A PARSER COPING WITH SELF-REPAIRED JAPANESE 
UTTERANCES AND LARGE CORPUS-BASED EVALUATION 
Yuji Sagawa Noboru Ohnishi Noboru Sugie 
Dept. of Information Engineering, Nagoya University, Japan 
INTRODUCTION 
Self-repair(Levelt 1988) is a repair of ut- 
terance by speaker him/herself. A truman 
speaker makes self-repairs very frequently in 
spontaneous speedt. (Blackmer and Mit- 
ton 1991) reported that self-repairs are made 
once every 4.8 seconds in dialogues taken 
fi'om radio talk shows. 
Self-repair is one ldnd of "permissible ill- 
formedness", that is a human listener can 
feel ill-formedness in it hut he/she is able to 
recognize its intended meaning. Thus your 
partner does not need to interrupt dialogue. 
Itow do you feel if your partner interrupts 
dialogue every 5 seconds to ask "What do you 
mean?" or so? You will give up dialogue or 
choose means of writing. Speaking without 
self-repair is the most difficult modality of 
natural language communication. 
The goal of our work is to make a di- 
alogue system coping with self-repaired ut- 
terances. In this paper we propose a 
parser called SERUP(SElf-Repaired Utter- 
ance Parser), which plays a major part in un- 
derstanding a self-repaired utterance. That 
is, because our approach is to translate a 
self-repaired utterance (Ex.1) into a well- 
formed version that does not contain self- 
repair (Ex.2) and parse the well-formed one, 
we do not need to change the subsequent pro- 
cesses. 
\[Ex.1\] And fi'om green left to pink, 
er, from blue left to pink (from 
(Levelt 1988)) 
\[Ex.2\] And fi'om blue left to pink 
SERUP uses some linguistic clues to trans- 
late utterances, those include a repetition, an 
unknown word and/or an isolated word. We 
describe how SERUP uses these clues. 
To evaluate SERUP, we analyze a large 
corpus that contains spontaneous dialogues 
over telephone. From the result, we estimate 
that SI';RUP works well with 88.1% of 1,082 
self-repairs in the corpus. 
I~ELATED WORKS 
(Hindle 1983) and (Langer 1990) proposed 
parsers coping with self-repaired utterances. 
But they assumed that an interruption point 
has already been detected. Hindle thought 
prosodic cues carl be used in detection, but 
it is not clear if they can always succeed. 
Langer thought editing expressions can be 
used, but they are not always used in self- 
repair. 
Recently, (Shriberg, Bear, and l)owdlng 
1992) proposed a pattern matching method 
and used it ill GEMINI system(Dowding 
et al. 1993). Tills is similar to our method, 
but the corpus(MADCOW 1992) used is less 
spontarleous than ours. (Subjects pressed a 
button to begin speaking to the system) 
(Nakatani and Hirschberg 1993) proposed 
a speech-first method in which prosodic cues 
are used mainly. We also think prosodic cues 
are important. But wc think people use lin- 
guistic cues mainly because they can under- 
stand self-repaired utterances in transcripts. 
All these works are done on English. 
(Langer also treats Germany) Because there 
are many syntactic differences (e.g., left 
l)ranching v.s. right branching), it is not 
593 
clear if their approach is applicable 
Japanese. 
OUTLINE OF SEt,UP 
to 
Fig.1 shows the outline of SERUP. 
Normal Parser is a parser that parses well- 
formed utterances. When Normal Parser 
fails to parse an utterance, the utterance is 
passed to SR-reconstructor that detects a 
self-repair in it and translates it into well- 
formed version. The translated utterance is 
returned to Normal Parser and parsed again. 
Because an utterance can contain two or 
more self-repairs, translation is repeated un- 
til Normal Parser succeeds in parsing or 
translation fails. In the latter case, the utter- 
ance has another ill-formedness or self-repair 
that the SR-reconstructor cannot cope with. 
