An Acquisition Model for both Choosing and Resolving 
Anaphora in Conjoined Mandarin Chinese Sentences 
Benjamin L. Chen 
Application Software Department 
Computer and Communication Research Laboratories 
Industrial Technology Research Institute 
W300/CCL, #1 RA~D Rd. I, 
Hsinchu, 30045, Taiwan, R. O. C. 
e-mail: blchen@jaguar .ccl.it ri.org.tw 
Von-Wun Soo 
Department of Computer Science 
National Tsing-Hua University 
Hsinchu, 30043, Taiwan, R. O. C. 
e-mail: eoo@es.nthu.edu.tw 
Abstract 
Anaphoric reference is an important linguistic phe- 
nomenon to understand the discourse structure and 
content. In Chinese natural language processing, 
there are both the problems of choosing and resolv- 
ing anaphora. In Mandarin Chinese, several linguists 
have attempted to propose criteria to ezplain the phe- 
nomenon of anaphora but with controversial results. 
On the other hand, search-based computational tech- 
niques for resolving anaphora are neither the best way 
to resolve Chinese anaphora nor to facilitate choosing 
anaphora. Thus, to facilitate both choosing and resolv- 
ing anaphora with accuracy and efficiency, we propose 
a case-based learning model G-UNIMEM to automat- 
ically acquire anaphorie regularity from a sample set 
of training sentences i, which are annotated with a 
list of features. The regularity acquired from training 
was then tested and compared with other approaches 
in both choosing and resolving anaphora. 
Keywords: anaphoric reference, semantic 
roles(case), natural language acquisition, ease- 
based learning. 
i Introduction 
In discourse, there may be anaphora in two consec- 
utive sentences. When anaphora appear in a pair of 
consecutive sentences, the two consecutive sentences 
are called conjoined sentences. In real life conver- 
sation, we frequently choose and resolve anaphora 
to understand the utterances. There are primarily 
three types of anaphora in Mandarin Chinese: zero 
(ellipsis), pronominal (using pronoun) and nominal 
anaphora\[4\]. Let's take the conjoined Chinese sen- 
tencez in (B) to illustrate the phenomenon. The con- 
1This paper is partially supported" by the Minis- 
ter of Economic Affairs, R.O.C. under the project no. 
33H3100 at ITRI and by National Science Councial of 
R.O.C. under the grant no. NSC81-0408-E007-02 at 
National Tsing Hua University. The authors also want 
to thank Dr. Martha Pollack for valuable comments 
during 1991 UCSC linguistic institute. 
joined sentence in (C) is the English translation of the 
Chinese sentences m (B). 
(B) Yueh-haa sheng-bing. \[ \] i-thing 
hui-chia-le. 
John got sick \[ \] already 
gone home 
(C) Because John .as sick, he has gone home. 
Because the anaphora in (B) is a zero anaphora and 
there is no zero anaphora in English, the antecedent 
of zero anaphora in (B) must be resolved first before 
choosing an appropriate pronominal anaphora in Chi- 
nese to English translation. In the translation from 
(C) to (B), it is not good to directly translate an En- 
glish pronoun to a Chinese pronoun. A better way 
is to resolve the anaphora ire (C) and then choose an 
appropriate type of anaphora in Chinese. 
In natural language processing, better results seems 
to be attainable if rich linguistic or domain knowl- 
edge is available. However it generally costs much 
and doesn't seem to be realistic. The same situa- 
tion applies for resolving and choosing anaphora in 
Mandarin Chinese. If we only used search-based ap- 
proaches(those that merely used heuristic and algo- 
rithmic methods without much linguistic knowledge), 
the performance was limited. However, when we in- 
tended to adopt linguistic knowledge, we found lin- 
guists' theories tended to be controversial and less 
computable. Thus, it motivated us to pursue an ac- 
quisition model that could acquire linguistic regularity 
from corpora and then used the regularity to resolve 
and choose anaphora. 
2 Review of previous 
approaches 
2.1 Search-based approaches 
Both history list \[1\] and Hobbs's naive syntactic 
.algorithm \[7\] are search-bused approaches for re~lv- 
