Shoulder Angel

Shoulder Angel:

Proactive Machine Augmentation for Interaction with Information

Luke Guolong Wang

Idea & Motivation:

Apart from what we can first-handedly see and hear, our perception of the world depends on information presented to us by the mass media. We therefore effectively live in a world of information that is mostly “weaved” by the mass media. However, the mass media tends to cover stories that are more likely to go viral or stories that resonate with what we may want to hear down inside, rather than pursue accuracy of depiction. We cannot deny the fact that such tendencies will sway the way we perceive the world, and go on to affect our judgements and decisions to different extends subliminally.

So what if, with the development of artificial intelligence technologies, a proactive machine agent is able to complement our current information intake with a broader context and alternating views? What if a machine agent was able to “read” alongside us, while scanning the sea of information available online, and proactively enlighten us about the biases and tendencies of the content that we are consuming, real time?

Such is the motivation of this exploration, but we must note that it would be unreasonable to expect to achieve a “perfectly accurate” perception of our surrounding world through a machine agent, as there will always be issues in our society that are controversial and be highly dependent on point of view. Rather, the hope of this exploration is to think about machine agents that can proactively assist us in broadening our horizons and heightening our awareness.

 

Background Work – Biased information, motivations for bias, and implications of consuming biased information

         There are two important psychological models regarding information processing and attitude change that are particularly relevant to our exploration. One is the heuristic-systematic model of information processing, developed by Shelly Chaiken in the early 80s, and the other one is the elaboration likelihood model of persuasion developed by Richard Petty and John Cacioppo in the mid 80s.

Systematic processing is processing of information that involves comprehensive and analytic cognitive processing of information. Heuristic processing, on the other hand, makes use of knowledge structures preconceived as judgement criteria. For example, opinions such as “experts can be trusted”, “what the majority believes in should be more correct”, or “longer articles should be more comprehensive and accurate” can all be foundations of heuristic processing. It is quite obvious that systematic processing of information is much better if accuracy of judgement and perception were of single priority, but in reality, we do not always have the mental energy to scrutinize every single issue, and thus heuristic processing still plays a big role in many situations. For example, you may meet a young gentleman at a cocktail party, and after hearing that the young man studied physics at MIT, most people will come away thinking that this person must be smart, even though deep down inside, we all know that such deductions from this piece of information are highly irrational.

The elaboration likelihood model explains the motivations behind our choices for using a more analytical and deliberate approach to processing information versus a more effortless approach based on previous heuristics. We are more likely to pursue an analytical and deliberate approach when there is sufficient motivation and ability, and absence of either will lead us to pursue a less critical approach.

The reality happens to be that most of the time, when we are casually consuming media content, there is no strong motivation tied to our interests to think deliberately, and it is also hard for us to assess information spanning a wide range of topics that we do not possess expertise in. Thus, this means that we are particularly susceptible to being unconsciously affected by biased information.

Furthermore, from the media perspective, as illustrated by Bernhardt in the Journal of Public Economics [1], in the political sphere, there is motivation for media outlets to maximize their profits (reader base) by suppressing information that the audience may not want to hear and promoting those that are more welcome. We can imagine that in sectors other than politics, such effects probably exist too. This means that there is possibly an undesired “positive feedback effect” where biased information misinforms our citizens and the stereotypes and heuristics that arise from such biased information will then motivate media outlets to cater to such archetypes and reinforce this process further.

So what are the undesirable effects for media bias, is there motivation for us to try and mitigate this effect? As portrayed, if somewhat extremely in Entman’s article[2], media biases have very long-term implications on public framing, priming, and agenda-setting, thus greatly affecting the influence of different schools of opinions, and this, in the political sector, can have great impact on power dynamics. Bernhardt’s article mentioned above also holds a similar belief where bias leads to polarization and also electoral mistakes when severe.

Additionally, as Shelly mentions in her work[3],  systematic processing of information is related to heuristic processing, as the notion of information source credibility is important in our logical assessment of information, but such notions of credibility is, to different degrees, based on heuristics. Therefore, there is also the implication that biased information will affect our deliberate inspections of information in the long term by affecting our heuristic processing of information.

Given these implications of biased information, we propose “Shoulder Angel” as a machine augmentation that assists in your information processing, allowing you to exercise a more “systematic processing” of information when you are only browsing information casually and thus are more heuristic and casual in judgement.

