This application of artificial intelligence (AI) demonstrates two things: the astonishing types of questions algorithms can answer today, and (implicitly) what challenges and limitations we currently still face despite the current renewed hype in AI and deep learning.
The authors develop and implement a three-phase algorithm for answering opinion questions (for example, “Is Obamacare unconstitutional or not?”) based on the analysis of large sets of video clips.
Phase one detects sentiment-oriented (that is, positive, negative, or neutral) speeches in video streams. Phase two then locates (within these speeches) segments where opinion holders (as opposed to other types of speakers like moderators) express their views. The third phase then tries to match potential answers (that is, spoken words of identified opinion holders) to the original question by treating this as a form of translation problem from a “question” vocabulary to an “answer” vocabulary.
Whereas steps one and two use existing algorithms, the third step extends a published two-layer neural network to a four-layer neural network endowed with an increased number of neurons in and more connections between the layers. Despite the authors’ claim, however, this is not really a deep learning neural network because intra-layer connections or advanced features (for example, neurons possessing memory) are missing altogether. In total, the model contains 506,840 parameters, which have been learned from various (pre-analyzed) data sets.
A comparison with four existing models demonstrates that the performance of the new model is approximately 20 percent better than the second best. The study also finds that users prefer the identified video answers to the extracted text answers with statistical significance.
The work is geared toward the professional or academic fluent in Bayesian statistics as applied to sentiment analysis and neural networks because prior models used or even crucial acronyms (for example, discounted cumulative gain (DCG)) are not explained.