There are two main problems in trans- 
lation. One is to det, errnine an interrup- 
tion point, and the other is to determine a 
reparandum. If these two problems can be 
solved, then the process of translation is car- 
ried out as follows. 
1. Remove editing expressions such as er, 
rio, I mean. 
2. Supersede the reparandum with repair 
part. 
For more detail of SERUP, see (Sagawa, 
Ohnishi, and Sugie 1993). 
CLUES TO TRANSLATION 
In this dlapter, we will describe a classifi- 
cation of self-repaired utterances. They are 
Inputs 
..... v 1¢ ...... 
ISa.rec:?nstructor ~ -- 
L ..... 
tree "fail" 
Fig. 1: The outline of SERUP 
classified by clues usable to determine an in- 
terruption point and a reparandum. 
Table 1 shows the classification. Cate- 
gories printed in italics have no clue, i.e., 
SERUP fails to parse utterances in those cat- 
egories. 
with repetition 
A self-repair is mostly made in a way to re- 
pair a word or a phrase just before an inter- 
ruption(Levelt 1988). So words or phrases 
around an interruption are in the same cate- 
gory. For example, in \[Ex.l\] speaker repairs a 
prepositional phrase "from green left to pink" 
to "from blue left to pink", It is rare that 
he/she just repairs a noun "green" to "blue". 
In such self-repairs, a repetition of a word 
or a phrase often exists. In self-repairs which 
are intended to correct an error (such as 
\[Ex.1\]), words or phrases around the error 
may be repeated.. 
In \[gx.1\], "from" and "left to pink" are re- 
peated. In sell-repairs which are intended to 
add some information to the item just men- 
tioned, the item may be repeated as in \[Ex.a\]. 
\[Ex.3\] \[ want a fight, one way flight 
(from (Shrlberg, Bear, and l)owd- 
ing 1992)) 
ILl this example a word "flight" is repeated. 
A repetition is made with the same con- 
stituent or an item in tile same category, such 
as "orange" with "apl)le". 
There are four possible structures around 
an int, errupl, ion of a self-repa.ir with a repet;i- 
tion. l"ig.2 shows them. 
reparundum 
R X R g 
.............................. J ............................... 
n: ropotilion 
Fig. 2: Possible structures around interrup- 
tion 
594 
A is a case of a simple repetition. B, C 
and l) are cases in which some words ex- 
ist between repetition. With cases II and 
C, positions of repetition directory indicate 
where an interruption occurs and which is a 
reparandum, but with D case, do not. 
SERUI ~ can cope with cases A, tt and C. 
with syntactic break 
A self-repalr comes with an interruption of 
utterance. Because an interruption may oc- 
cur anywhere in an utterance (even within a 
word), self-repaired utterance can contain a 
syntactic break. 
If this 1)reak can be detected, we can iden- 
tify an interruption point. 
same fl'agment repetition 
When a sl)eaker interrupts an utterance 
within a word, a fl'agment of the interrupted 
word is left. But he/she sometimes starts the 
repair with a word that begins with the slune 
fragment as in \[Ex.4\]. 
\[F,x.4\] ten, tenji tanntou ,m kata t,o 
'.l'his can be treated as A repetition, but 
to investigate a within-word interruption~ we 
treated it as a separate category. 
In this case, an interruption point is just 
after a repeated fragment. And if within- 
word interruptions are only made to repair 
an interrupted word, a rtq)armldum can be 
identified as the repeated fragment 
with unknown word 
Sometimes a fi'agment left clm be detected 
as an unknown word. For example, if a word 
"ketueki(blood)" is interrul)ted and a fi'ag- 
meat "ketue" is left, this fragment (:an be 
detected because there is no Japanese word 
"ketue'. 
