FAST2: An intelligent assistant for finding relevant papers

The SE literature is complex and immense. Most papers are rarely read so researchers and practitioners routinely miss important related work. Automatic active learners can help readers faster find more relevant works. But those tools offer no guidance on when to stop reading. Nor are those tools robust in the face of poor initial selection of papers, which leads to much longer reading time than average. This paper introduces the FAST2 text miner that addresses these problems, as follows. Firstly, FAST2 employs better tactics to quicker find a good initial set of relevant documents. As shown in this paper, with a little domain knowledge, FAST2 can find quickly find highly informative sets of initial papers, which leads to far more robust results while, at the same time, having to read fewer papers (by an average of 10% to 40%). Secondly, as reading progresses, FAST2 offers a new semi-supervised estimation model called SEMI that estimates the remaining number of relevant studies. Users can use SEMI to better understand when can they safely stop reading. When compared with the prior state-of-the-art estimator, SEMI is far more accurate.