A Wordnet from the Ground Up

3.4. Measures of Semantic Relatedness 85• neutral – several LUs on the list are in some relation, but the linguist might missthem,• useless – at most a few LUs may be related.The results of the manual evaluation appear in Table 3.15.PoS very useful useful neutral useless no relationsVerb [%] 17.8 37.6 20.0 15.6 9.0Adjective [%] 19.2 26.3 29.7 14.4 10.4Table 3.15: Manual evaluation of MSR for verbs and adjectives performed for (Broda et al., 2008) onthe MSRs from that timeSelected lists for verbs and adjectives are shown in Tab. 3.16. The English translations“select” the meaning common to the grouping that the list suggests.In nearly half of the cases, the linguist can find valuable hints on the list generatedon the basis of MSRs. Suggestions should help notice specific or domain-restricteduses of LUs. The manual evaluation suggests MSR accuracy much lower than forthe WBST, but the latter operates on generic semantic similarity rather than specificsemantic relations. However, besides the relatively large percentage of the MSRlist (x,k)lists evaluated as very useful and useful, the percentage of correct hints – pairs ofLUs in some wordnet relation – was still significantly below the 50% threshold ofacceptance for the MSRlists (x,k) as a valuable tool for linguists. In order to increasethe accuracy of automatic extraction of instances of wordnet relations, we need to lookfor additional knowledge sources. In the case of adjectival LUs and verbal LUs theuse of lexico-syntactic patterns is hardly possible, cf Section 4. Only antonymy foradjectives seems to be marked by specific language expressions. If we redefine thesupport task to expanding a wordnet, the wordnet structure already in place becomesan additional source of knowledge. We will explore this approach in relation to verbalLUs in experiments presented in Section 4.5.3.In the case of all three parts of speech, the constructed MSRs have the accuracy inWBST+H and EWBST (nominal MSR only) which surpasses that of the MSRs basedon the algorithms proposed in literature. Due to the nature of the applied tests, we canconclude that all three MRS extracted have the ability to distinguish among semanticallyrelated and unrelated LUs with an accuracy that is relatively close to the average resultsof humans in the same task. Let us note that the results may have been different hadthe users been trained linguists. Nevertheless, having these three good MSRs, we havestill not achieved a tool of practical importance for semi-automatically expanding thecore plWordNet: generated MSRlist (x,k) lists include too many LU pairs which donot belong to any wordnet relation. Our experience makes us pessimistic about the