Simple and efficient machine learning frameworks for identifying protein-protein interaction relevant articles and experimental methods used to study the interactions

Table 3 Runs submitted for ACT

Run number	Label	Classifier Algorithm	Type of features used	Number of features	Training data
1	NBM-12-1k-td	NBM	Unigrams and Bigrams	1000	Training+Development
2	NBM-12-400-td	NBM	Unigrams and Bigrams	400	Training+Development
3	NBM-12-1k-d	NBM	Unigrams and Bigrams	1000	Development
4	SVM-12-400-d	SVM	Unigrams and Bigrams	400	Development
5	SVM-12-400-td	SVM	Unigrams and Bigrams	400	Training+Development
6	NBM-1-1k-td	NBM	Unigrams	1000	Training+Development
7	NBM-1-400-td	NBM	Unigrams	400	Training+Development
8	NBM-1-1k-d	NBM	Unigrams	1000	Development
9	SVM-1-400-d	SVM	Unigrams	400	Development
10	SVM-1-400-td	SVM	Unigrams	400	Training+Development

For the BioCreative III challenge, each participating team was allowed to submit 10 runs for ACT. Five runs could be submitted offline and the other five runs could be submitted online, using XML-RPC. Runs 1-5 were submitted offline, while runs 6-10 were submitted online. For all runs, we used mutual information feature selection algorithm, as it gave better performance than chi-square score. We submitted 10 runs, listed here.