Proceedings of the 7th SIGdial Workshop on Discourse and Dialogue, pages 76–79,
Sydney, July 2006. c©2006 Association for Computational Linguistics
Semantic tagging for resolution of indirect anaphora
R. Vieira1, E. Bick2, J. Coelho1, V. Muller1, S. Collovini1, J. Souza1, L. Rino3
UNISINOS1, University of Denmark2, UFSCAR3
renatav@unisinos.br, eckhard.bick@mail.dk, lucia@dc.ufscar.br
Abstract
This paper presents an evaluation of indi-
rect anaphor resolution which considers as
lexical resource the semantic tagging pro-
vided by the PALAVRAS parser. We de-
scribe the semantic tagging process and a
corpus experiment.
1 Introduction
Bridging anaphora represents a special part of the
general problem of anaphor resolution. As a spe-
cial case of anaphora, it has been studied and dis-
cussed by different authors and for various lan-
guages. There are many problems in develop-
ing such studies. First, bridging is not a regu-
lar class, it seldom contains cases of associative
and indirect anaphora (defined in the sequence);
lexical resources such as Wordnet are not avail-
able for every language, and even when available
such resources have proven to be insufficient for
the problem. In fact, different sources of lexi-
cal knowledge have been evaluated for anaphora
resolution (Poesio et al., 2002; Markert and Nis-
sim, 2005; Bunescu, 2003). At last, corpus stud-
ies of bridging anaphora usually report results
on a reduced number of examples, because this
kind of data is scarce. Usually bridging anaphora
considers two types: Associative anaphors are
NPs that have an antecedent that is necessary
to their interpretation (the relation between the
anaphor and its antecedent is different from iden-
tity); and Indirect anaphor are those that have
an identity relation with their antecedents but the
anaphor and its antecedent have different head-
nouns. In both associative and indirect anaphora,
the semantic relation holding between the anaphor
and its antecedent play an essential role for res-
olution. However, here we present an evalu-
ation of the semantic tagging provided by the
Portuguese parser PALAVRAS (Bick, 2000)
(http://visl.sdu.dk/visl/pt/parsing/automatic) as a
lexical resource for indirect anaphora resolution.
We focus on indirect anaphors for two reasons,
they are greater in number and they present better
agreement features concerning human annotation.
2 Semantic Annotation with Prototype
Tags
As a Constraint Grammar system, PALAVRAS
encodes all annotational information as word
based tags. A distinction is made between mor-
phological, syntactic, valency and semantic tags,
and for a given rule module (or level of analysis),
one tag type will be regarded as primary (= flagged
for disambiguation), while tags from lower lev-
els provide unambiguous context, and tags from
higher levels ambiguous lexical potentialities.
Thus, semantic tags are regarded as secondary
help tags at the syntactic level, but will have un-
dergone some disambiguation at the anaphora res-
olution level. The semantic noun classes were
conceived as distinctors rather than semantic de-
finitions, the goal being on the one hand to cap-
ture semantically motivated regularities and rela-
tions in syntax, on the other hand to allow to dis-
tinguish between different senses, or to chose dif-
ferent translation equivalents in MT applications.
A limited set of semantic prototype classes was
deamed ideal for both purposes, since it allows at
the same time similarity-based lumping of words
(useful in structural analysis, IR, anaphora reso-
lution) and context based polysemy resolution for
an individual word (useful in MT, lexicography,
alignment). Though we define class hypernyms
as prototypes in the Roschian sense (Rosch, 1978)
76
as an (idealized) best instance of a given class of
entities, we avoided low level prototypes, using
<Azo> for four-legged land-animals rather than
<dog> and <cat> for dog and cat races etc.).
Where possible, systematic sub-classes were es-
tablished. Semiotic artifacts <sem>, for instance
are sub-divided into “readables” <sem-r> (book-
prototype: book, paper, magazine), “watchables”
<sem-w> (film, show, spectacle), “listenables”
etc. The final category inventory, though devel-
oped independently, resembles the ontology used
in the multilingual European SIMPLE project
(http://www.ub.es/ gilcub/SIMPLE/simple.html).
