When you’re a multinational tech kaiju spending billions on research and development per year, it’s only natural to think of some creative ways to get more bang for your buck. Google, which expensed some $16.6 billion (15% of net revenues) on R&D in 2017, has been a leader in harnessing the wisdom and manpower of crowds to achieve more cost-efficient research outcomes. The company’s engineers and managers have demonstrated exceptional flexibility in tailoring the structures of crowd engagements to incentivize optimal participation for specific projects, such as open competition for a large bounty (the Little Box Challenge) or harvesting location data from a passive crowd of providers (Google Traffic, Parking, Trips, various business data, and so on.)
In 2011, when Google wanted to expand Voice Search to new languages, international program managers enlisted the help of local users in several countries for the Word of Mouth Project. “Really, we couldn’t have done it without them,” wrote Linne Ha, one of the project managers, in her notes on the experience. Rather than paying to license speech and text data from an external repository, Google realized that better and more robust data sets could be obtained directly from users. Using phones provided by Google, hundreds of volunteers recorded voice samples alongside their friends and family, and then later assisted in beta testing as the launch window approached. With thousands of hours of raw data, Google was able to build a world-class, fully localized product that was dynamic enough to interpret regional dialects and idiomatic speech.
Incentives and market design
Google’s brand, built on its mission to provide “access and technology for everyone,” is now one of the most recognized in the world. The program managers were able to leverage the brand to recruit volunteers who were simply big fans of Google and eager to contribute their time for the intangible satisfaction of assisting the company. While it’s likely that many participants were also motivated by the perceived utility of accessing localized Voice Search in the future, there is no evidence to suggest Google offered financial incentives beyond a free phone for data recording.
With a large pool of motivated volunteers to choose from, Google was able to bypass the typical market design challenges that commonly appear in crowdsourcing efforts. Researchers could handpick the users they perceived to be highest in quality and commitment. As no financial rewards were involved, there was no incentive for a mass of low quality users to enter the labor pool and bid down the price for services. Instead, the participants were primarily self-selecting, with an intrinsic motivation that dovetailed with that of Google’s research managers: to help build a highly functional product.
Task design and management
The success of the Word of Mouth project depended on the collection of linguistic data of sufficient quantity and quality to train Google’s machine learning models. As non-expert (crowdsourced) contributors are often likely to generate a higher degree of error or variance in their work than experts or professionals, there was a risk that a portion of the data would not be useful (resulting in wasted man-hours) or even serve to corrupt the data pool and negate the effects of value-added work. To overcome these challenges, Google applied a number of techniques to produce a positive outcome:
- Task decomposition. In order to reduce cognitive load and limit user error, Google segmented language input operations into the smallest and simplest steps possible. Researchers relied on instructing users with examples, finding that it was easier for participants to imitate examples than to interpret and follow complicated guidelines.
- UI simplification. Google knew that a streamlined user interface would reduce mental stress on users, allowing them to devote more attention to the task at hand. Ample testing before roll-out served to sustain clean and focused interactions between the users and collection platform.
- Responsive feedback. Adequate supervision and feedback served to limit noisy data from entering the data pool and allowed contributors to flag possible failure conditions in the collection systems. Google researchers solicited quantitative and free-form feedback through the UI as well as offline, communicating extensively with the users. This reactive attention to the users also served to validate their participation and sustain continued engagement.
Value creation, capture, and growth
As most people are aware, Google’s business model depends primarily on keeping users on web and application portals under Google’s sphere of control and influence for the delivery of targeted performance advertising. Products such as localized Voice Search have allowed Google to increase user engagement and ad delivery in many new markets, making it one of the most valuable and profitable companies in the world. Additionally, the speech recognition technology advanced through the Word of Mouth project and its successors has been a critical component of the next wave of products in the Google ecoystem such as Google Assistant. While the methodology used for Word of Mouth is just one expression of Google’s multistrategy approach to crowdsourcing at the project level, it is a great example of the company’s ingenuity in leveraging crowd support to execute cost-efficient research at scale.