Considering the limited number of examples for each pattern, we developed pattern-independent methods for classifying the semantic relations us­ing the provided argument annotations and the context from the sentence.

We built two types of classifiers: binary that focuses on building a classifier for a specific rela­tion (SRNREL) and multi-class methods that build classifiers for all the SRN relations (SRN). Table 1 presents the accuracy of the classifiers built using different machine learning algorithms.

The classifiers were built using lexical, seman­tic, and syntactic features of the arguments, their phrases, their clauses, their common phrase/clause, and their modifier or head phrase. The system uses WordNet, an in-house Named Entity Recognizer, and an in-house Syntactic Parser for determining the values of some of these features. Table 2 pre­sents the list of features used by the SRN classifi­ers.

Argument's lexical, semantic, and syntactic features: the surface form, the label (POS tag or phrase label), the named entity (human, group, location, etc), the WordNet hierarchy (entity, group, abstraction, etc), the Semantic Scattering class (e.g. object, substance, etc), the grammatical role (sub­ject or object of the clause), the syntactic parser structure, the POS Pattern (the sequence of POS of the words from the argument), and the phrase pattern (the sequence of labels of the phrases, words from the argument);_ Argument's phrase features: surface form, label, gram­matical role, named entity, POS pattern, Phrase patterns; Argument's Modifier/Head features: the label, surface forms, NE, and WN Hierarchy for the first modifier, post modifier, pre-modifier, and head; Arguments' common tree node features: label, named entity, grammatical role, POS pattern, and phrase pattern, the tree path between arguments, and their order in tree; Arguments' clause: label, verb, voice, POS pattern, phrase pattern._

The second type of methods we used, were for particular patterns frequent in the training corpus. Table 3 shows the list of most frequent patterns in the training corpus. For having general pattern and covering the arguments that correspond to more than one node in a tree, we considered as argument the noun phrase that contains the nominal instead of the node for the nominal. 40 LCC relations and respectively Part-Whole rela­tion for the most frequent patterns from the SRN corpus (Table 3).

Pattern name

Noun compounds: NN1 NN2

If you are cleaning a <e1>coffee</e1> <e2>maker</e2> that hasn't been cleaned regularl271076ad y, repeat this step again with a fresh vinegar and water mixture._

Of-genitives:

S-genitives: NP1 's NP2

Prepositional constructions: NP1 IN NP2

Verbal construc­tions:

Verbal preposi­tional construc­tions:

The incoming <e1>chairman</e1> of the <e2>committee</e2> is promising an array of oversight investigations that could provoke sharp disagreement with Republicans and the White

This is the <e1>government</e1>s <e2>effort</e2> to encourage more employers to open up childcare centres at the respective ministries and gov­ernment departments._

I believe that unless we take this issue seriously, the red squirrel is facing eventual <e1>extinction</e1> from the <e2>woods</e2> of Scotland. On both of my systems, the <e1>reboot</e1> produced the omi­nous <e2>message</e2> 'Missing operating system '._

Manila radio station DZMM quoted survivors as saying that the <e1 >fre</e1 > started with an <e2>explosion</e2> in the cargo hold and spread across the ship within min­utes.

most frequent patterns found in the

For the pattern-dependent methods we adapted some of our existing binary and multi-class classi­fiers to work with the SRN relations.

For the SRN system we used only one binary classifier built for the Part-Whole relation (relation 6) using the ISS learning algorithm and trained/tested on the examples used in (Girju, Badulescu, and Moldovan, 2006) and different multi-class classifiers for the first 4 patterns from algorithms trained on a corpus annotated with 40 semantic relations (extracted from Wall Street Journal articles from the TreeBank collection and LATimes articles from TREC 9 collection) that includes the 7 SRN relations (or equivalents). (Badulescu, 2004) gives more details on this list of relations (definitions, examples, distribution on corpus, etc). Table 4 shows the accuracy of these classifiers on other WSJ and LAT articles for the the list of 40 LCC relations and the Part-Whole Rela­tion.