Walking the Graph of Language: On a Framework for Meaning and Analogy
Nal Emmerich Kalchbrenner

Abstract:
We introduce a computational framework for generating representations
of linguistic concepts. The concepts we consider are the meanings of
words and the verbal analogs corresponding to n-tuples of
words. Representations of meanings can be compared to estimate their
degree of synonymy. Likewise, representations of verbal analogs can be
compared to estimate the strength of the analogy between them. The
framework automatically constructs from a corpus of language large
graphs with words as vertices and conceptual connections as edges;
these graphs are dubbed word-graphs. Focusing on representations of
verbal analogs of word pairs, we present two main algorithms for the
extraction of such representations from a word-graph. One algorithm
relies on path distance measures and random walks over the
word-graph. The other algorithm relies on spreading activation and
algebraic vector operations. Tested on a standardized set of verbal
analogy problems, one of the algorithms attains accuracy that is
statistically not significantly different from the
state-of-the-art. Further, the experiments yield a novel theoretical
insight into the workings of verbal analogy and its representation.