Vanguard Magazine

T TECHNOLOGY 36 AUGUST/SEPTEMBER 2015 www.vanguardcanada.com Seated by the laptop at the front of the room was Pierre Isa- belle, Research Officer, Multilingual Text Processing and to my right, Andrew Scheidl, Multimedia Analytic Tools for Security Program Lead. Scheidl opens with an explanation of NRC. "We're really about bringing science to improve the lives of Ca- nadians. We do that in part by inventing helpful technologies, but also by developing and transitioning the right ones to help Canadian companies gain competitive advantage." And today, they've invited us to experience a fascinating set of those technologies. About two years ago, Isabelle explains, they began thinking about how to apply some of the language tech- nologies they had created for other sectors to the defence and security domains. "In that respect, the notion of Big Data was already very vis- ible, very important," Isabelle explains. "We thought maybe people needed to examine social media. Social media ooze text, and this is what we're good with: sorting text, organizing text, translating text, searching text for sentiment and emotions and extracting relevant information." With that in mind, Isabelle decided to submit a project pro- posal about mining social media using the latest of the NRC's natural language technologies. Isabelle is proud of the NRC's accomplishments in this domain, referring to it as "world class." What followed was a collaborative project within the Canadi- an Safety and Security program, undertaken jointly between the NRC, Thales Canada, and MediaMiser. (The Centre for Security Science itself is a collaborative office of Defence Research and De- velopment Canada (DRDC) and Public Safety). Scheidl informs me that the project is actually just wrapping up. "This is a two-year project," he says. "And it ends this month, officially. It's been a great experience for our researchers, drawing together a number of our state-of-the-art technologies in a single platform." In order to understand the project, officially known as "Counter- ing Security Threats using Natural Language Technology," Isabelle begins by forming a picture in your mind. Imagine massive amounts of rapidly-flowing, publically available information from Facebook, Twitter, blogs and newsfeeds gushing through gigantic pipes. Using existing software supplied by MediaMiser, an Ottawa- based media monitoring company, Isabelle is able to filter out potentially relevant text as it flows. "Different datasets get extracted from the web, and that's a decision we make early because we can't collect the whole pipe. It's much too big," he says. It's at this point the technologies Isabelle and his team have created come into play. As the relevant text comes in, it gets ma- chine translated (if necessary), summarized, and interesting enti- ties like people, places and organizations are detected. But that's not all. Probably the most interesting component of this process is that the text is automatically analyzed for sentiment. "By senti- ment," Isabelle begins, "we mean, is the attitude of the author negative, positive, or neutral? As the information passes — on the fly — computers tag it with the appropriate human emotion. "We attach real names of emo- tions to text, or sentences in the text, like, 'This is expressing an- ger. This is expressing fear.' That sort of thing," Isabelle says. At that moment, it isn't readily apparent just how powerful a tool that can be. But Isabelle isn't finished explaining the process. After the texts are "annotated" they finally end up in a special purpose database at Thales Canada's research facility in Quebec where the information is indexed in order to become searchable. Thales Canada provides the visualizations of the analyzed data we are shown. "Each of these technologies are already proven," Isabelle admits, "so the goal of this project is not to develop them, but merely to show their usefulness in the context of pulling out information for security purposes." Thales' web-based interface allows you to search through that information the same way you would do a Google search… except that this search yields some pretty powerful insights. It's here, in this room within building M-50, that I find myself looking out at the frontier of big data analytics. The view is fasci- nating, but also a little frightening at the same time. Isabelle motions to the screen in front of me. It's a projected view of his laptop monitor displaying a web-based interface, the program developed to search through the annotated data. I jump ahead, asking him if he could use it to find a single person — someone like Parliament shooter Michael Zehaf-Bibeau, or Mar- tin Couture-Rouleau — who was planning on carrying out a ter- rorist attack. The simple answer is that in the future we will become increas- ingly good at detecting threats ahead of time, even when posters try to hide in the mass. For example, even though social media us- ers often turn off the geolocation function on their mobile devic- es, it is now possible to geolocalize their posts within an average distance of eight kilometers. "We could probably do even better than that" said Isabelle. With the interface, users can perform fine-grain queries on cer- tain assets, precisely narrowing down the things that are of inter- est. "The collections can be queried on all sorts of dimensions," Isabelle says. "Ordinary words, series of words, hashtags, authors, places...a given time period or specific location." For example, users could type in something like this: "Give me all the text that mentions the Pan Am Games with negative emo- tions." The results would yield a trove of information that can be re- stricted to specific dates and/or geographic location. Therefore, the system could be used to figure out how Torontonians feel about the Pan Am Games in the week leading up to the event, or alternatively, to find those people who may pose a potential threat. The NRC team makes clear that for this technology demonstra- tor, the focus is on foreign events. Isabelle brings up a "card" containing information they've been Even though social media users oen turn o the geolocation function on their mobile devices, it is now possible to geolocalize their posts within an average distance of eight kilometers.

• Cover