Natural Conversation and NVC

What Is Natural Conversation

Natural conversation was defined by Stubbs (1983) as "without any intervention from the linguist" and "is spontaneous in the sense of unplanned, and which is composed in real time in response to immediate situational demands". Social situations that have intervention from an experimenter are therefore not necessarily natural. Also, people who are informed they are being recorded alter their behaviour. This change is known as reactivity and presents a major experimental difficulty to investigation of natural conversation.

The options available for recording social situations are:

• Record people surreptitiously and ask for permission after the fact. (This raises ethical concerns.)
• Record people with prior consent but at the risk of the data not reflecting the "natural" state due to reactivity.

We have opted for the second option while attempting to minimise effect of experimenter by placing the least constraints on participants, while still satisfying the need for usable data. This will allow the study of social situation sensitive non-verbal communication given experimental constraints.

Data Capture

Our data capture consisted of inviting two colleagues of roughly equal social status to our media lab and asking them to be seated at a table. They then conversed for at least 12 minutes. The subject of conversation was not specified. The seated position made recording of the conversation and later processing more practical.

Multi-observer Annotation

Annotation of video performed by multiple people usually results in a low interannotation agreement. Annotators are the observers of the video trying to find examples of a particular social signal, or who try to classify a social signal into one of several categories. The annotators are sensitive to many social and cultural factors (Reidsma et al. 2008). An attempt to annotate video using a single observer would only result in a narrow view of the social signals contained in the social situation. We overcome this by using multiple annotators and recording their demographic information.

A web based survey was created to annoy users to conventiently view short clips of view and to rate them on an 11 point scale. For the analysis presented here, there were 21 participants. Each clip had an average of 3.8 annotators providing scores. This process resulted in a set of 527 clips with multiple annotations of 4 different categories: "I understand what you", "I agree with you", "I am thinking" and "I am asking a question".

Last update: July 2009.