blog

I have three representations of Dr Who. S4E1 sitting in front of me:

1. A Semantic annotation of the episode based on the BBC Stories Ontology, both by Michael O. Jewell, Paul Rissen, and Toby Harris{{1}}.
2. The script for the episode, by Russel T Davies
3. A transcript of a couple of very rowdy screenings of the episode I organised at The People Speak HQ during which people heckled at the screen using short messages, images and video.

What’s hurting my brain at the moment is a question of representation. In this triple, if ‘represents’ is the predicate, which is the subject and which is the object?

• Is the Semantic annotation a representation of Dr Who S4E1: Partners in Crime the TV show, or is it a representation of the experience and interpretation of the person watching and annotating it? Or both?
• In the same way, is the transcript of the conversation a representation of people’s experience of watching the episode and making social sense of it together, but with a lot more context?
• Is the episode itself a representation of the shooting script?

Which philosophical texts can I turn to to help me make sense of this?

But most crucially (for my purposes), how can I best understand the similarities and differences between 1 (the semantic annotation) and 3 (the conversational transcript)?

I had a few ideas about this, mostly based on text-mining the conversation transcript via concept-extraction services such as LUpedia or Alchemy API to see if snatches of conversation can identify related entities within the annotation’s timeline, but feedback from the wonderful Jo Walsh was sceptical of this approach.

Basically, her critique was that

1. Using text-mining/concept extraction privileges text, whereas the heckle stream is very visual, and that seems important.
2. Entity-recognition/tagging services will yield a very variable quality of metadata. They’re designed to look for something specific in text and match it, and tend to require quite a bit of context (more than 140 characters of text)
3. Asking the question “to what extent can this be considered metadata” will get very inconclusive answers, which will question the point of asking the question in the first place.

I think I agree with point 3 – which questions the point of this blog post, but I think I still need some kind of bottom-up analysis of the relatedness of the data, and although I’d like to just disregard the slightly solipsistic question of what is representing what, it would be nice to be able to attribute any philosophical assertions to someone other than myself!

[[1]] Here’s the OWL for the episode. Here’s the n3 formatted annotation of the episode [[1]]

This scenario is designed to elicit and capture conversation between a group of people who are watching a specific episode of Dr. Who together.

The aim is to be able to compare existing formal metadata for this episode with this speculative ‘conversational metadata’, and evaluate it as an alternative representation of the same media object: Dr Who, Season 4, Episode 1, Partners in Crime.

The Setup

Two groups of eight people are invited to watch of an episode of Dr Who together on a large screen, during which they use their laptops and a simple text/image/video annotation interface to type short messages or send images onto the screen where they are visible as an overlay on top of the video of Dr Who.

The room is laid out in a ‘living room’ arrangement to support co-present viewing and interaction between participants, with comfortable seating arranged in a broad semi-circle, oriented towards a large projected video screen about ten feet away. Each participant is asked to bring their own laptop, tablet PC, or other wifi-enabled device with a web browser.

After making sure that all participants are on the network, there is an introductory briefing where they are given a presentation explaining the aims of the project and that they are free to walk around, use their laptops or just talk, and help themselves to food and drink during the screening.

The Annotation Tool

The system that the participants are using on their laptops/tablets or mobile phones has a simple web-based client, enabling viewers to choose a colour to identify themselves on the screen, and then type in 140 characters of text or search for images and video, before sending them to the main screen.

The Display Screen

The video of Dr Who is projected on a ‘main’ screen, alongside text, images and video clips sent by viewers in a fullscreen browser window. The images and videos sent by users have a coloured outline, and text-bubbles are coloured to indicate who posted them.

Images and videos run underneath the video in a ‘media bar’, while text bubbles posted by users drop onto the screen in random positions, but can be re-arranged on the screen or deleted by a ‘facilitator’.

Rationale

This ‘conversational scenario’ is a hybrid of various methods in which researchers have contrived situations to elicit data from participants. Before making any claims about the data gathered, some clarification of the purpose and methods of the scenario are necessary.

Ethnographic Studies of Social TV have tended to use audiovisual recordings of TV viewers in naturalistic settings as their primary source, and analytical methods such as Conversation Analysis and participant observation have been used to deepen their understanding of how people use existing TV devices and infrastructures in a social context.

HCI approaches to designing Social TV systems have built novel systems and undertaken user testing and competitive analysis of existing systems in order to better understand the relationship between people’s social behaviours around TV, and the heuristics of speculative Social TV{{1}} devices and services.

Semantic Web researchers have opportunistically found ways to ‘harvest’ and analyse communications activity from the Social Web, as well as new Social TV network services that track users’ TV viewing activity as a basis for content recommendations and social communication.

All of these approaches will be extremely useful in developing better conversational annotation systems, and improving understanding and design of Social TV for usability, and for making better recommendations.