For the sake of rule based inheritance reasoning,
semantic prototype classes were bundled using a
matrix of 16 atomic semantic features. Thus,
the atomic feature +MOVE is shared by the dif-
ferent human and animal prototypes as well as
the vehicle prototype, but the vehicle prototype
lacks the +ANIM feature, and only the bun-
dle on human prototypes (<Hprof>, <Hfam>,
<Hideo>,...) shares the +HUM feature (human
professional, human family, human follower of a
theory/belief/conviction/ideology). In the parser,
a rule selecting the +MOVE feature (e.g. for sub-
jects of movement verbs) will help discard com-
peting senses from lemmas with the above proto-
types, since they will all inherit choices based on
the shared atomic feature. Furthermore, atomic
features can themselves be subjected to inheri-
tance rules, e.g. +HUM −> +ANIM −> +CON-
CRETE, or +MOVE−> +MOVABLE. In Table 1,
which contains examples of polysemic institution
nouns, positive features are marked with capital
letters, negative features with small letters1. The
words in the Table 1 are ambiguous with regard
to the feature H, and since it is only the <inst>
prototype that contributes the +HUM feature po-
tential, it can be singled out by a rule selecting
’H’ or by discarding ’h’. The parser’s about 140
prototypes have been manually implemented for a
lexicon of about 35.000 nouns. In addition, the
±HUM category was also introduced as a selec-
tion restriction for 2.000 verb senses (subject re-
striction) and 1.300 adjective senses (head restric-
tion).
While the semantic annotation of common
nouns is carried out by disambiguating a given
lemma’s lexicon-listed prototype potential, this
strategy is not sufficient for proper nouns, due
1furn=furniture, con=container, inst=institution
Ee = entities (±CONCRETE)
Jj = ±MOVABLE
Hh = ±HUMAN ENTITY
Mm = ±MAS
Ll = ±LOCATION
polysemy spectrum
Ee j Hh m Ll faculdade
E H L <inst> univ. faculty
e h l <f-c> property
Ee j Hh m Ll fundo
e h L <Labs> bottom
E H L <inst> foundation
e h l <ac> <smP> funds
Ee j Hh Mm Ll ind´ustria
E H m L <inst> industry
e h M l <am> diligence
E Jj Hh m L rede
J h <con> net
j H <inst> <+n> network
J h <furn> hammock
Table 1: Feature bundles in prototype based poly-
semy
to the productive nature of this word class. In
two recent NER projects, the parser was aug-
mented with a pattern recognition module and a
rule-based module for identifying and classify-
ing names. In the first project (Bick, 2003),
6 main classes with about 15 subclasses were
used in a lexeme-based approach, while the
second adopted the 41 largely functional cate-
gories of Linguateca’s joint HAREM evaluation
in 2005 (http://www.linguateca.com). A lexicon-
registered name like Berlin would have a stable
tag (<civ> = civitas) in the first version, while
it would be tagged as either <hum>, <top> or
<org> in the second, dependent on context. At
the time of writing, we have not yet tagged our
anaphora corpus with name type tags, and it is
unclear which approach, lexematic or functional,
will work best for the resolution of indirect and
associative anaphora.
3 Indirect Anaphora Resolution
Our work was based on a corpus formed by 31
newspaper articles, from Folha de S˜ao Paulo, writ-
ten in Brazilian Portuguese. The corpus was au-
tomatically parsed using the parser PALAVRAS,
and manually annotated for anaphoricity using
the MMAX tool(http://mmax.eml-research.de/) .
Four subjects annotated the corpus. All annota-
tors agreed on the antecedent in 73% of the cases,
in other 22% of the cases there was agreement be-
tween three annotators and in 5% of the cases only
two annotators agreed. There were 133 cases of
77
definite Indirect anaphors (NPs starting with def-
inite articles) from the total of 1454 definite de-
scriptions (near to 10%) and 2267 NPs.
The parser gives to each noun of the text (or to
most of them) a semantic tag. For instance, the
noun japonˆes [japanese] has the following seman-
tic tags ling and Hnat, representing the features:
human nationality and language respectively.
