Richard Socher

Socher received a computer-science PhD from Stanford in 2014 and later taught there as an adjunct professor. He led MetaMind as CEO and CTO before its acquisition by Salesforce in 2016. At Salesforce he served as chief scientist and executive vice president, overseeing research and applied AI. His current biography identifies leadership at Recursive, which pursues automated knowledge discovery through self-improving AI, and You.com, which supplies web-search APIs to language models. He also founded and manages the investment firm AIX Ventures.[1]

Learn meaning at the phrase level

The 2013 Sentiment Treebank paper, on which Socher is first author, addresses a limitation of representing words independently: combining words can reverse or qualify a sentence’s sentiment. The authors labeled phrases inside sentence parse trees and trained a recursive neural tensor network to combine smaller representations into larger ones. This made it possible to test whether a model handled negation and phrase composition, rather than only recognizing positive and negative words. The reported gains concern the paper’s sentiment benchmarks.[3]

Use global word statistics to learn vectors

GloVe, coauthored with Jeffrey Pennington and Christopher Manning, learns numerical representations of words from how often they occur near other words across a corpus. Its training objective fits information in the word co-occurrence matrix so related statistical patterns correspond to structure in the vector space. Those vectors became reusable inputs to language-processing systems. The paper evaluates analogies and downstream tasks; it does not establish that a vector captures every aspect of a word’s meaning or a contextual interpretation.[2]