06.08.2015 Views

A Wordnet from the Ground Up

A Wordnet from the Ground Up - School of Information Technology ...

A Wordnet from the Ground Up - School of Information Technology ...

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

74 Chapter 3. Discovering Semantic Relatednesscriteria can be based on <strong>the</strong> number of LUs described (<strong>the</strong> number of non-zero cellsfor <strong>the</strong> given feature), total feature frequency (with all LUs: a sum over <strong>the</strong> column)or statistical analysis, as in (Geffet and Dagan, 2004) discussed in a while.Hindle (1990) applied Mutual Information (MI) to compute feature weights inrelation to particular target LUs. Weights express <strong>the</strong> strength of association betweena target LU and a feature. Lin (1998) used a slightly modified version of MI based on<strong>the</strong> ratio of <strong>the</strong> information shared and total description. Lin and Pantel (2002) appliedPointwise Mutual Information modified by a discounting factor during LU semanticsimilarity computation – see <strong>the</strong> generalised version in (Mohammad and Hirst, 2006).Geffet and Dagan (2004) introduced Relative Feature Focus (RFF), a feature-weightingfunction based on a two-step transformation. First <strong>the</strong>y extract Lin’s MSR, filtering outfeatures with <strong>the</strong> overall frequency below 10 and MI weight below 4. Next, <strong>the</strong>y recompute<strong>the</strong> value of a feature f in LU u as <strong>the</strong> sum of MSR values of LUs most relatedto u such that <strong>the</strong>y have a non-zero value for f. The final MSR(RFF) is calculated <strong>from</strong><strong>the</strong> new feature values. Weeds and Weir (2005) proposed “a flexible, parameterisedframework for calculating” MSR, based on <strong>the</strong> idea of casting <strong>the</strong> problem as a CooccurrenceRetrieval Model (CRM). CRM describes semantic relatedness of two LUsin terms of weighted precision and recall of feature sharing between <strong>the</strong>m. Of severalweighting functions applied, <strong>the</strong> best results came with MI and t-score measures.The method of transformation is often tightly coupled with <strong>the</strong> computation of<strong>the</strong> final MSR value, e.g. (Lin, 1998, Weeds and Weir, 2005), but vector similaritymeasures independent of <strong>the</strong> weight function are also applied, e.g. cosine measuresor Jaccard coefficient (Mohammad and Hirst, 2006). All <strong>the</strong>se transformations stillassociate <strong>the</strong> calculated value of a feature with <strong>the</strong> initial frequency. For example, in<strong>the</strong> case of Lin’s measure more frequent features do not only get values higher than lessfrequent features; <strong>the</strong> value level of frequent features is also higher than <strong>the</strong> value levelof those less frequent. We have noticed that this phenomenon negatively affects <strong>the</strong>accuracy of MSR (Piasecki et al., 2007a). In a new weighting function proposed, <strong>the</strong>z-score measure was combined with a Rank Weight Function [RWF], a transformation<strong>from</strong> values to ranks.The main idea behind RWF is to put more emphasis in <strong>the</strong> description of LUmeaning on <strong>the</strong> identification of features most relevant to this LU. The calculation of<strong>the</strong> exact values of <strong>the</strong> strength of <strong>the</strong>ir association with <strong>the</strong> target LU is less important.We believe that <strong>the</strong>se values are largely <strong>the</strong> artefact of <strong>the</strong> biased corpus frequencies,so one should not depend on <strong>the</strong>m too strictly during row vector similarity calculation.The particular order of relevance of <strong>the</strong> features delivers clearer information. In RWF,<strong>the</strong> meaning of <strong>the</strong> given LU is described by an ordered set of relevant features, and<strong>the</strong> meaning of two LUs can be compared on <strong>the</strong> corresponding sequences of featuresordered by relevance. In order to keep <strong>the</strong> correlation between relevance and feature

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!