 

Related Work

Digital assistants have played a role in automating customer service queries for quite a time, sorting out calls for the likes of banks and airlines. More recently, personal assistants and digital chatbots have become a hot field, with Google, Microsoft, Apple, Amazon, and Facebook all releasing personalized digital assistants of their own.

While these new digital assistants are indeed designed for consumers, and meant to provide as natural a form of interaction as possible, they are still mostly passive agents that cater to your specific wishes such as scheduling a meeting or playing a song.

The exception is Google Now, already available on most Android phones on the leftmost screen, which does proactively present you articles and information such as scores for basketball matches depending on your past search history.

However, what we are hoping for is a proactive agent that compliments your current information influx with different viewpoints and a broader context, not just bolster it with similar content.

    Picture1  Picture2

 

Design and Usage Scenario

       We’ve considered many different formats of information expression of our “Shoulder Angel” from infographic charts to index-like numbers to plain text, but in the end we’ve come to believe that plain text may actually be the best choice.

While graphs and charts offer more information bandwidth, and are able to depict information accurately, it is also true that there is a certain learning curve involved in understanding infographics. Likewise, index-like numbers are favored by professionals in social science areas as a concise and relatively accurate way to grasp situations, but the general public will probably find this format quite unintuitive.

Thus, somewhat following the trend of tech companies moving towards chatbots in their AI efforts nowadays, we chose a plain text, conversational format for our “Shoulder Angel”, this choice was made largely based on pursuits for intuitiveness and a close, “human” feel.

Usage Scenarios:

We highlight a couple of possible usage scenarios below, the first of which demonstrates a scenario when our user has a newfound interest in the debates surrounding the national elections. Shoulder Angel will pick up this trend, and provide feedback about our user’s penchants that he/her may not have been aware of by himself/herself. Shoulder Angel also aims to provide this feedback in a non-rigid, somewhat casual and witty fashion, which should help to make this feedback more humane and easier to relate to.

Picture3

Another usage scenario is when our user is writing an email to a friend/colleague, and shows subliminal cues of strongly biased opinions. Shoulder Angel will whisper to you about your

Picture4

Shoulder Angel will also notify you of biases and stance of media outlets, so that you can be consciously aware of how the emphasis of your content has been intentionally selected to portray a certain biased inclination. This will allow you to be more mentally aware and consciously keep a check on what you read in your mind, real time.

Picture5

Implementation

       The full vision of our Shoulder Angel would require several very difficult technical challenges in natural language processing that are beyond the scope of this design project to realize, and thus we’ve only achieved a very elementary mockup of our final vision.

We had two simple schemes for training our machine to “understand” media content. One was a document summarization model that relied on TextRank[4], a graph-based model for ranking the most important sentences in a document, thus achieving a “summarization” of the document. The second was a Latent Dirichlet Allocation[5] over the content to model the topic distribution, so that our machine can “understand” what the key topics are in the content we are consuming.

These primitive levels of “understanding”, particularly the latter, will allow for a limited number of pre-engineered logic based responses, but more natural responses to summaries will need to rely on large amounts of corpus training data.

 

Future Work

Despite difficulties, the realization of our Shoulder Angel concept should be doable with existing state-of-the-art deep learning natural language processing techniques, as products such as Google Now and Siri are achieving astonishing levels of conversational capability, thus we would like to dive deeper into more sophisticated models of implementation with larger datasets in the future to realize a fuller version of our Shoulder Angel concept.

 

Citations

[1] Political polarization and the electoral effects of media bias, Bernhdart, 2008, http://www.sciencedirect.com/science/article/pii/S0047272708000236

[2] Framing Bias: Media in the Distribution of Power, Entman, 2007 http://onlinelibrary.wiley.com/doi/10.1111/j.1460-2466.2006.00336.x/full

[3] Heuristic processing can bias systematic processing: Effects of source credibility, argument ambiguity, and task importance on attitude judgment, Chaiken, Shelly, 1994 http://search.proquest.com/docview/614307529?accountid=12492

[4] TextRank: Bringing Order into Texts, Mihalcea,Tarau, 2004 https://web.eecs.umich.edu/~mihalcea/papers/mihalcea.emnlp04.pdf

[5] Latent dirichlet allocation, Andrew Ng, Michael Jordan, 2003, http://dl.acm.org/citation.cfm?id=944937