Although the conversational scenario described borrows from each of these methods, it’s primary objective is to gather data from people’s mediated conversations had around a TV in order to build a case for seeing and using it as metadata.

System design, usability, viewer behaviour, user profiles, choices of video material, and the effect those issues have on the quality and nature of the captured metadata are a secondary concern to this first step in ascertaining whether conversations can be captured and treated as metadata pertaining to the video in the first place.

[[1]]I am using the term Social TV, following one of the earliest papers to coin the phrase by Oehlberg et. al (2006) to refer to Interactive TV systems that concentrate on the opportunities for viewer-to-viewer interaction afforded by the convergence of telecoms and broadcast infrastructures. Oehlberg, L., Ducheneaut, N., Thornton, J. D., Moore, R. J., & Nickell, E. (2006). Social TV: Designing for distributed, sociable television viewing. Proc. EuroITV (Vol. 2006, pp. 25–26). Retrieved from http://best.berkeley.edu/~lora/Publications/SocialTV_EuroITV06.pdf [[1]]

Annotation of a conversation would usually be a post-hoc chore undertaken by someone charged with watching a documentary or ethnographic video and ‘making sense’ of the diffuse multifariousness of a conversation. Heckle‘s approach, that each visual/textual interjection might be used as an annotation, attempts to turn annotation into an augmentation of the experience of the conversation. Because it is concurrent and live, the participants who heckle may notice and incorporate all kinds of contextual markers outside the view of the video camera, as well as bring their own diverse interpretations and experiences into the heckled conversation.

Most crucially, the variety of ways that people use the Heckle system mirrors the diversity of people’s verbal and non-verbal contributions to the live conversation. The stream of images, video, text and links that result can be seen as a parallel conversation that ‘annotates’ the conversation around the Talkaoke table, but also interacts with it in real time: the representation of the conversation itself becomes conversational.

Research Strategy

There are so many questions to be asked about this approach: about the user interfaces, about how and whether Heckle does really ‘augment’ the experience, lead to further engagement, and how it influences people’s interpretation and behaviour. However, with the time and resources available at this stage, the goals will have to be very limited and specific.

My research task at hand is to enquire about this ‘conversational metadata’: what is it? To what extent can it be considered ‘metadata’? What objects does it’s metadata relate to; to the conversation around the Talkaoke table, or to the people Heckling? And to what extent does it correlate (or not) with other forms of annotation and representation of these objects?

Asking this question will involve re-purposing the Heckle system to create scenarios in which this correlation can be measured.

To be specific, rather than using Heckle to annotate a live conversation around the Talkaoke table, I will be using it to annotate a group of people watching Dr. Who Season 4 Episode 1, ‘Partners in Crime’.

This is an opportunistic choice of programme, suggested to me by Pat Healey because he happens to have supervised my MAT colleague Toby Harris on the BBC Stories project, with Paul Rissen and Michael Jewell to annotate this episode of Dr. Who in an exemplary ‘top-down’ fashion, developing and then using the BBC stories ontology.

The plan, then, is to gather a group of people to sit watch TV together, and to provide them with The People Speak’s Heckle system as a means of interacting with each other, layered on top of the Dr Who video. The resulting conversational metadata can then be compared to the detailed, semantic annotatations provided by the BBC Stories project.

Evaluation Strategy

There are a number of possible methods to use to make this comparison, although it will be hard to tell which to use before being able to look at the data.

It may be useful to simply look at mentions of characters, plot developments, and other elements in the BBC Stories ontology, and see whether they appear at equivalent moments in the BBC Stories annotation and the heckled conversation. A basic measurement of correlation could be gathered from that kind of comparison, and would indicate whether the two forms of metadata are describing the same thing.

Similarly, it might be useful to demonstrate the differences between the conversational metadata and the BBC Stories version by looking for conversational annotations that relate to the specific context of the experience of watching the episode: the space, the food, the sofa. These would (of course) be absent from the BBC Stories annotation.

However, another strategy, which Toby Harris and I concocted while playing with LUpedia, a semantic ‘encrichment’ service that takes free text, and attempts to match it with semantic web resources such as DBPedia, and return structured metadata.

If it would be possible to feed LUpedia the BBC Stories ontology, and then feed it with the episode of Dr Who in question as a dataset, it should be possible to submit people’s heckles to it, and see if LUpedia returns relevant structured data.

If LUpedia can enrich people’s Heckles with metadata from the BBC Stories dataset, that should indicate that the heckles are pertinent to the same object (in this case, the episode of Dr Who), and might therefore be seen as conversational metadata for it{{1}}.

