A Wordnet from the Ground Up

58 Chapter 3. Discovering Semantic Relatednessdifferentiate the LUs in a synset from all other LUs similar but not synonymous, amongthem co-hyponyms. Any such MSR must therefore distinguish closely related LUs, notonly those with very different meaning.In modifying the WBST+H test we assumed that we needed to construct the answerset A so that non-synonyms are closer in meaning to the correct answer a i than it isthe case in WBST+H. Obviously, they cannot be synonyms of either a i or Q, but theyought to be related to both. We need to select the non-synonyms among LUs similarto s and to Q. In order to achieve this, we have decided to leverage the structure ofthe wordnet in the determination of similarity and to construct a semantic similarityfunction SSF W N based on the plWordNet hypernymy structure:SSF W N : S × L → R (3.2)where S is a set of synsets, L — lexical units, R — real numbers.SSF W N takes a synset S (e.g. including Q and a i ) and a lexical unit x (e.g.a detractor), and returns the semantic similarity value.During the generation of the modified Enhanced WBST [EWBST], non-synonymsare still selected at random but only from the set of LUs broadly similar to Q and a i .The acceptable values of SSF W N (S Q , x) are lower than some threshold sim t if thesynset S Q contains Q and a i , and x is a detractor. We tested several wordnet-basedsimilarity functions (Agirre and Edmonds, 2006), here implemented using plWordNet’shypernymy structure, and achieved the best result in a generated test with the followingfunction:SSF W N = p min(3.3)2dp min is the length of a minimal path between two LUs in plWordNet, and d is a maximaldepth of the hypernymy hierarchy in the current version of plWordNet. The similaritythreshold sim t = 2 for this function has been established experimentally. To achieveconsistency between tests generated from different versions of plWordNet, we decidedto set the sim t to value corresponding to four arcs in hypernymy hierarchy.The hypernymy structure of nouns in plWordNet does not have a single root,because in plWordNet we have not introduced any artificial common root nodes for allnominal LUs 4 Many methods of similarity computation require a root, however, so wehave introduced a virtual one for the sake of the similarity computation, and linked toit all trees in the hypernymy forest.We noticed that the random selection of LU detractors based any similarity measuretends to favour LUs in the hypernymy subtrees other than Q, if Q is located near theroot. The number of LUs linked by a short path across the root is much higher than4 The same is the case for verbal and adjectival LUs, whose hypernymy structures are also partial andquite shallow.