mg anaphora. However, it's not quite obvious to tell 
which was better than the other with only few exam- 
pies. Thus, we collected 120 testing instances to test 
them. Those instances were selected from linguists' 
ACTES DE COLING-92, NANTES, 23-28 Aotrr 1992 2 7 4 PROC. OF COLING.92, NANTES, AUO. 23-28, 1992 
examples, textbooks, essays and novels. Half of them 
contained zero anaphora the other pronominal. 
The result showed that the correct number was 
111(92.5%) with Hobbs's syntactic algorithm and 
87(72.5%) with the history list approach if first 
matched were selected. There was 109(90.8%) cor- 
rect for history list if the last matched were selected. 
It seemed that both approaches were applicable to re- 
solve anaphora, tiowever, when there are several NPs 
with the same semantic features, both approaches may 
get into troubles. \]~harthermore, both cannot be used 
to choose anaphora. 
2.2 LinguisCs criteria 
Among linguists' works \[31 \[51 \[lll \[121 \[141, Tai's 
criteria \[14\] was applicable to both choose and resolve 
anaphora. Others' suffered from difficulties of extract- 
ing features or resolving anaphora. Table 1 shows 4 
co-references for Tai's citeria, which are all applica- 
ble when co-referred NPs are human. For example, 
consider the following conjoined sentences: 
Tai:\[ Lao Zhang \] dao-le Meiguoyihou. \[ \] jiac-le 
hen-duo pengyou. 
John came U.S.A after \[\] made 
many friends 
Since John came to the U.S.A., he has made many 
friends. 
The subject in the first sentence is human and co- 
referred by the subject in the second sentence, so this 
is a subject-subject co-reference. According to Table 
1, zero anaphora is preferred to the pronominal one 
and nominal anaphora is not permitted in this exam- 
ple. Though Tal didn't propose the criteria for resolv- 
ing anaphora, it was possible to get these criteria just 
by transforming the choosing criteria in reverse order. 
After Tai's criteria were applied to choose and re- 
solve anaphora on the 120 testing instances, we got 
the success numbers 86(71.7%) and 65(54.2%) respec- 
tively. The results failed to meet our satisfaction. 
Through above paragraphs, it appears that search- 
based methods have their limitations due to lack of 
enough linguistisc knowledge and Tai's criteria seems 
to be applicable to both choose and resolve anaphora. 
It might be that Tai's criteria were too general to 
lead to a high success rate. More reliable method 
to acquire regularity might be required to promote 
the success rate. We hypothesized the regularity of 
anaphora could be accounted by causal relations be- 
tween the features in the conjoined sentences and the 
antecedents. In the following section, an acquisition 
model is introduced. 
3 G-UNIMEM: A Case-Based 
Learning Model 
In natural language acquisition problem, the re- 
striction of positive-only examples \[2\] has prohibited 
many machine learning models as a feasible natural 
language model. However, a case-baaed learning ap- 
proach such as Lebowitz's UNIMEM \[9\] \[I0\] seems to 
be a candidate due to its capability to form concepts 
incrementally from a rich input domain. Neverthe- 
less, to apply UNIMEM directly to the acquisition of 
anaphoric regularity in Mandarin Chinese is still not 
sufficient. We have therefore modified UNIMEM into 
G-UNIMEM. 
G-UNIMEM, a modified version of UNIMEM, is 
an incremental learning system that 
uses GBM(Geueralized-based Memory) to generalize 
concepts from a large set of training instances. The 
program was implemented in Quintus PROLOG and 
on SUN workstation. 
G-UNIMEM differs from UNIMEM in two respects. 
Firstly, if a drinker got drunk many times after tak- 
ing either whiskey and water or brandy and water, 
he would induce that water made him drunk with 
UNIMEM. This is intuitively incorrect. Whereas, 
with G-UNIMEM, he would induce that whiskey and 
water, brandy and water or water would cause him 
drunk. In this case, G-UNIMEM retains the possible 
causal accounts without committing to erroneous con- 
clusion. Secondly, G-UNIMEM can extract explicit 
