16th Annual Conference of the International Speech Communication Association

Dresden, Germany
September 6-10, 2015

Verbal Intelligence Identification Based on Text Classification

Roman Sergienko, Alexander Schmitt

Universität Ulm, Germany

This paper analyses and compares term weighting methods for automatic verbal intelligence identification from speech. Two different corpora are used; the first one contains monologues on the same topic; the second one contains dialogues between two or three people. The problem is described as a text classification task with two classes: low and high verbal intelligence. Seven different term weighting methods were applied for text classification using the k-NN algorithm. The best result is obtained with the ConfidentWeights method as a term weighting method for the dialogue corpus. The best classification accuracy equals 0.80 and the best macro F1-score equals 0.79. The numerical results have shown that highest scores can be obtained when using a very small number of terms which characterize only the class of higher verbal intelligence.

Full Paper

Bibliographic reference.  Sergienko, Roman / Schmitt, Alexander (2015): "Verbal intelligence identification based on text classification", In INTERSPEECH-2015, 2524-2528.