Table 3:
Local perplexities at repetition DF positions.
A trigram model with DF modeling for repetitions was trained and tested as described above. Table 3 shows word perplexities for positions relative to repetition events. REP1 refers to the second instance of a repeated word in a one-word repetition; REP1+1 and REP1+2 denote the first and second word, respectively, after such a repetition. REP2 and REP2+1 refer to the repeated words in a two-word repetition; REP2+2 denotes the word following a two-word repetition. Unlike the case of FPs, the Cleanup Model is generally beneficial in REP contexts, reducing the joint perplexity (all the above positions relative to REP) from 85.9 to 76.6.
We also tested whether the words following the repetition might be better predicted by the REP event itself, rather than the actual words being repeated, analogous to what we found for filled pauses; this turned out not to be the case.
Table 4:
Local perplexities at deletion DF positions.