causal rules from memory hierarchy. 
Similar to UNIMEM, G-UNIMEM organizes input 
training instances into a memory hierarchy according 
to the frequencies of features. However, its goal is 
to explicitly express the generalized causal relation- 
ships between two specified types of features: cause 
features and goal features. Since there may be incon- 
sistency due to lack of cause features, further refine- 
meat is needed to obtain consistent causal relations. 
Thus, there are four different modules in G-UNIMEM 
to complete different functions in order to achieve this 
purpose. 
3.1 The classifier 
The classifier is the first module that processes 
all training instances for G-UNIMEM. Its function is 
close to UNIMEM that organizes a hierarchy structure 
to incrementally accommodate a training instance and 
at the same time generalize the features based on sim- 
ilarities among training instances. The forming hier- 
archy is organized as either a g-c-hierarchy or a c-g- 
hierarchy depending on the setup of system, which is 
defined in Definition 1. In Appendix A we show the 
basic classifier algorithm. 
Definition 1 A g-c-hierarchy is the hierarchy that 
every generalized goal feature resides in a GEN-NODE 
and there is no generalized cause feature that resides 
between the root node and this GEN-NODE. A c-g- 
hierarchy doesn't allow any generalized goal feature to 
reside in the GEN-NODE between the root node and 
any GEN-NODE where generalized cause features re- 
side. 
Figure 1 and Figure 2 show the forming g-c- 
hierarchy and c-g-hierarchy respectively after 13 anno- 
tated training sentences are entered into G-UNIMEM. 
Generally, g-c-hierarchy would be chosen since it re- 
tained all possible causal accounts. For example, the 
drinker with g-c-hierarchy would induce that whiskey 
and water, brandy and water or water would cause 
him drunk; whereas, he would induce whiskey and 
Ac'rY.s DE COLING-92, NANTES, 23-28 AOU'r 1992 2 7 5 FROC. OI' COLING-92, NAhrrES, AUG. 23-28, 1992 
water, brandy and water with e-g-hierarchy. The c- 
g-hierarchy is more efficient since no rules are needed 
to be generated. Fig. 3 and Fig. 4 show the updating 
of a GBM before and after inserting a new training 
instance. 
3.2 The rule generator 
Once a hierarchy has been constructed by the clas- 
sifter, the causal rules can be extracted. The rule gen- 
erator module serves as the role to extract causal rules 
from the hierarchy. It generates all causal rules from 
the hierarchy as the regularity is retrieved for predic- 
tions. 
In Fig. 6, if a testing instance is given for choos- 
ing anaphora with a query feature list \[ (g,type(*?)), 
(g,ante(theme)), (c,fl(theme)), (c,anaphor(theme)), 
(c,s2(obj)), (c,p(pv))\], the retrieval process is searched 
with a post-order traverse, namely, in the order se- 
quence of the node number 1, 2, 3, 4, 5 and 6. Since 
there may be more than one candidate, the system can 
be setup to select either the first or the most specific 
one. If the first one is preferred, type(nil) is yielded 
as the prediction. If the most specific answer is pre- 
ferred, all possible rules will be tried and the one with 
the most number of contingent features matched will 
be the answer(i.e, type(pronoun) ). 
The sample rules generated from Fig. 1 are shown 
in Fig. 5. Before generating rules, the GBM is ad- 
justed so that all children of a GEN-NODE are or- 
dered according to their confidence scores of features. 
Then all rules are generated in a post-order traversal. 
3.3 The rule filter 
The rule filter removes those rules that are ill- 
formed and useless. For example, the causal rule 5 
in Fig. 5 has no causes which is not a well-formed 
rule. It also detects conflicting rules. Conflicting rules 
are those that have different goal feature descriptions, 
which are accounted by the same cause. For example, 
the rule I and rule 6 in Fig. 5 are conflicting. These 
rules will be detected in this module and then to be 
resolved by the feature selector. 
3.4 The feature selector 
