With­ out any ablation, both precision and recall of the noun com­ pound interpretation are around 80% across all three knowl­ edge bases. Ablating level one causes a big drop in both pre­ cision and recall. Ablating level two introduces a big gap be­ tween precision and recall because the algorithm does not find interpretations for many noun compounds. As lower levels in the ontology are ablated one at a time, the impact diminishes and performance improves to the level of a knowledge base with no ablations.

The contribution of the first two levels of the ontology is observed across all three data sets and knowledge bases. This pattern strongly suggests that top levels of the ontology are most important for the noun compound interpretation task, which is likely due to some combination of two factors:

1. Top levels of the ontology include concepts that make important ontological distinctions. For example, ablat­ ing the level that introduces Entity and Event blurs the distinction between obviously different concept, which causes the search to stop with erroneous results. Con­ sequently, many more interpretations are returned, and because we only use the first interpretation, the possibil­ ity of it being correct is reduced.
2. Although they contain relatively few axioms, these ax­ ioms are important for the task. The top-level ontology contains the most frequently used axioms, such as that "every Action involves an object that is acted upon". These axioms are used in the search as a step along the way. Deleting these axioms makes it difficult to find an interpretation for many of the noun compounds, thereby causing recall to Jag behind precision.

5 Conclusion

This paper reports an encouraging result: interpreting noun compounds does not require detailed knowledge of the con­ stituent nouns. Rather, it requires only that the nouns be cor­ rectly placed in a taxonomy, and that the taxonomy include the ontological distinctions and axioms commonly found in domain independent upper levels. These requirements are easily met in the context of knowledge acquisition, which is our focus.

We reached this conclusion using a novel experimental method. We measured the contribution of each level of the ontology to the task of interpreting noun compounds. We ab­ lated successive levels of the ontology one at a time, thereby conflating ontological distinctions and removing the axioms associated with concepts at each level. We found that the up­ per levels of the ontology for the KBs we used are the most important for noun compound interpretation. As lower levels in the ontology are ablated one at a time, the impact dimin­ ishes to nil. A more detailed version of this paper can be found in [Fan et al., 2003].

Acknowledgments

Support for this research is provided by a contract from Stan­ ford Research Institute as part of DARPA's Rapid Knowledge Formation project.