Contextual Sensing and Sentiment Classification

Editor's Notes

• #2: My name is Adrienne, and I’m from a startup in Seattle called ARO. I’m a scientist at ARO, and I’m part of the team that makes sense of all of the sensor data that can be generated and collected by your smartphone. My background is not in sentiment detection, but in HCI and ubiquitous computing. As a company, we generate customer insight from sensor data, but we’re really excited about the capability and advances in the sentiment detection field, as well as how the kind of context we provide can help sentiment detection.
• #3: Let’s start with an example. This tweet comes in. Let me show you what you’re missing to make sense of this tweet. Linguistically, this is a neutral statement. However, greater context and “common sense” can help humans detect relevant sentiment.
• #4: In addition to the location that is included with the tweet, including social network information (that is usually modeled as group membership), can provide some form of context and improve the ability to make sense of this content.
• #5: But let’s also take into account the day of week and time. By itself, it may add some value. But Tuesday at 8:30 has different meaning to different people.
• #6: One of the things we do is capture and make sense of individuals’ routines, such as where they work, what they do on the way to work, how long someone typically spends at one location, and what might be coming up next.
• #7: In combination with information about an upcoming meeting at work and additional data such as traffic and weather, the speaker is likely more stressed than neutral.
• #8: Let’s consider an alternative set of context. A Saturday, out of routine, a Macklemore concert coming up, and the sentiment is likely more positive than neutral.
• #9: But let’s consider another sentiment classification task. This time, we’re looking at a Yelp review. In this case, a person’s review can be given more or less weight based on where they’ve gone in the past.
• #10: Why is the state of the art insufficient?This is a technical problem: we’re training on insufficient data. The (over-simplified) approach to sentiment detection is counting positive and negative words.
• #11: Context is hard– and it doesn’t come from humans. It comes from the sensors and information streams that people are carrying in their pocket.
• #12: This is ARO’s specialty: Making sense of all the noisy sensor data that’s on our phone. AccelerometerWifiAcousticsLight sensor
• #13: So what do we do with this? How is it available to you? We have consumer apps, an SDK for OEMs, and API, a CIP, and a SS-user data provisioner.
• #14: Other people who are interested in this kind of space include Google, which we see from Google Now; NFL, who is interested in what football fans do on game days, where they watch the game, etc; and Disney, who wants to know more about how people go around the parks so that they can optimize the experience for everyone.
• #15: Saga is alifelogging app. Saga keeps track of where you go, what you do, and learns about you over time. Saga can be a smart assistant, learning about you over time, and providing relevant information when you care about. Saga presents your lifelog back to you, as a beautiful diary of what you’ve done with your life. Saga learns things like where you live, where you work, and a typical commute for you. We can also identify what kinds of places you like to visit, and how this changes over time. While Saga is primarily about providing a tool for individuals to keep track of information about themselves (we never sell this data!), it represents our core capability: capturing and making sense of a person’s context.
• #16: Why do we care about sentiment? For lifelog users, we’d like to be able to include information about whether deviations from routine are positive or negative.
• #17: Now, just to come back to the example we started with: We’d love to help improve sentiment detection by providing access to the context information tha.
• #18: Contact us to talk more!