Any two conflicting rules are resolved by the fea- 
ture selector through augmenting the two rules with 
mutual exclusive contingent cause features, which are 
prepared in advance. Dominant features were used 
in initial regularity acquisition stage; whereas contin- g 
ent features were used in feature selection stage. The 
ominant features such as goal features are assumed to 
be those that must be present in every anaphoric rule. 
Contingent features are optional. Fig. 6. shows the 
GBM with g-c hierarchy after feature selection pro- 
c¢08. 
4 Tests using sentences anno- 
tated with mixed features 
We trained G-UNIMEM with 30, 60, 90, 120 in- 
stances using those features mentioned by Tai, and 
used all the 120 instances as testing instances. It 
showed that the approach using Tai's criteria was not 
promising. There are two reasons. First, none of the 
success rates was as high as those using the history list 
approach or IIobbs's algorithm. Second, many con- 
flicting rules remained conflicting due to either that 
no further features from feature selection were avail- 
able or too many specific training leading to too many 
specific rules. These factors decreased the success rate. 
4.1 Selecting mixed features 
Since Tal's features were not sufficient, more se- 
mantic features were considered. Among several lin- 
guists' works, we tentatively selected some computa- 
tional feasible syntactic and semantic features from 
different sources \[3\] \[5\] \[11\] \[12\] \[13\] \[14\] \[15\] as in Ta- 
ble 2. An example with annotated features is shown 
below. Tile notation \[ \] represents zero anaphora. 
(C)\[Lao zheng\]i qu-le ji-ge \[nurenli.\[ \]j hen hui zuo-cai. 
John married a woman t \] wetlcan cook. 
agent theme agent 
hm sub,by 
nondefinite 
John married a woman, and the woman cooked 
well. 
The training feature list for the sentences (C)is : 
\[(g ante(theme)), (g,type(nil)), (c,fl(agent)), 
(c,f2(theme)), (c,anaphor(agent)), (c,p(bv)), 
(c,s2(sub)), (c,h(hm)) (c d(nondefinite)) 
where the notations g and c represent goa and cause 
features respectively. 
4.2 Testing using mixed features 
After semantic features has been determined, we 
trained G-UNIMEM with 30, 60, 90, 120 instances 
and used all the 120 instances as testing instances each 
time.We hypothesized to choose semantic roles(i.e. 
ease) as dominant cause features. The features such as 
ante(CASE), type(X), anaphor(CASE) and fi (CASE) 
are dominant features and the number of fi is vari- 
ant. The hypothesis was motivated by Sidner \[13\] who 
used semantic roles to determine focus and resolve def- 
inite anaphora. The others such as h(Hm), p(POS), 
s2(SYN); d(D), con(s) belong to contingent features. 
4.3 The experimental results 
It is interesting that the success numbers in Ta- 
ble 3 increased with the number of training instances. 
Finally, our results showed that experiments with c- 
g-hierarchy had a little high accuracy rates (95.8% 
for resolving and 90.8% for choosing anaphora with 
120 training instances) than thoee with g-e-hierarchy. 
Both accuracy rates were higher than those with TaPs 
criteria \[14\]. Thus, G-UNIMEM with semantic roles 
as dominant features promised much higher accuracy 
rate. 
In Appendix B we show some sample rules acquired 
in Horn-like clauses. After examination, either the 
agent or ~heme of first sentence is most likely to 
AcrEs DE COLING-92, NANTES, 2.3-28 AOOT 1992 2 7 6 PROC. OF COLING-92. NANTES, AUG. 23-28, 1992 
act as antecedents of anaphora. Tiffs phenomenon is 
in coincidence with the investigation on anaphora by 
Sidner. That is, the agent often appeared as actor 
focus and theme as default focus . This is similar to 
Tai's criteria but is in more compact interpretation. 
5 Discussion 
There are two concerns in implementing G- 
UNIMEM: 
(1) The feature set : Is the assignment of dominant 
features and contingent features objective? If there 
is any contingent feature in the assignment that obvi~ 
ously improves the accuracy rate, it shonld be assigned 
as dominant feature. We use statistical methods \[8\] to 