<word id="word_28">
<n can="japonˆes" gender="M" number="S">
<secondary_n tag="Hnat"/>
<secondary_n tag="ling"/>
</n>
</word>
The approach consists in finding relationships
with previous nouns through the semantic tags.
The chosen antecedent will be the nearest expres-
sion with the largest number of equal semantic
tags. For instance, in the example below, the
anaphor is resolved by applying this resolution
principle, to japonˆes - a l´ıngua.
O Eurocenter oferece cursos de japonˆes em Kanazawa.
Ap´os um mˆes, o aluno falar´a modestamente a l´ıngua.
The Eurocenter offers Japanese courses in Kanazawa. Af-
ter one month, a student can modestly speak the language.
As both expressions (japanese and language)
hold the semantic tag “ling” the anaphor is re-
solved. For the experiments, we considered as cor-
rect the cases where the antecedent found automat-
ically was the same as in the manual annotation
(same), and also the cases in which the antecedent
of the manual annotation was found further up in
the chain identified automatically (in-chain). We
also counted those cases in which the antecedent
of the manual annotation was among the group of
candidates sharing the same tags (in-candidates),
but was not the chosen one (the chosen being the
nearest with greater number of equal tags).
Indirect anaphora
Results # % of Total
Same 25 19%
In-chain 15 11%
Total Correct 40 30%
In-candidates 9 7%
Unsolved 40 30%
Error 44 33%
Total 133 100%
Table 2: Indirect anaphor resolution
Table 2 shows the results of the indirect anaphor
resolution. In 19% of the cases, the system found
the same antecedent as marked in the manual an-
notation. Considering the chain identified by the
system the correct cases go up to 30%. The great
number of unsolved cases were related to the fact
that proper names were not tagged. Considering
mainly the tagged nouns (about 93 cases), the cor-
rect cases amount to 43%). This gives us an idea
of the quality of the tags for the task. We further
tested if increasing the weight of more specific
features in opposition to the more general ones
would help in the antecedent decision process. A
semantic tag that is more specific receives a higher
weight The semantic tag set has three levels, level
1, which is more general receives weight 1, level 2
receives 5, and level 3 receives 10. See the exam-
ple below.
<A> 1 Animal, umbrella tag
<AA> 5 Group of animals
<Adom> 10 Domestic animal
In this experiment the chosen candidate is the
nearest one whose sum of equal tag values has
higher weight. Table 3 shows just a small im-
provement in the correct cases. If we do not
consider unsolved cases, mostly related to proper
names, indirect anaphors were correctly identified
in 46% of the cases (43/96).
Indirect anaphora
Results # % of Total
Same 24 18%
In-chain 19 14%
Total Correct 43 32%
In-candidates 6 5%
Unsolved 40 30%
Error 44 33%
Total 133 100%
Table 3: Indirect anaphor - weighting schema
Since there is no semantic tagging for proper
names as yet, the relationship between pairs such
as S˜ao Carlos - a cidade [S˜ao Carlos - the city]
could not be found. Regarding wrong antecedents,
we have seen that some semantic relationships are
weaker, having no semantic tags in common, for
instance: a proposta - o aumento [the proposal -
the rise]. In some cases the antecedent is not a
previous noun phrase but a whole sentence, para-
graph or disjoint parts of the text. As we con-
sider only relations holding between noun phrases,
these cases could not be resolved. Finally, there
are cases of plain heuristic failure. For instance,
establishing a relationship between os professores
78
[the teachers], with the semantic tags H and Hprof,
and os politicos [the politicians], with the seman-
tic tags H and Hprof, when the correct antecedent
was os docentes [the docents], with the semantic
tags HH (group of humans) and Hprof.
4 Final Remarks
Previous work on nominal anaphor resolution has
used lexical knowledge in different ways. (Poe-
sio et al., 1997) presented results concerning the
resolution of bridging definitions, using the Word-
Net (Fellbaum, 1998), where bridging DDs en-
close our Indirect and Associative anaphora. Poe-
sio et al. reported 35% of recall for synonymy,
56% for hypernymy and 38% for meronymy.
