We saw that global and local scopes can matter in finding similarities. We discuss a different paper in this approach: This is the paper on a Graph Based Contextualization method using Distributional Thesauri by Biemann and Riedl. This paper discusses a new contextual similarity finding method that generates for each term occurrence in a text, a ranked list of terms that are semantically similar and compatible with the context. The framework is instantiated by the definition of the term and context. The dependencies are parsed to get the context. Contextual similarity based approach has trumped the non-contextualized baseline across all parts of speech. Moreover this is an unsupervised generative method for similarity in context  and does not rely on the existence of lexical resources as a source for candidate expansion. This approach made use of an existing distributional thesaurus (DT) that has entries consisting of a ranked list of the globally most similar terms for a target term. This approach uses the DT in a graph representation and moves from a global notion to local. The prevalent way to represent word meanings has been to use the vector space model which uses a corpora. However this approach does not require a domain knowledge and is unsupervised in the sense that it does not require training.
They introduce an operation called the holing which splits any sort of observations on the dependencies into pairs of terms and context features. Then these pairs are used for calculating the similarity and for finding the contextual similarity. This operator is used in the system that requires preprocessing steps. The notation is the @ symbol. This is used in more unified representations of dependencies.

The distributional similarity is represented using three graphs. This operation involves using a bipartite term-context feature graph, with T the set of terms, C the set of context features, and e(t,c,f) where f = count(t,c) .count is the count of terms and count of c Based on the graph TC , a first order graph e(T,C, sig) is created and the significance score sig calculated. The significance score sig is calculated for each pair (t,c) using a formula defined in Evert 2004 using counts.
Then all the edges with score (t,c) that is < 0 is removed and only the p most significant pairs per term t and removing the remaining edges. Additionally overly general features that have more than a thousand words are removed.
The second order similarity graph between terms is defined as SO(T,E) for t1,t2 belonging to T with the similarity score which is the number of salient features two terms share. This graph gives the distributional similarity.