We present a knowledge-rich methodology for disambiguating Wikipedia categories with WordNet synsets and using this semantic information to restructure a taxonomy automatically generated from the Wikipedia system of categories. We evaluate against a manual gold standard and show that both category disambiguation and taxonomy restructuring perform with high accuracy. Our approach produces an integrated resource, bringing together the fine-grained classification of instances in Wikipedia and a well-structured top-level taxonomy from WordNet.
The need for structured knowledge in AI is a long-standing issue, starting from McCarthy (1959) to current research. Previous efforts have concentrated on manually collecting knowledge, such as the WordNet project, which provides a semantic lexicon for English. However, WordNet contains very little domain-oriented knowledge and is populated with only a few thousand instances.
To overcome the limitations of manually assembled knowledge repositories, research efforts in AI and NLP have focused on automatically harvesting knowledge from semi-structured knowledge repositories such as Wikipedia. However, questions remain whether these automatically-induced knowledge resources achieve the same quality as manually engineered ones.
The Wikipedia categorization system is a thematically organized thesaurus, but it copes badly with very general concepts. The upper regions of the Wikipedia categorization are almost exclusively thematic, and no subsumption relation can be found while remaining inside the category network. For instance, COUNTRIES is categorized under PLACES, which in turn is categorized under GEOGRAPHY and NATURE.
We tackle these issues by proposing a two-phase methodology. The method starts with WikiTaxonomy, although in principle any taxonomy can be input. In a first step, the taxonomy is automatically mapped to WordNet. This mapping can be cast as a Word Sense Disambiguation (WSD) problem: given a Wikipedia category, the objective is to find the WordNet synset that best captures the meaning of the category label.
The optimal mapping is found based on a knowledge-rich method which maximizes the structural overlap between the source and target knowledge resources. As a result, the Wikipedia taxonomy is automatically 'ontologized'. Secondly, the mapping outcome of the first phase is used to restructure the Wikipedia taxonomy itself. Restructuring operations are applied to those Wikipedia categories which convey the highest degree of inconsistency with respect to the corresponding part of the WordNet subsumption hierarchy.
Our approach to taxonomy mapping and restructuring provides three contributions: first, it represents a sound and effective methodology for enhancing the quality of an automatically extracted Wikipedia taxonomy; second, as an additional outcome, we are able to populate a reference taxonomy such as WordNet with a large amount of instances from Wikipedia; finally, by linking WikiTaxonomy to WordNet we create a new subsumption hierarchy which includes in its lowest regions the fine-grained classification from Wikipedia, and in its upper regions the better structured content from WordNet.
Our methodology takes as input a Wikipedia taxonomy. First, it associates a synset with each Wikipedia category in the taxonomy. Next, it restructures the taxonomy in order to increase its alignment with the WordNet subsumption hierarchy. We take as input WikiTaxonomy, which can be viewed as a forest F of category trees. Each vertex in the tree represents a Wikipedia category. The label of this category is often a complex phrase, e.g. JAZZ HARMONICA PLAYERS BY NATIONALITY. In order to produce a mapping to WordNet, we need to find the lexical items heads(c) best matching each category label c, e.g. JAZZ HARMONICA PLAYERS can be mapped to any WordNet sense of player.
Terms in WordNet are first searched for a full match with the category label, e.g. plant for PLANTS. If no full match is found, we fall back to the head of the category. First, the lexical heads of a category label are found using a state-of-the-art parser. Then, we take as head of a category the minimal NP projection of its lexical head, e.g. public transport for PUBLIC TRANSPORT IN GERMANY. Such NP is found in the parse tree by taking the head terminal and percolating up the tree until the first NP node is found. If no such minimal NP can be found in WordNet, we take the lexical head itself, e.g. plants for EDIBLE PLANTS. In case of coordinations we collect both lexical heads, e.g. building and structure for BUILDINGS AND STRUCTURES IN GERMANY.
For each category tree T ∈ F and for each category c ∈ T, we first produce a mapping of c to a WordNet synset. We then restructure the taxonomy in order to increase its alignment with the WordNet subsumption hierarchy. This is done by applying restructuring operations to those Wikipedia categories which convey the highest degree of inconsistency with respect to the corresponding part of the WordNet subsumption hierarchy.
