Title: Implementing Latent Semantic Analysis Functions in R

Author: Fritz Günther

Affiliation: University of Tuebingen

Abstract: Latent Semantic Analysis (LSA, Landauer & Dumais, 1997) belongs to a
class of models for semantics called Distributional Semantic Models. In such a
model, words and documents are represented through vectors in a
high-dimensional vector space, derived from a large corpus of text documents.
This representation allows for a formal computation of the semantic similarity
of words and documents, for example by using the cosine value between two such
vectors.

The presented package comes with three different 300-dimensional LSA spaces
based on blog entries, newspaper articles or literature, each consisting of
about 50,000 documents in German language. It also provides a collections of
functions to perform computations on the basis of these spaces:

For instance, the semantic neighbourhood of a given word and document can be
computed. Such a semantic neighbourhood can also be plotted in a
three-dimensional approximation of this part of the LSA space. Furthermore,
pairwise similarities can be computed for two given vectors of words, as well
as the similarity between two whole text documents and the similarities
between a whole text document and a vector of words. The latter can for
example be used for an automatic chracterization of books using a list of
keywords.

The package also includes functions for the computation of a document's
coherence (adapted from Landauer & Dumais, 1997) and for several computation
methods for vectors of two-word phrases (adapted from Mitchell & Lapata,
2008), including the predication process by Kintsch (2001). Such vectors for
two-word phrases can also be used in the aforementioned functions of the
package.

References:

Kintsch, W. (2001). Predication. Cognitive science, 25, 173-202.

Landauer, T. K., & Dumais, S. T. (1997). A solution to Plato's problem: The
Latent Semantic Analysis theory of acquisition, induction and representation
of knowledge. Psychological Review, 104, 211-240.

Mitchell, J., & Lapata, M. (2008). Vector-based models of semantic
composition. In J. Moore, S. Teufel, J. Aller, & S. Furui (Eds.). Proceedings
of ACL-08: HLT (pp. 236-244). Columbus, OH: ACL Press.