Appendix: Tagsets

The full initial 37 tagset set found from the WSJ is given in table 6. The best tagset, shown in table 7, is formed by collapsing these tags into 23 tags. As ex, fw and 2 are typically unreliably predicted, and quite rare, they are not included in the implementation released with Festival, and POS tags of this type, if predicted, are treated as part of the nn_nnp_nnps_nns group. Although this marginally reduces the accuracy for our test set it reduces the size of models and hence seems worthwhile in a run-time system.

Table 6: The original WSJ tagset

Tag	Function
CC	coordinating conjunction
CD	cardinal number
DT	determiner
EX	existential ``there''
FW	foreign word
IN	preposition
JJ	adjective
JJR	adjective, comparative
JJS	adjective, superlative
MD	modal
NN	non-plural common noun
NNP	non-plural proper noun
NNPS	plural proper noun
NNS	plural common noun
of	the word ``of''
PDT	pre-determiner
POS	posessive
PRP	pronoun
puncf	final punctuation (period,
	question mark and exclamation mark)
punc	other punction
hline RB	adverb
RBR	adverb, comparative
RBS	adverb, superlative
RP	particle
TO	the word ``to''
UH	interjection
VB	verb, base form
VBD	verb, past tense
VBG	verb, gerund or present participle
VBN	verb, past participle
VBP	verb, non-3rd person
VBZ	verb, 3rd person
WDT	wh-determiner
WP	wh-pronoun
WRB	wh-adverb
sym	symbol
2	ambiguously labelled

Table 7: The best clusterd tagset. The names directly show the clustering of the original WSJ set.

Tag

cc

cd

dt

ex

fw

in

jj_jjr_jjs

md

nn_nnp_nnps_nns

of

pdt

pos

prp

punc_puncf

rb_rbr_rbs_rp

to

uh

vb_vbd_vbg_vbn_vbp_vbz

wdt

wp

wrb

sym

2