In this case, an interruption point is just 
after an unknown word. And the repa.ran- 
dum can be determined if the same condition 
as the above case is sufficed. 
with isolated word 
A fragment left by a within-wor.'l interrup- 
tion is not alwa.ys detected as the same fl:a,g~ 
meat repetition or an unknown word. For 
example, it fragment "hen" can be left when 
"hontou"(real) is interrupted, but this string 
can l)e a wor(1 meaning "book". 
But such a word is always "isolated", tlutt 
is, both two subtree.s in fig'l fail. 
In this (:as(.', an interruption point is just 
after an isolated word. An(l repara.ndum can 
be determined if the same condition as the 
above ('ase is sufficed. 
without repetition of a stem 
Because Japanese inflectional morphology 
is complicated, speakers often make inflection 
errors. To rel)alr such errors a speaker often 
starts a relmi," without repetition of a stem 
as i,, \[l':x..q ,,or as i,, \[I,:×.6\]. 
\[ICx.5\] itada i, ker, ,o ka 
\[F,x.f;\] itada i, ita.da keru ,o kn 
In these examph'.s, "ita.da" is it stem and 
the Sl:)eak(w first tries to say "itada ita" or 
"ita(la i re" and then changes to "ita(la keru". 
I,t the case of \[l';x.6\], a repetition of a stem 
can be used as a (:lue. In the. case of \[Ex.5\], 
existence of an affix without a stem indicates 
an interruption point and a reparandum. 
fresh start 
l"resh start is a rel)air with a complet(;ly 
diffe.r(mt utterance. A fragment of utterance 
I)efore interrulfl.iotl is ignored. SI';ltUP tries 
the detection of fresh stm't if all possible (:lues 
are n()t, fou,(l. It tries to pa.rse the fragment 
of utterance without a first word of it. it 
rel)eats this trial until I)ar~ing succeeds. 
A 
~~ uttaranco l 
isolated-word 
l"ig. 3: An isolated word 
595 
others 
SERUP cannot cope with utterances of all 
these categories. 
changed to well-formed 
A self-repaired utterance is occasionally 
parsed successfully as a well-formed utter- 
ance that has a meaning that the speaker 
does not intend. For example, in \[Ex.7\], 
a fragment "kyou" of a word "kyousan 
"(cosponsorship) is treated as a word "kyou 
"(today), and parsed successfully but the 
meaning of it is "cosponsor today". 
\[Ex.7\] kyou, kyonsan suru 
Some of these utterances can be detected 
as an error in semantic interpreter. And wc 
think prosodic cues can be used effectively, 
because a fragment "kyou" and a word "kyou 
" is pronounced differently. So far, SERUP 
cannot cope with such utterances, because it 
uses well-formed first method. 
dividing word 
In \[Ex.8\] the speaker starts repair within 
word. 
\[Ex.8\] junji, bi ni desu ne 
The speaker tries to say "junbi ni desu ne", 
but makes a lexical error "junji". IIe starts 
the repair with a fragment "bi" of "junbi', 
instead of a complete word "junbi". This is 
a very rare case. 
repetition with different cate- 
gory 
Speakers occasionally repair with different 
category of words. A human listener can 
draw some inference and find relation be- 
tween words, but automatic detection is dif- 
ficult. 
ambiguous repair 
In \[gx.9\], it is ambiguous what kind of self- 
repair is made. 
\[gx.9\] apointo wo, nl, er, suuzitu 
tyuu ni 
The speaker may repair a particle "wo" 
with "ni", or repair a fragment "ni" of a 
word "nisanniti" that has the same meaning 
of "suuzltu"(some days). We cannot solve 
this anablguity automatically. 
LARGE COIl.PUS-BASED 
ANALYSIS 
To investigate effectiveness of SERUP we 
analyzed a large corpus called ADD(Ehara 
et al. 1990). ADD contains one million words 
of dialogues about registration to an inter- 
national conference over telephone. ADD is 
created at ATR Interpreting Telephony Lab- 
oratories. 
There are 1.,082 self-repairs in the corpus. 
