Artificial Intelligence (AI) has been around for decades and for some, just the mention of AI makes them think of sci-fi movies where robots take over the world. That image represents a type of Artificial General Intelligence (AGI) that does not exist yet. Less advanced versions of AI (aka, “Narrow AI”) do exist and are already being applied to humanitarian relief and development.
At present, when we talk about AI we’re largely talking about supervised learning (or “deep learning”) where we train machines to identify objects and “learn” through labeling data. When the computer detects that object or data pattern again, it will associate it with the label it has been assigned. This type of Narrow AI allows for business applications to become good at reasoning and problem solving or for automating tasks of low complexity1.
Pervasive data from 5G networks, the coming Internet of Things (IoT), and increases in computing power are all factors fueling the growth of AI technologies and inspiring new products and services. Andrew Ng, an author and computer scientist, refers to AI as, “The new electricity: Just as electricity transformed numerous industries starting 100 years ago, AI is now poised to do the same.”2
Recognizing the potential benefit AI may hold for the humanitarian and development space, the Mercy Corps Technology for Development (T4D) team, began an exploration of different AI technologies this past year. We ran three pilot programs with the goal of identifying and prioritizing areas meriting further exploration. The three use cases we chose to pilot were sentiment analysis, translation, and AI learning tools that can help users recognize hate speech on social media platforms3.
On this site across the next few weeks, I will publish a series of articles that look at the three AI pilots from this last year as well as the lessons we have learned and will be leveraging going forward. In this first installment, let’s look at our work with the Google Assistant in the context of translation.
Case Study 1: Google Assistant Interpreter Mode
In Spring 2019, Mercy Corps kicked off a pilot program to explore machine translation at our Mercy Corps Northwest4 (MCNW) office in Portland, Oregon. Mercy Corps Northwest invests in communities, uncovers breakthrough solutions and fights for equitable opportunities across the Pacific Northwest. MCNW’s programs support up to 2,100 small business owners each year; reduce prison re-entry rates for participants by 50%; improve services at more than 30 local nonprofits; and support under-resourced communities with disaster preparedness knowledge. MCNW also works with immigrant and refugee-owned small businesses to provide support, including training, Individual Development Accounts (IDAs), and small grants. The goal of the pilot was to enhance the services offered by MCNW by providing translation for the program participants. Google Home devices with the Google Assistant Interpreter Mode enabled were selected to help the MCNW staff focus on better communication with the diverse populations they serve.
Viable translation is a service Mercy Corps’ employees require on a regular basis. It is often incredibly expensive, time consuming, and sometimes difficult to find, given the diversity of the populations with which we work. The MCNW pilot gave us a chance to explore the possibilities for AI to support translation efforts, and prioritize testing software in development that supports these activities.
MCNW installed two Google Home devices in the MCNW office. Those devices were used by MCNW small business clients for a period of 5 months, ending in March 2020 when COVID-19 began. The initial trial device supported only Spanish and Russian, but MCNW prioritized Arabic support as critical for their clients. With Google’s help, we were able to quickly add this language to the device. The approach proved not only workable, but popular with the users.
By using translation, via AI on device, MCNW’s employees and clients did not have to rely on hiring a translator and incurring extra costs and scheduling effort. The devices also provided a level of privacy and independence that wasn’t possible when relying on a live translator. The cost and the deployment effort required were quite favorable. Users were consistently happy with how easy it was to use. All 30 users provided positive feedback on both their experience and the tool’s effectiveness at managing detailed conversations. The user experience is ambient and non-intrusive, which is helpful in facilitating natural conversations.
This pilot has built momentum for both Mercy Corps and Google, who worked together to collaborate on features and design. Mercy Corps isn’t providing volumes of data, but the subtler user interactions, and design for specific field use cases, has been quite helpful for Google. Mercy Corps’ work helped provide a guidepost for why these tools are relevant, and how users would prefer to use them.
The future Google roadmap will include greater server-side privacy and protections, such as GDPR compliance, so that other markets can take advantage of the translation devices. The roadmap also includes offline features — a key development for the other contexts in which Mercy Corps works. Future project work could include continuing our partnership with Google (and their partners) to test the tool in more diverse contexts in the field.
Overall, T4D sees great value in growing the use of this translation technology. This pilot emphasized the power of committing to user centered design and the utility of simplified AI technology applications. As this tool develops, particularly as it includes offline functionality and an increasing number of languages, it has significant implications for representation and inclusion among our target populations, in Portland and abroad.
The Technology for Development (T4D) team supported the Mercy Corps Northwest Team in the delivery of this pilot through generous partnership with Cisco, under a 5-year program aimed at using technology to deliver aid and development assistance faster, better, and to more people.
Note: This article is excerpted and expanded from work originally done by Karen Smetana during her tenure at Mercy Corps.
[1] By way of example, look at the Google predictive typing feature in Gmail, or Amazon’s shopping suggestions.
[2] See, https://www.gsb.stanford.edu/insights/andrew-ng-why-ai-new-electricity
[3] We purposefully decided to pilot AI prototypes that could get us quick learnings, not necessarily large AI platforms that require massive data, large formatting or collection efforts, or anything with a significant cost.