A deletion-only DF model was trained on 1.4 million words of DF-annotated
transcripts.
In order to perform an analysis by DF type and position,
the models were tested on 17,800 words of similarly annotated data.
Only sentence and one-word deletions are reported in
Table 4,
since the test data contained only a single two-word deletion.
The second row for ``DEL model'' gives the perplexity based only on
the word following a deletion, without including the probability for the
deletion event itself.
This shows that the context modification has the
intended effect of making the next word more likely on average.