analyze if contingent features actually improve accu- 
racy rates. If there is no obvious improvement with 
contingent features, the division of dominant and corr- 
tingent features is acceptable. 
We made the null hypothesis "G-UNIMEM with c- 
g-hierarchy doesn't have obvious improvement with 
contingent features" and the alternative hypothesis 
"G-UNIMEM with c-g-hierarchy has obvious improve- 
ment with contingent features". We titan got two 
test values from test statistics: tl = 0.8472 and t2 
< 0. Both test statistic.q were less than t~ = .05 (= 
1.734 with d.f. = 18). Thus, the null hypothesis "G- 
UNIMEM with c-g-lfierarchy doesn't have obvious im- 
provement with contingent features" was not rejected, 
which justified that G-UNIMEM using semantic roles 
as dominant features was valid. 
(2)The sample size : Compared with actual linguis- 
tic domain, the 120 training and testing instances are 
small. A large corpus is desirable to test the system's 
performance. If it becomes available, our resnlts would 
be more objective and reliable. 
6 Conclusion 
We have illustrated a way of using machine learn- 
ing techniques to acquire anaphoric regularity in con- 
joined Mandarin Chinese sentences. The regularity 
was used to both choose and resolve anaphora with 
considerable accuracy. Table 4 shows a comparison 
between different approaches. 
In comparison to other approaches, tire proposal of 
using G-UNIMEM as the acquisition model and using 
semantic roles as dominant features is practical and 
serves multiple purposes. 
A(:Iq~S I)E COLING-92, NANTES, 23-28 Aolrr 1992 2 7 7 PRO{:. ov COLIN(L92, NANTES, AUG. 23-28, 1992 
g,type(nil)):11 1' (g,type(pronoun)):2 I 
\ \[ (g,ante(theme)):2 
i\ I (c,anaphor(theme)):2 (g,ante(theme)):7 
\\[ ic,fl(thenm)): 2 (c,fl(theme)):7 \[\~ 
c ,anaphor(t heine) ): 7 i \. 
(g,ante(agent)):4 \[ 
/I (c,fl(agent)):4 "--l~x~ 
tc,anaphor(theme)):2 (c,f2(theme)):2 i \[ , .\] I (c,anaphor(agent)):~ 
Fig. 1. A g-c-hierarchy of GBM 
(g,ante(theme)):3 
inst:\[(g,type(pronoun))\] j 
Fig. 4 A new GBM after inserting a new 
instance \[(g,type(pronoun)),(g,ante(theme))\] 
1: \[(g,type(nil)),(g,ante(theme))\] :- \[(c,:fl(gheme)),(c,anaphor(theme))\] 
2: \[(g,type(nil)),(g,aute(agent))\]:- \[(c,:fl(agent)),(c,anaphor(theme))\] 
3: \[(g,type(nil)),(g,ante(agen¢))J:- \[(c,fl(agent)),(c,:f2(theme)), 
(¢, a.aaphor (agent) )'1 
4: \[(g,type(nil)), (g,ante(agent))\] :- \[(c,f l(agent))\] 
5: \[(g,type(nil))\] :- \[\] 
(c,fl(theme)):9 \] I (c'fl(agent)):4 16: \[(g,type(pronotm) ), (g, a.ate (gheme))\] :- 
c,amaphor(theme)):9\[ / \[(c,:f 1 (theme)) (c,aaaphor(th~e))\] 
(g,ante(theme)):9 \[ Fig. 5 The sample rules generated :from 
\[ (c,anaphor(theme)):2 " Fig. 1. 
\[ (g,type(nil)):2 
~g,ante(agent)):2 , 
3 ...... (c,anaphor(agent)):~ 
lg,type(pr°n°un))i~ (c,f2(theme)):2 \[ (g,type(nil)):2 | 
g,type(nil)):7 \] (g,ante(agent)):2 ,. \[ 
Fig. 2. A c-g-hierarchy of GBM 
(g,type(nil)):2 
(g,ante(theme)):2 1 
Fig. 3 A GBM with two training instances 
I/c,anaphor/theme/):  ' / / 
k (c,anaphor\[theme)):2 \ (g,type(nil)):2 \ 
2 ~ \] (g,ante(agent)):2 \ 
(g,type(pronoun)):2' 5 (c,anaphor(agent)):~ 
(g,ante(theme)):2 (c,f2(theme)):2 | (c,s2(obj)):2 
(g,type(nil)):2 / (c,p(pv)):2 (g,ante(agent)):2 \] 
1 
(g,type(nit)):7 
(g,ante(theme)):7 1 
Fig. 6 The c-g hierarchy after conflict 
resolution from Fig. 2 
AcrEs DE COLING-92. NANTES. 23-28 AO0r 1992 2 7 8 PROC. OF COLING-92, NANTES. AUG. 23-28, 1992 
Table 1: TaPs criteria for choosing anaphora 
I .... wneos_e \] zero--5-~ro~ n~onoun~rouou~-~~_nom,nal j 
I Not permitted \[ nominal \] zero \[ ~- ~----ze-m \] 
no~. u j~~J. c~-#&Tgf&__ ~ ~ i~pre~rre~ g87 
T~m:es and efe e ces ~ . feature ~ notation .... 
semantic antecedent ante CASE)~ 
semanttc role \[ ti(CASE) 
'~ I --anaphoriCXSE) -- ~an 
or no~~~n(nonhm~-- 
syntactic anaphora \[ type(X) 
1___ subject .... \[ sz(suz~, ... 
I position of anap\[i-o~--p~gv~o~-- 
\] I bv: before verb; pv: post-verb 
\[_____.~finite . \[ d(uondefinite) __ 
I co.ne~or I c°n/s) 
notation :'GAS rE-rE-rE-rE-rE-rE-~g-presents a variable for a semantic role an~ order in a sequence of roles. 