[[1]]My conversational metadata will probably also describe the interactional experience of watching the show, and other contextual references that will be absent from the BBC Stories annotation. However, it is important to show that the two types of metadata relate to at least one of the same objects. If this is not demonstrable, it does create some ‘fun’ philosophical problems for my research such as what conversational *does not* refer to. That one might be harder to answer.[[1]]

 

In my last post about the Social TV research context I explained that I decided to focus on generating and evaluating conversational video metadata by eliciting mediated conversations through SocialTV.

Before drilling down into that choice to distill a set of research questions, however, there’s a more basic question about the context of research into SocialTV metadata: why gather SocialTV metadata at all?

The Notube project has a diagram on this issue, which shows the progression of a piece of TV content along a timeline from pre-broadcast to media archive. The assertion is that having more metadata about user preferences enhances the value of TV content because it provides many more opportunities to recommend programmes. Although the Notube project adds a great deal to existing research on recommendation systems, Notube’s central focus follows most existing research in focusing on developing more accurate and sophisticated user profiles (in this case through aggregating media consumption habits from heterogeneous sources on the web).

The way the diagram is re-used in my slide above emphasises the corrolary of the point that it is intended to illustrate: that having more TV programme metadata would also enhance the frequency and accuracy of recommendation for content (and thereby its value) throughout its lifecycle {{2}}.

If metadata (about profiles or programme content) can enhance the value of a TV show and facilitate production, discovery and delivery of TV programmes because it increases the likelihood of that programme being recommended, then it should follow that the greater the richness and referential diversity of that metadata, the more ‘recommendable’ the programme becomes.

This suggestion presupposes that the metadata in question is relevant, rather than a random spamming of references, intended to maximise recommendation in all possible contexts. So the question then becomes: what determines relevance in this context, and, even more importantly, relevance to what?

It may seem self-evident that the metadata about users should be relevant to their consumption habits, and that these habits, recorded and aggregated should constitute their preferences or dispreferences. Even if this widespread assumption holds true, what should programme metadata relate to? Are broadcaster’s assertions about their programes necessarily relevant to the way those programmes are discovered and interpreted? And crucially, in SocialTV – which almost by definition is about the interaction between viewers brokered via a networked TV infrastructure, how do those metadata correlate with the way TV programmes are used as a prop, touch-point, or stimulus to conversation between viewers{{1}}?

Current content discovery heuristics tend to rely on user profiles and broadcaster-provided metadata, without necessarily taking into account the context and quality of interactions between users.

The hypothesis of this research project is that metadata derived from conversations between viewers via SocialTV can provide a crucial additional component to support the interactional possibilities of of SocialTV.

To test this, an experimental scenario will be developed to involve concurrent, co-present viewers of a TV programme in a public multimedia chat system, designed to elicit metadata from their conversation. The transcript of their interactions will then undrego Conversation Analysis. This analysis will provide a baseline to evaluating to what extent conversations around a SocialTV experience can be correlated to a detailed and highly granular ‘top down’ semantic annotation of a TV programme.

An analysis of the data may also be used to test several related hypotheses:

• that conversational metadata are likely to have more divergent subject matter and more external reference than a-priori programme data about actors, characters and plot developments
• that conversational metadata are likely to be more responsive to ways in which the context of the conversation changes{{3}}.

  

So in terms of my earlier exploration of SocialTV as a research context, I can start to narrow my focus onto a few areas.

To achieve the study outlined above, I am proposing to deploy a multimedia discussion interface that will allow co-present concurrent viewers of a TV programme to interact and converse as freely as possible. Although it may have significant design issues, the state of the art in this context is definitely the ‘hot topic’ of Second Screen/Companion Devices. This is not the central point of the study, however, so I will be approaching this part of the project as a design process – building on prior art – which is abundant at the moment – and iterating out a customised version as quickly as possible to find something basic that works well enough for me to get the conversational data I’m looking for.

The other research objectives in the slide above are already in order of priority: eliciting and capturing discussion is the most pressing need for the system. Other functions and features are interesting, but probably out of scope for the time being. However, I might well post some ideas to this blog about how conversational metadata might underpin new approaches to searching, filtering, annotating and segmenting video.

[[1]] There’s still the thorny issue of what determines ‘relevant’ in this context. There may be perceptual studies comparing recommendation engines or other ways of trying to determine relevance of metadata statistically. However, for the purpose of this study, the question is moot. Relevance, especially in terms of conversational metadata, is subjective unless based on evidence of how it is used to broker or support an interpersonal interaction via SocialTV.[[1]] [[2]] This is also core to Notube’s aims and methods: using highly granular metadata about programmes to create ‘serendipitous’ trails through content that can break through the sameness of personalisation recommendation systems that can tend to channel users into relatively static, homogenous clusters of content. [[2]] [[3]]For example, a programme might be broadcast 50 times over 15 years. Conversational metadata associated with the programme might change significantly over time and in the different contexts in which it is watched, accumulating qualified layers of annotations, whereas broadcaster-provided metadata is likely to remain static.[[3]]