Deep Learning of Charisma
| Type of publication: | Idiap-Com |
| Citation: | Carron_Idiap-Com-03-2020 |
| Number: | Idiap-Com-03-2020 |
| Year: | 2020 |
| Month: | 8 |
| Institution: | Idiap |
| Abstract: | The study of charisma in the field of leadership studies has been steadily growing after it was shown to be a good predictor of success in the workplace and political setting. In order to study what charisma is, one must be able to define character traits that can be observed and quantified. In this paper, we will take a look at elements of speech that are used to measure how charismatic a speaker is. Modern studies have defined a total of nine Charismatic Leadership Tactics with metaphors, lists or contrasts being a few of them. Until now, these elements of speech were marked by hand by trained human annotators from transcribed speeches they wished to analyze. This is a long and costly process that takes a toll on the person executing it and is prone to bias. Here, we propose some methods using machine learning to perform the classification of text into these different classes. We explore state-of-the-art models and analyze their efficiency in performing this classification task. We begin by building a simple classifier with recurrent networks and linear projections which performs some amount of classification but can be largely improved on. We then introduce the mechanism of attention which greatly improves the performance by being able to focus on relevant words in the input sentence. Taking advantage of attention by increasing the number of heads looking at the sentences doesn’t have a noticeable effect on the results. Lastly, we use transformers which are said to attain state-of-the-art performance in text processing tasks and see that a BERT model slightly improves on our simple attention model. |
| Keywords: | |
| Projects |
Idiap |
| Authors | |
| Added by: | [ADM] |
| Total mark: | 0 |
|
Attachments
|
|
|
|
|
Notes
|
|
|
|
|
