Evidence Based Approach for Sentence Extraction from Single Documents

Abstract

We present an evidence based sentence extraction model which is an application of subjective logic in a document computing scenario, to rank sentences according to their írnportanec in a document. Elements from the Dempster-Shafer belief theory arc used by this model to measure the subjectíve belief or opinion about a sentence. The important sentences extracted by this model can be seen to summarize a document partially. For qualitative analysis, this method is compared with two different open source summarizers along with human extracted sentences which are used as benchmarks for this purpose.

This model also iruproves the effect of signal to noise ratio on sentence rank by applying the whole evidence based model on a reduced data set to evaluate its stability and accuracy. Since evidence based models are cornputationally very expensive, here we show that one third of the words of a document are sufficient to rank sentcnces similarly to human judgements, but if reduced further, the accuracy drops. The results show thatour evidence based model outperforms standard summarizers when evaluated with human ranked sentences.