Why Study Video Game Livestreams?

Topics: Machine Learning, Data Gathering. Audience: Knowledgeable.

Seeing as I have built this blogging platform on my website it seems only fitting to fill it. That said, I don’t plan to blog often and if my previous blogs are anything to go by this will be blog 1 of 2, but I digress. My research is into modelling video game lives streams, e.g. Twitch.tv. For those of you to who Livestreaming is a new term let me briefly explain. Livestreaming describes a paradigm where the events being broadcast are occurring live (as in live TV) and accessible through the internet. In the video game livestreams world a particular layout is common, where the broadcaster (streamer) broadcasts their game screen with an overlay of their webcam. Streamers then play their game while providing commentary to the viewers. Viewers then interact with the streamer (and other viewers) by posting messages in an open chat, where the streamer can read and respond to the messages.

A streamer (P4wnyhof) playing Player Unknown’s Battlegrounds
A streamer (P4wnyhof) playing Player Unknown’s Battlegrounds

So, why study streamers?

Livestreams have many many compelling reasons to study them. Here are a few which are most interesting for my research:

A map of the interactions in a stream
A map of the interactions in a stream

Data is hard

Hopefully, the above arguments have convinced you that studying video game streams has huge potential with implications not just in player experience modelling but also behaviour, emotion, and interaction modelling. However, it is not all plain sailing, stream data can be very difficult to work with as I will outline here:

Chat is a special beast

You may have noticed that in the above section makes no mention of chat. That is because I felt it needed its own section given what a challenge it is. Traditionally Natural Language Processing is focused on ‘proper’ English, e.g. Open AI’s headline-grabbing language model GPT-2. However, in a setting such as ours, we do not have the luxury of ‘proper’ English. In fact, Twitch chat is a very unique form of communication. Largely, communication occurs through the use of emoticon, emoji, emotes, and ASCII art (all variations of a common theme - communicating with images rather than words). In fact, in our data, we have found that the most common ‘word’ (token) is “lul” which, when sent on Twitch, actually takes the form of an emote. Furthermore, this communication is inherently temporal in that messages are sent in response to what the streamer is saying, what is occurring in the game, or other viewers. This is very different from the more commonly studied areas of NLP such as the language model work mentioned above. Emotes are also loaded with meaning despite being a single token. For instance, it is often common for emotes to constitute part of the identity of a streamers viewer community. This can manifest itself in many ways, e.g., emotes can act as a pseudo-shibboleth, identifying viewers as either ‘in’ the community or outside of it. My aim is to examine chat next as it is an area of stream data that I have not currently focused on and is arguably the most difficult to model due to it being so unlike the data used in the existing literature.

A small selection of some of the twitch specific emotes avalaible
A small selection of some of the twitch specific emotes avalaible

In closing

I hope this short blog post has shown you why studying livestreams is so attractive and exciting. Granted it is an extremely difficulty paradigm but in part that is what makes it so interesting to study - not only can advances in modelling livestreams help the video games research community it has the potential to have farther reaching implications. It is my hope and goal that through a collaborative effort from a range of disciplines, e.g. Games research, Computer Vision, Natural Language Processing, Affective Computing, we can begin to address some of the challenges outlined above and make progress on modelling this unique data source.