THE HIVE THINK TANK: DEEP LEARNING AND NLP BY SALESFORCE, LINKEDIN, BAIDU, AND VENTUREBEAT
November 15, 2016 – By Travis Miller
From left to right: Carl Case (Baidu), Dan Bikel (LinkedIn), Richard Socher (Salesforce), Jordan Novet (VentureBeat), TM Ravi (The Hive)
The Information Age is, by nature, immersed in the complexities of linguistics. Be it the Google search bar, New York Times notifications, or Siri’s home button, language-driven human-computer interaction is moving our world of data forward. Amidst this connectivity, improving a machine’s ability to interpret human languages evokes promises of better services and new products.
A machine’s understanding of complex language structures is driven by the concepts and practices of natural language processing (NLP). In recent years, NLP has drastically improved through deep learning algorithms, as interpretation tasks are impelled by computers leveraging large amounts of data. To provide insight into this forward momentum, The Hive Think Tank hosted a Deep Learning and NLP panel moderated by Jordan Novet of VentureBeat on November 3, 2016. During the discussion, each of the panelists expounded upon their respective recent projects. Richard Socher, Chief Scientist of Salesforce, noted that among several areas of study his team is focusing on multi-task learning. He believes it to be the crucial next step for NLP: building a model that can continually become more intelligent without cycles of re-training. Dan Bikel, Principal NLP Scientist at LinkedIn, shared how his team is using traditional entity linking techniques to realize relationships between LinkedIn entities. These linguistically-driven connections are building the company’s “Economic Graph”, with a vision to elevate the interconnectedness of the working world. Finally, Carl Case, Research Scientist at Baidu Silicon Valley AI Lab, discussed Baidu’s recent deep dive into conversational interfaces. Center stage in their efforts is “Melody”, a deep learning-driven medical chatbot reminiscent of WebMD.
Drawing from their experience in these projects and beyond, the panelists discussed a wide array of topics. They pointed to the various ways in which deep learning is elevating NLP potential, noting that it annuls the need for burdensome feature engineering processes. They compared different data labeling strategies (Is a supervised dataset worth the intensive effort? Are Amazon Mechanical Turk and CrowdFlower valuable means to an end?). Opinions were shared on emerging trends in deep learning, and answers were given to the audience’s diverse set of NLP questions. The panel recording can be viewed here and pictures from the event here.
Learn about our upcoming events on The Hive Think Tank Meetup page.