Nathalie Delfosse, Tsuyoshi Morimoto
Evaluation of metrics for text comparison
with implication for recognition
Abstract:This report describes work done on some texts from the ATR database,
evaluating the relationship between text similarity and speech recognition
accuracy. The work is in two parts: the first part is an application of the
quantification theory IV, using a set of Japanese conversation texts with a
corresponding set of English texts, and testing different definitions for the metric
of the quantification. The correlation between distances is measured, as is the
correlation between Japanese and English texts. Bigrams and trigrams (both a
simplified form of syntax) have strong correlations in Japanese and in English,
but vocabulary and simplified syntax appear only weakly correlated.
In the second part of this report, only Japanese texts are used. Two
probabilistic grammars have been trained on different sets of conversation texts.
For one target text, the recognition rate is higher if the distance to the training
set is smaller. For one training set, the distance of the target texts to this set and
the recognition rate seem not to be strongly correlated.