With these self-repairs, we investigate the 
categories they belong to. Table 1 shows the 
result. 
DISCUSSION 
In sum, SERUP seems to cope with 
953(88.1%) of self-repairs. We think SERUP 
is effective to Japanese self-repaired utter- 
&rices. 
Most of utterances that SERUP cannot 
~ope with are in tile category "Changed to 
well-formed". As we mentioned, these utter- 
ances might be processed successfully with 
semantic constraints or prosodic cues. If we 
could implement them, SERUP would cope 
with 1,064(98.3%) self-repairs. 
CONCLUDING 1-I.EMARKS 
We proposed SERUP, a parser coping with 
self-repaired Japanese utterances. SERUP 
uses some linguistic clues and translates a 
self-repaired utterance into well-formed ver- 
sion and parses it. The result of large corpus- 
based analysis suggests that 88.1% of 1,082 
self-repairs can be processed by SERUP. 
Our future directions are to test the sys- 
tem with large grammar and lexicon and to 
incorporate prosodic processing. 
596 
Table 1: The result of ananlsys 
With repetition 
A repetition 
B repetition 
C repetition 
D repetition 
same constituent repetition 141(13.0%) 
same category repetition 108(10.0%) 
same constituent repetition 96(8.9%) 
same category repetition 2(0.2%) 
same constituent repetition 136(12.6%) 
same category repetition 3(0.3%) 
same constituent repetition 4(0.4%) 
same category repetition 0(0%) 
With syntactic break 
..... Same fragment repetition 
With unknown word 
With isolated word 
Without repetition of a stem 
Fresh restart 
105(9.7%) 98(9.1%) 
235(21.7%) 
23(2.1%) 6(0.6%) 
Changed to well-formed 11.1(10.3%) 
Others Dividing word 4(0.4%) 
Repetition with diIferent categozy 5(0.5%) 
Ambiguous repair 4(0.4%) 
Total of suceessable 953(88.1%) 
Total 1,082 
1Zeferences 
Blackmer, E. R. and J. L. Mitton (1991). 
Theories of monitoring and the timing of 
repairs in spontaneous speech. Cognition 
39,173-194. 
Dowding, J. el al. (1993). Gemini: A natu- 
ral language system for spoken-language 
understanding. In Proceedings of lhe 31st 
Annual Meeting of A CL. 
Ehara, E. et al. (1990). Contents of the ATR 
dialogue database. Technical Report TR- 
1-0186, ATP~ Interpreting Telephony Re- 
search Laboratories. 
lIindle, D. (1983). Deterministic parsing of 
syntactic non-fluencies. In Proceedings of 
the 21st Annual Confe;ence of the ACL, 
pp. 123-128. 
Langer, It. (1990). Syntactic normalization of 
spontaneous speech. In COLING 90, pp. 
180-183. 
Levelt, W. J. M. (1988). Speaking: From 
Intention to Articulation, Chapter 12, 
pp. 458-499. Cambridge, MA: The MIT 
Press. 
MADCOW (1992). Multi-site data collection 
for a spoken langnage corpus. In Proceed- 
ings of the DARPA Speech and Natural 
Language Workshop, pp. 7-14. 
Nakatani, K. and J. IIirschberg (1993). A 
speech-first model for tel>air detection and 
correction. IT, Proceedings of the 31st An- 
nual Meeting of A CL, pp. 46-53. 
Sagawa, Y., N. Ohnishi, and N. Sugie (1993). 
Repairing self-repairs in Japanese. In Pro- 
ceedings of Natural Language Processing 
Pacific Rim Symposium (NLPRS '93), 
Fukuoka, pp. 191-198. 
Shriberg, E., J. Bear, and J. Dowding (1992). 
Automatic detection and correction of re- 
pairs in human-computer dialog. In Pro- 
ceedings of the DARPA Speech and Natu- 
ral Language Workshop, pp. 419-424. 
597 
