A Wordnet from the Ground Up

88 Chapter 3. Discovering Semantic Relatednesshand, the one-sense-per-discourse heuristic states that a lemma is used only in one ofits senses in one discourse (Agirre and Edmonds, 2006). Combining both heuristics,we can assume that polysemous lemmas will be used in one dominant sense in onecluster.One can hope to alleviate the ambiguity present in MSR by incorporating knowledgeof the domain of the documents that contain the given lemma. Document hierarchycould also be used as a base structure for a wordnet (or only parts of a wordnet).The approach based on document clustering in sense discovery is discussed inSection 3.5.1. Another remedy for inherent polysemy in any MSR can be a specializedalgorithm, for example Clustering by Committee [CBC] (Pantel, 2003). Section 3.5.3describes an adaptation of CBC to Polish, and an extension.3.5.1 Document clustering in sense discoveryDocument clustering in our work had two reasons. First, we wanted to explore thepossibilities of extracting knowledge about polysemy of lemmas from document groups.One-sense-per-discourse heuristic suggests that a polysemous lemma will appear in agiven domain only in one of its meanings. On the other hand, document clusters can belabelled with keywords – most representative words for a document group. Arrangingdocument clusters in a hierarchical tree could form the basic structure for a wordnet.There are many clustering algorithms. Following a review of the possibilities (Jainet al., 1999, Forster, 2006, Broda, 2007) we chose two algorithms for further analysisand experiments. We looked at following properties of clustering algorithms: the abilityto cluster high-dimensional data (such as documents represented by vectors), the abilityto detect clusters of irregular shapes and the possibility of building hierarchical trees.There are many ways of representing documents for clustering (Forster, 2006,Broda, 2007). In this work we used the Vector Space Model. In this model documentsare represented as vectors in high-dimensional space. Each dimension of the space correspondsto occurrences of a specific word. Vectors store data describing occurrencesof words in documents.RObust Clustering using linKs [ROCK] (Guha et al., 2000) follows the agglomerativeclustering scheme. Initially, each document is in a one-element cluster. Pairs ofthe most similar clusters are merged iteratively. The algorithm differs from others inhow the merging is decided. ROCK selects for merging a cluster that maximises thenumber of links between documents. To avoid oversized clusters (or even putting alldocuments into one cluster), the algorithms imposes an expected number of links fora cluster of a given size.The notion of links can be explained by common neighbours. Neighbourhood isdefined using a similarity function: if two documents are similar enough, they are consideredneighbours. If links replace similarity in clustering, global information about