(Schulte im Walde, 1997) evaluated the bridg-
ing cases presented in (Poesio et al., 1997), on
the basis of lexical acquisition from the British
National Corpus. She reported a recall of 33%
for synonymy, 15% for hypernymy and 18% for
meronymy. (Poesio et al., 2002) considering syn-
tactic patterns for lexical knowledge acquisition,
obtained better results for resolving meronymy
(66% of recall). (Gasperin and Vieira, 2004)
tested the use of word similarity lists on resolv-
ing indirect anaphora, reporting 33% of recall.
(Markert and Nissim, 2005) presented two ways
(WordNet and Web) of obtaining lexical knowl-
edge for antecedent selection in coreferent DDs
(Direct and Indirect anaphora). Markert and
Nissim achieved 71% of recall using Web-based
method and 65% of recall using WordNet-based
method. We can say that our results are very sat-
isfactory, considering the related work. Note that
usually evaluation of bridging anaphora is made
on the basis of a limited number of cases, because
the data is sparse. Our study was based on 133
examples, which is not much but surpasses some
of the previous related work. Mainly, our results
indicate that the semantic tagging provided by the
parser is a good resource for dealing with the prob-
lem, if compared to other lexical resources such as
WordNet and acquired similarity lists. We believe
that the results will improve significantly once se-
mantic tags for proper names are provided by the
parser. This evaluation is planned as future work.
Acknowledgments
This work was partially funded by CNPq.

References
Eckhard Bick. 2000. The Parsing System PALAVRAS:
Automatic Grammatical Analysis of Protuguese in
a Constraint Grammar Framework. Ph.D. thesis,
Arhus University, Arhus.
Eckhard Bick. 2003. Multi-level ner for portuguese in
a cg framework. In Nuno J. et al. Mamede, editor,
Computational Processing of the Portuguese Lan-
guage (Procedings of the 6th International Work-
shop, PROPOR 2003), number 2721 in Lecture
Notes in Computer Science, pages 118–125, Faro,
Portugal. Springer.
Razvan Bunescu. 2003. Associative anaphora reso-
lution: A web-based approach. In Proceedings of
the Workshop on The Computational Treatment of
Anaphora - EACL 2003, Budapest.
Christiane Fellbaum, editor. 1998. WordNet: An Elec-
tronic Lexical Database. MIT Press, Cambridge,
MA.
Caroline Gasperin and Renata Vieira. 2004. Us-
ing word similarity lists for resolving indirect
anaphora. In Proceedings of ACL Workshop on Ref-
erence Resolution and its Applications, pages 40–
46, Barcelona.
Katja Markert and Malvina Nissim. 2005. Comparing
knowledge sources for nominal anaphora resolution.
Computational Linguistics, 31(3):367–401.
Massimo Poesio, Renata Vieira, and Simone Teufel.
1997. Resolving bridging descriptions in un-
restricted texts. In Proceedings of the Work-
shop on Operational Factors In Practical, Robust,
Anaphora Resolution for Unrestricted Texts, pages
1–6, Madrid.
Masimo Poesio, Ishikawa Tomonori, Sabine Shulte im
Walde, and Renata Vieira. 2002. Acquiring lexical
knowledge for anaphora resolution. In Proceedings
of 3rd Language resources and evaluation confer-
ence LREC 2002, Las Palmas.
Eleanor Rosch. 1978. Principles of categorization.
In E. Rosch and B. Lloyd, editors, Cognition and
Categorization, pages 27–48. Hillsdale, New Jersey:
Lawrence Erlbaum Associate.
Sabine Schulte im Walde. 1997. Resolving Bridging
Descriptions in High-Dimensional Space Resolving
Bridging Descriptions in High-Dimensional Space.
Ph.D. thesis, Institut f¨ur Maschinelle Sprachverar-
beitung, Universit¨at Stuttgart, and Center for Cogni-
tive Science, University of Edinburgh, Edinburgh.
