Article Preview
TopIntroduction
Sensing affect related states, including interest, confusion, or frustration, and adapting behavior accordingly, is one of the key capabilities of humans; consequently, simulating such abilities in technical systems through signal processing and machine learning techniques is believed to improve human-computer interaction in general (Schuller & Weninger, 2012) and computer based learning in particular (Aist, Kort, Reilly, Mostow, & Picard, 2002; Forbes-Riley & Litman, 2010). Important abilities of affective tutors or lecturers, besides emotional expressivity (Huang, Kuo, Chang, & Heh, 2004), include the choice of appropriate wording, which has been found to be highly important in computer based tutoring to support the learning outcome (Narciss & Huth, 2004). Furthermore, there is increased evidence for the influence of affect related states on the learning process (Craig, Graesser, Sullins, & Gholson, 2004; Bhatt, Evens, & Argamon, 2004; Forbes-Riley & Litman, 2007). In particular, previous studies highlighted the relation between system responses in a tutoring dialogue and student affect (Pour, Hussein, Al Zoubi, D’Mello, & Calvo, 2011); it turned out, for example, that dialogue acts of an automated tutor influence student uncertainty (Forbes-Riley & Litman, 2011). However, these studies do not take into account the linguistic content of lectures as a whole; hence, we aim to bridge this gap by addressing the automatic assignment of categorical affective ratings by a large audience to on-line lectures from the TED talks website (www.ted.com/talks). This prediction is based on learning the relation between linguistic features of the speech transcripts and the ratings given by the audience, which comprises many thousands of internet users in our case. Such automatic predictions can be immediately useful to evaluate the quality of lectures given by a distant education system, and to gain insight into which lecture topics or lecturing strategies are related to certain affective states. The aspect of predicting the induced affect from the lecturers’ speech has—to our knowledge—not been addressed in a systematic fashion so far: Rather, in (Forbes-Riley & Litman, 2011), features from student responses to the system and abstract goals of the dialogue manager are used to analyze student affect. In this respect, that study is somewhat related to sentiment analysis (Schuller & Knaup, 2010) or opinion mining (Turney, 2002), where the goal is to deduce the affect of the users from written reviews. However, in our study we aim at predicting the users’ affective ratings based on the lectures themselves. This also distinguishes our contribution from the large body of literature on prediction of (ordinal-scale) movie ratings—for a recent study on the public Internet Movie Database (IMDB), we refer to (Marovic, Mihokovic, Miksa, Pribil, & Tus, 2011). In that field, in contrast to our study, the vast majority of approaches seem to be exploiting similarities in user profiles rather than features of the rated objects (instances in terms of machine learning), such as in (Marlin, 2003).