Better Approaches to Analysing Twitter Reply Chains

The final speaker in this session at AoIR 2016 is my DMRC colleague Brenda Moon. She points out that hashtag studies on Twitter are subject to significant limitations because they capture only those tweets that have been explicitly marked with those hashtags, but may not also examine the broader conversation that might unfold around those hashtagged tweets without being itself hashtagged. There is a need here to move beyond quantitative and computational analysis of these datasets as well – so the challenge here is to identify reply chains and to examine them more qualitatively.

The present study focusses on discussion relating to Uber in Australia; it commenced by capturing tweets containing the term 'Uber' that were posted by Australian users, and followed the reply chain from these tweets both forwards and backwards. This approach identified some 17% of additional tweets beyond the seed dataset containing the term Uber – demonstrating the volume of content usually lost in conventional hashtag and keyword studies.

The key moment selected for further analysis from this dataset is a controversy around Uber search pricing on New Year's Eve 2015; in this timeframe, the reply chain approach found 17% more tweets and 12% more users than then keyword tracking itself identified. What emerges from a network analysis of this is that the reply chains around the seed tweets are usually relatively short, but some longer chains also emerge.

This analysis also points us towards a number of particularly extended reply chains that are worth a closer reading. So, for instance, one longer reply chain starts with a post by a prominent Australian journalist discussing surge pricing, which kicks off a further discussion of his points that more ordinary voices are also participating in. Additionally, from such analyses it is also possible to explore the various hashtags being used in these conversations.

Such qualitative observation and exploration of social media datasets therefore adds substantial additional information and detail. It especially enables the identification of conversational engagement, and the visualisation of the interaction networks is especially useful in discovering new themes being addressed by participants. There is also a need for further tool development, especially if it also enables the use of media objects rather than keywords or hashtags as seeds for the conversation network analysis.