Tans 
= Chm~ 
-- ru~-lg 
number 
Accuracy rate 
(for resolving) 
Accuracy rate (for choosing 
~g number 
(after resolving) 
Table 3: Comparison of G-UNIMEM using ~ against Tai's criteria 
Group Candidate E 3 ~Y-~"qS"0- ~ --~-'E~20 "~ TaP~4-- 
choice criteria 
30** 
specific 58/62 87/96 104/110 114/120=95.0% 
,,one 58/62 86/96 104/110 111/120=:92.5% 
trst\]'?/~ 53/60 83/99-~ 95/109 109/120:-'90.8°70- ~TT2ffg"TE=71.7~U-o 
specific 56/65 83/100 88/110 101/120=84.2% 
4 -----~*~-- 
notation: * ." for choosin~r~ ~g; accuracy ~g~'cc~s- 
that have applicable rules; none, in column 2 means no contingent features are used. 
Table 4: Comparison of different approach~ 
MethOds ~_~os~ 
1. --history list L~___~_~e search 
\[" llobbs"s syntactic -- ~'-----/'-~-~~e~y-~-y~ 
\]. al~;orithm ~ J ~s mter,'g----- c 
oo~ pre ,ct~~ 
\[.~ Chen's criteria \[ eho~ pl_~.dlc~ not easy J 
\] G-UNIMEM & dominant \] choose & relive \] predict \] easy J 
I features \] ~_~ J 
AcrEs DE COLING-92, NANTES, 23-28 noun" 1992 2 7 9 PROC. OF COL1NG -92, NANTES, AUG. 23 -28, 1992 
Appendix A. The basic classifier algorithm 
input: The current node N of the concept hierarchy. 
The name I of an unclassified instance. 
The set of I's unaccounted features F. 
Results: The concept hierarchy that classifies the instance. 
Top-level call: classifier( Top-node, I, F) 
Variables: N, N', C and NC are nodes in the hierarchy. 
G, H, and K are sets of features. 
J is an instance stored on a node. 
P~ is a variable of set. 
Classifier(N, 1, F). 
Let G be the set of features stores in N. 
Let H be the features in F that match features in G. 
Let K1 be the features in F that do not match features in G. 
Let K2 be the features in G that do not match features in H. 
Let H', KI' and K2' be the sets of features after Adjust(H,K1,Ki,H',KI',Ki') 
/* adjust goal and cause features for 
g-c-hierarchy or c-g-hierarchy */ 
if N is not the root node, 
then if H is empty set/* no features match */ 
then return False 
else if both H' and KI' are not empty sets 
then ~/* split node N */ 
spht N into N' and NC where NC is a child of N'; 
N' contains features in H' with confidence scores 
and I as a instance with features KI'; 
each confidence score in tl' is increased by 1; 
the remaining features and instances belong to NC; 
return Split. ) 
else if It' and H are equal/* all features match */ 
then increase each confidence score in N by 1. 
for each child C of node N/* continue match remaining features */ 
call Classifier(C, I, KI') and collect returns to the set It, 
if any Classifier(C, I, KI') call return True or Split then break. 
if 1% is \[False \] /* All trials fail, try to do generalization */ 
then for each instance J of node N 
call Generalize(N, J, I, KI') and collect returns to the set 1%, 
if any Generallze(N, J, I, KI') call return True then break. 
if tt is \[ False \]/* All trials fail, insert I as an instance of N */ 
then store I as an instance of node N with features KI'. 
return True. 
Appendix B. Sample rtllen of regularity ~ith high probability of appearance 
in Horn-like clauses 
\[an~;e (agent), gype(nil)\] :- \[anaphor(agent) ,f2(theme) ,ft (agent)\] 
\[ante(agent) ,type(nil)\] :- \[anaphor (agent) ,f l(agent)\] 
\[ante(theme) ,type(nil)\] :- \[p(bv). s2 (sub) ,d(nondefinite), anaphor (agent), f I (agent), f2 (theme)\] 
\[ante(theme). type(nil)\] :- \[h(nonhm), anaphor (theme). f i (agent), fi(theme)\] 
\[ante(theme) .type(nil)\] :- \[anaphor (theme) ,fl (theme)\] 
\[ante(ar E) .type(pronoun)\] :- \[anaphor(agent), f2(pred) , fl(arg)\] 
\[ant • (agent). type (pronoun) \] : - \[h(hm). d(definite), con(s) 0anaphor(agent). f2 (theme), f i (agent)\] 
\[unt e (art) ,type(pronoun)\] :- \[k(hm), anaphor (a~ent), f 2(pred) ,f l(arg)\] .. 
\[ante(theme). type(pronoun)\] :- \[s2(obj ) ,p(pv) ,d(definite),anaphor(theme) ,f I (agent) ,f2(theme)J 
\[ant e (art) .type(pronoun)\] :- \[h(hm), anaphor (theme). f2 (pred) ,f I (arg)\] 
Acrf~s DE COLING-92, NAN'IT.S, 23-28 AoU'r 1992 2 8 0 PROC. OF COLING-92, NANTES. AUG. 23-28, 1992 

