Mobile messaging platforms and social networking services combined with recent advancements in natural language understanding and processing (NLU/P) have helped create an emerging market for social bots. Today, social bots can do everything from manage your calendar to order you an Uber or provide fashion advice in the course of a conversation. As social bots become intertwined, and integral, within social network services and everyday digital interactions it’s quite likely that conversational social bots will be the first exposure many people have with anything approximating artificial intelligence (AI).
Looking back, the origin of social bots is inexorably rooted in the development and adoption of social networking services and messaging platforms spanning several decades. Social bots found on platforms such as Facebook Messenger, Telegram, and Kik today share distinctive pedigree with party-line Bulletin Board Systems (BBS), CompuServe, and AOL. Slack, one of this era’s most popular “unicorns,” traces its lineage to Internet Relay Chat (IRC) developed nearly 30 years ago. The expansion of internet availability in developing economies, rapid growth of mobile computing, messaging platform proliferation and solidification of social networking services as an always-on, ever-present component of daily life have created a rich ecosystem in which social bots now pervade.
Dozens of companies backed by venture capital are developing technology to enable social bots to become as universal as the messaging platforms on which they are designed to operate. Bolstered by significant investments from the world’s foremost technologically innovative companies and emerging research from leading academic institutions, these startups are poised to alter the medium by which human-machine information transference occurs, thus ushering in a new era of machine-augmented-human capabilities.
One of the most prolific use cases for social bots is automation of what are often mundane, repetitive tasks. There is no shortage of social bot startups seeking to automate tasks such as meeting scheduling (X.ai), business data and information retrieval (butter.ai), expense management (Birdly), and nominating candidates for major elections (LaPrimaire). It is not unrealistic to expect that social bots in the near-term will be capable of supporting inquiries such as, “tell me how many widgets I sent to Region X in the last six months” with relevant and timely data retrieval. In fact, this capability already exists inside many innovative organizations. AT&T[1] has deployed several hundred social bots that automate a variety of mundane tasks including data entry into legacy systems. To be sure, this capability represents a significant potential effect on how humans interact with data, how analytical work is accomplished, and how fast machine-augmented-humans go from question-to-answer.
It’s no secret that brand marketers excel at idea diffusion across myriad mediums. Whereas digital marketing is a relatively recent phenomenon, the idea of shaping consumer behavior is at the very least hundreds of years in the making. Social bots expose rich interaction possibilities between consumers and brands allowing marketers to exploit characteristics of human behavior that would not be possible without a highly interactive, conversation-based engagement. For instance, “Lt. Reyes” from Call of Duty: Infinite Warfare enabled a fairly robust, albeit domain constrained, conversation via Facebook Messenger ahead of the game’s release last year. Another example is Taco Bell’s “TacoBot,” a Slack-based bot that is simultaneously humorous and efficient in converting dialogue to revenue. These examples are important in that brands had never before had access to consumers via such an interactive medium.
Another emerging — significant — use is its ability to influence. Shaping public discourse has existed for approximately as long as humans have had the ability to communicate. At the very least, there exists a rich history of the uses of strategic disinformation and discourse manipulation since ancient Greece. What has not been possible until very recently, however, is the ability to automate a virtually anonymous, global, coordinated influence operation targeting something on the order of 300 million people, simultaneously (Twitter boasts 300+M monthly active users) with only a handful of human actors. Further, never before has it been possible to leverage computationally intensive methods for mathematical exploitation of social networks to maximize information diffusion at the cost of less than a marginally reliable used car. In short, in less than a decade, machines have become the most capable agents of influence-shaping ever employed.
While social bots may be “conversing”, there is substantial variation in their conversational ability predicated largely on their intended purpose. Commerce-oriented social bots may have a very narrow topic domain to consider and no requirement for “small talk” or exhibition of specific personality traits, such as the Nike shoe bot. Conversely, persona-based social bots used in marketing a new movie, video game, or toy may be required to navigate several topical domains like gameplay, military vernacular, and fashion while maintaining a cohesive personality during dialogue exchange.
The Making of a Bot
At the core, a conversational social bot is a representation of dialogue structure subjected to conversational regulation governing how an interaction begins, proceeds, and ultimately concludes. Depending on the requisite breadth and depth of the dialogue structure, various mathematical approaches to achieving conversational functionality can be employed, often together. Dialogue model choices coupled with the social bot use case directly influence computational complexity, with current research focused on maximizing the effectiveness of various combinations of neural networks (“deep learning”) and additional methods, the most popular of which include:
- Recurrent Neural Networks (RNN): A flavor of neural network well-suited for working with arbitrary sequence data (i.e. dialogue text) because it possesses internal “memory” enabling statefulness. Whereas a feed-forward neural network (such as a Convolutional Neural Network (CNN)) can manage only a fixed number of computational steps, in a specific order, RNNs are quite a bit more flexible and make use of previous computational steps on the way to generating an output. Their ability to “remember” recent information and utilize it in a current computational task makes them useful components of a conversational bot toolkit.
- Long-Short Term Memory (LSTM): Think of LSTM as an RNN with longer-term memory. Implementations of LSTM are virtually innumerable, with several variations having likely been published while you were reading this Technology Brief. LSTM utilizes a “gated cell” that enables it to make decisions about what inputs to store, forget and eventually output. This capability proves to be very useful in interactions with the types of dialogue structure used in most conversational social bots where the objective is to generate a response (output) from a learned corpus of dialogue (context).
- Graph Propagation: Semi-supervised learning (SSL) techniques are attractive for a broad range of use cases because they require relatively small volumes of labeled observations. A subclass of SSL is graph-based, which scales well to large datasets and high-dimension problem spaces (i.e. language corpora). What’s more, graph-based SSL is particularly well-suited for label propagation leveraging traversals of the standard nodes/edges/vertices which is great for applications such as semantic cluster label propagation.
The Bots’ Next Steps
Undoubtedly, social bots have an important role in the next generation of human/machine interface. Their utility is multiplying and now is the time to begin considering their current and future role inside your organization. A decade ago it was bleeding-edge if you could order a pizza from a website and now you can do it using the pizza emoji over SMS or by telling your Amazon Echo to do it — what role would the same type of task automation have in your team? As the conversational quality of social bots becomes increasingly indistinguishable from human agents in closed domains what capabilities does that introduce that have never before existed? And, perhaps most interesting to consider, what do we do when we are no longer reasonably sure we’re conversing via text with a human? What happens when diffusion of an idea or the shape of discourse is merely a function a social botnet is deployed to maximize? With enormous investments in social AI from the world’s leading technology companies, venture capital firms, and research institutions there’s little doubt that we will find out very soon.
[1] Rachel King, “AT&T Employees Automate Repetitive Tasks with Software Bots,” The Wall Street Journal, May 15, 2016. Online version: http://blogs.wsj.com/cio/2016/05/15/att-employees-automate-repetitive-tasks-with-software-bots/