Current distributed representations of words show little resemblance to
theories of lexical semantics. The former are dense and uninterpretable, the
latter largely based on familiar, discrete classes (e.g., supersenses) and
relations (e.g., synonymy and hypernymy). We propose methods t