References 

James Allen (1987), Natural Language Under- 
standing, The Benjamin/Cummings Publishing 
Co. 

Robert Berwick (1986), Learning From 
Positive-only Examples: The Subjective Princi- 
ple and Three Case Studies. In R.Michalski, 
J. Carbonell, and T. Mitchell(cal.) Machine 
Learning: An Artificial Intelligence Approach 
Vol. \[I, Chapter 21, Morgan Kanfmann Publish- 
ers, Inc. 

Ping Chen (1984), A Discourse Analysis of Third Person Zero Anaphora in Chinese, repro- 
duced by Indianan University Linguistics Club, 
Bloomington, Indiana. 

Ping Chert (1987), A Discourse Approach to 
Zero Anaph0ra in Chinese, ghonggua Ynwen, 
Beijiug. 

Chauncey C. Chu (1983), Definiteness, Pre- 
suppositimt, Topic and Focus in Mandarin Chi- 
ne~e, Student Book company, Taipei, Taiwan. 

John H, Gennari, Pat Lmlgley and Doug. 
Fisher (1989), Models of Incremental Concept 
Formation, Artificial Inielligence 40. 

J. llobbs (1978), Resolving Pronoun References, 
Lingua 44, North-tlolland Publishing Co. 

Richard Johnson and Court Bhat- 
tacharyya (1985), Statistics: Principles and 
Methods, Chapter 6 (Section ,I), 9, 10, 11, Pub- 
lished by Hohn Wiley & Sons. 

Michael Lebowitz (1986), Concept Learning 
in ~ Rich Input Domain: Generalization-based 
Memory. Ira It. Midralski, J. Carbonell, 
and T. Mitcholl(ed.) Machine Learning: An 
Artificial \[nlelligence Approach Vol. 1I, Chapter 
8, Morgan Kaufmann Publishers, Inc. 

Michael Lebowitz (1986), Integrated Learning: 
Cmttrolling Explanation, Cognitive Science 10. 

Cherry lug Li (1985), Participant anapho~ in 
Mandarin Chinese, University of Florida Press. 

Met Du Li (1986), Anapho,~c Structure of Chi- 
nese, Student Book CO., Taipei, Taiwan. 

C. Sidner (1983), Focusing in tim Compre ~ 
hens\[on of Definite Anaphora, Iu Computational 
Models of Discourse, M. Brady and R. Berwick, 
eds., MIT Press. 

J. lI. Tat (1978), Auaphoric Constraints in 
Mandarin Chinese Narrative Discourse. In J. 
Hinds (ed.) Anaphora in Discourse, Linguistic 
Research, Edmonton, Alberta. 

Ting-Chi chanties Tang (1975), A Case 
Grammar Classification of Chinese Verbs, Ha\[- 
Gun Book Company, Taipei, Taiwaa. 
