Tension Analysis in Survivor Interviews: A Computational Approach

Oral history plays a significant role for historians to understand the experience shared by the people from their past. One of the crucial benefits of oral history interviews, such as survivor interviews, is that it can shed light on important issues that might not have been present in previous historical narratives. Such interviews involve complex social interactions and different factors highly influence the interview situation such as complexity of human lives, age, intelligence, personal quality, etc. (Bornat, 2010). Both the interviewer and the interviewee contribute to these components during this dialogical process. In survivor interviews about holocaust, an interview goal is to better understand the interviewee’s related experiences. While such experiences are often associated with negative emotions such as fear, anger, or sadness and the interviewee is not willing to recall that memory and talk about it, the interviewer often needs to get to that and have the interviewee engage in this reflection. Tension can thus occur, as a result of conflict of interest or uneasiness. Researchers study these moments to gain an understanding of the conversational dynamics in these interview processes. For example, tension can be shown as reticence in the interview. Layman (2009) studied how reticence can cause conversational shifts by interviewees therefore putting a constraint on the responses the interviewees will offer to the interviewer. Layman (2009) discussed the importance of being aware of these situations so that the interviewer can better judge whether to press the interviewee.
In this work, we aim to identify the tensions in transcribed and translated interview transcripts of Rwandan genocide survivors. To approach this, we explored a list of potential indicators of tension including the use of hedging and boosting in the language, the sign of reticence, and emotion of the interviewees from their interview responses. Hedging refers to the technique of adding fuzziness to the propositional content by a speaker. People are known to use hedging to deal with controversial issues in conversations (Ponterotto, 2018). For example, the use of “I think…”, “Well, …” in interviews give interviewees the authority to shape their narratives. Phrases such as “In other words”, “In my understanding” can also be used to shift a topic either completely or partially. It can be used as a filler or delaying tactic. Boosting, using terms such as “absolutely”, “clearly” and “obviously”, is a communicative strategy for expressing firm commitment to statements. It restricts the negotiating space available to the hearer. It plays a vital role in creating conversational solidarity (Holmes, 1984) and in constructing an authoritative persona in interviews (He, 1993). Interestingly, if booster words are preceded by negated words such as (not, without), it can act as hedging (e.g., not sure). We identified lists of boosting and hedging words and phrases in this study to help us identify the tensions. We also developed a hedging detection tool to automatically identify the occurrences of hedging in the interview transcripts and achieved accuracy of 85.4% and F1-score of 81.9% for hedging sentences.
Tension can be defined as an emotion of physical and psychological strain joined by discomfort, unease, and pressure to look for alleviation via talking or acting (Nugent, 2013). To analyze whether and how the interviewee’s emotional aspect indicates the tension during the interview, we developed an emotion recognition tool to recognize the interviewee’s emotions from the interview transcript. Emotion recognition in computational linguistics is the process of identifying discrete emotion expressed by humans in text. Emotions lead directly to the past and bring the past somatically and vividly into the present (Misztal, 2003). Leveraging the high performance of deep learning compared to other machine learning approaches (Kim, 2014; Kalchbrenner et al., 2014; dos Santos and Gatti, 2014), we used a multi-channel convolutional neural network (CNN) model to recognize the emotions from the transcript. We achieved relatively high F1-scores in the emotion categories, compared to previous studies. For example, we achieved F1-scores of 72.6%, 73.6%, 76.8% and 46.0% for happiness, sadness, anger and fear, respectively.
We also considered prosodic cues such as laugh and silence as signs of reticence, but also acknowledge that further exploration is needed to interpret these cues. For example, while “laughter” may indicate invitations for the next question sometime, it may also represent hesitation or nervous deflection, i.e., the tension. We explored the traces of tension in the situations when an interviewee gives unusually long or short answers for a specific question type, shorter or longer than three standard deviation of the average length of responses of that question type (e.g., wh questions, yes/no questions, etc.).
Our algorithm of identifying the tensions in the interview transcripts combines all these aforementioned components. To evaluate the performance of this algorithm, oral history researchers first annotated tension points in the interviews, which provides us the “ground truth”. Then, we applied our algorithm to identify tensions from these interviews. With the two human labelled interviews available to us, the algorithm is able to identify six out of seven tension places in one interview, and three out of four in the other. The recall of the algorithm is thus satisfying. However, the algorithm suffers from low precision and we have 76 false positives in one interview and 56 in the other. On the other hand, given that the two interviews are very long (over 17,000 words for each) but there are only a few tension places identified by the researchers, this algorithm may work as a filtering tool in the tension annotation process. First, the algorithm identifies possible tension places, and then the researchers review them to identify the actual tension occurrences.
In conclusion, this study is interested in developing computational techniques to analyze where the interviewee’s tensions can be detected in the interview data in the context of survivor interviews about Rwanda Holocaust. This contributes to a better understanding of where tensions occur and how in the interviews. Such understanding helps researchers in the exploration of dialogical space created by the interviewer and interviewee in these conversational interviews.

Bibliography

Bornat, J. (2010). Remembering and reworking emotions: The reanalysis of emotion in an interview.
Oral History, 38(2), 43-52.

dos Santos, C., & Gatti, M. (2014). Deep convolutional neural networks for sentiment analysis of short texts.
In Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pp. 69-78.

He, A. W. (1993). Exploring modality in institutional interactions: Cases from academic counselling encounters.
Text-Interdisciplinary Journal for the Study of Discourse, 13(4), 503-528.

Holmes, J. (1984). Modifying illocutionary force.
Journal of pragmatics, 8(3), 345-365.

Kalchbrenner N., Grefenstette E., and Blunsom P. (2014). A convolutional neural network for modelling sentences.
In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 655–665.

Kim, Y. (2014). Convolutional neural networks for sentence classification.
In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pp. 1746–1751.

Layman, L. (2009). Reticence in oral history interviews.
The Oral History Review, 36(2), 207-230.

Misztal, B. (2003).
Theories of social remembering. McGraw-Hill Education (UK).

Nugent, Pam M.S., “TENSION,” in
PsychologyDictionary.org, April 29, 2013, https://psychologydictionary.org/tension/ (accessed November 27, 2018).

Ponterotto, D. (2018). Hedging in political interviewing.
Pragmatics and Society, 9(2), 175-207.