TOWARDS AN ANNOTATED DATABASE FOR ANAPHORA 
RESOLUTION 

R.DELMONTE, L.CHIRAN, C.BACALU 
Laboratory of Computational Linguistics 
Ca' Garzoni­Moro, San Marco 3417 
Universit "Ca Foscari" 
30124 ­ VENEZIA 
Tel. 39­41­2578464/52/19 
E­mail: delmont@unive.it 

ABSTRACT 

The basic data used in the research was made available by the GETARUN program developed at the 
University Ca' Foscari of Venezia [1] , which implements a parser in order to determine the morpho -- syntactic 
characteristics of any word from a text. We used a text taken from four Italian journals. The text is composed of 
different parts and contains over 40,000 words. 

The structure of the paper is as follows: Introduction; Section 2 -- Clitics and its Theoretical features; 
Section 3 -- Clitics and Database creation; Section 4 -- Results; Section 5 ­ Conclusions. In a separate Appendix 
we report excerpts from the Anaphoric Relations Database. 

1. INTRODUCTION 

From suggestions offered by research people who had already worked on this topic [M.Poesio, personal 
communication] we decided to limit our scope to only one type of anaphoric pronoun: the morphologically 
expressed or unexpressed class of clitic pronouns. These pronouns are different from the rest of the rich 
inventory available in Italian ­ and in the Romance language Romanian, which will eventually have to be 
compared to the current one ­ because they are monosyllabic, unstressed and fairly commonly used ­ but see the 
figures below. Personal emphatic pronouns, possessive pronouns, deictic pronouns, and quantified expressions ­ 
leaving aside interrogative and relative pronouns which require a completely different treatment in the syntax ­ 
have a different status and may alter the overall framework. So we decided to keep our database as much 
homogeneous as possible: but this notwithstanding, a number of preliminary decisions had to be taken already 
from the start. When we chose clitics as a comprehensive category we had first of all to decide whether we 
wanted to include the empty subject pronoun or not in the list of anaphoric relations to be studied. This was an 
easy decision to take, given the fact that morphologically expressed clitics are the morphological spellout of the 
antecedent together with some semantic information related to animacy and semantic role inherited from the 
governing predicate. So we had to conclude that the little_pro, i.e. the morphologically unexpressed subject 
pronoun used in Italian texts was an empty clitic lacking in one feature (gender) in case the main verb was not a 
compound, because in that case the past participle would have indicated that too. 
However, we had to exclude from the study sentential anaphors expressed in the form of deictic 
pronouns but also clitic pronouns; empty pronominals as placeholders for elliptical material. 

2. Clitics and their Theoretical features 

Clitics have been the subject of much discussion in the linguistics literature, as to their definition which is 
still a topic of debate. A possible definition might be the followin one: 
A clitic is a language element with wordlike status or form that resembles a word. 
A clitic usually cannot be used on its own as a word in a construction. Clitics are usually phonologically 
bound to a preceding or a following word. 

. Clitics are usually associated with certain properties such as the fact that they occur in a special position 
within a sentence. The following Italian example shows this: 
. Maria lo vede. / Mario sees it/him 
. In this case, the accusative clitic lo occurs on the left of the finite verb, while usually direct objects occur on 
the right. 
. Clitics cannot occur alone, they need a host to attach to; in most Romance languages this host is represented 
by the verb. 
. Clitics occur in a fixed order and this order is usually different from that of the corresponding full phrases: 
. Maria me lo spiega. / Mario to me it explains 
. In this example me represents a dative clitic and lo an accusative clitic. In Italian, the direct object usually 
precedes the indirect object if they are both represented by full phrases, while the example above shows that 
this is not the case if clitics are present. 
. Another property, which is usually associated with clitics, is that they cannot be stressed. 
. These properties are considered the basic characteristics, which distinguish clitics from other elements. 

A problem which is still open is the one related to the status of clitics, namely the question of whether they 
should be treated as affixes or as independent syntactic forms. This is because clitics have a status which is 
neither obviously that an independent word, nor that of an affix. 
Within early works in generative grammar, such as that of Kayne(1975), the assumption that clitics are 
syntactically independent elements is not questioned. More generally, the problematic status of clitics with 
respect to the interaction of syntax, morphology was to the large extent neglected. It is only with appearance of 
Zwicky (1977) that clitics are looked at from a broader perspective and that a classification of clitic types which 
takes into account their various syntactic, morphological and phonological properties is proposed. 

2.3. Types of clitics for Italian 
As to their case and related syntactic constituency, Italian clitics can be classified as follows: 
. Nominative 
. Accusative 
. Dative 
. Partitive or Genitive 
. Locative or Ablative 
As to their semantic import, Italian clitics can be classified as follows: 
. Impersonal 
. Reflexive 
. Pleonastic 
. Expletive

2.4. The analysis of si 
One particular type of clitic which occurs very frequently in our database is the ``si''­clitic. Its behavior is 
in many respects different from that of the other clitics. ``Si'' can have one of the following interpretations: 
1. Reflexive/reciprocal interpretation 
Example: 
Maria si lava. / Maria is washing herself 
In this case named ``accusative si'', the clitic represents a direct object which is coreferential with the 
subject. 
2. Impersonal interpretation 
Example: 
Si vedono le montagne. / One sees the mountains 
It is understood as a generic quantified subject, therefore it is ``nominative si'' 3rd person plural, 
+human. 
3. Dative­ Ethic ­ Beneficiary si 
Example: 
Maria si lava le mani. / Maria is washing her hands. 
4. Middle interpretation 
Example: 
Le fragole si mangiano spesso. / Strawberries are often eaten 
This construction seems like a passive ­ the property that the direct object functions as the superficial 
subject. In this case, the clitic cannot be interpreted as an argument, but it has a passivizing function. 
5. The ergative interpretation 
Example: 
La bottiglia si e' rotta. / The bottle broke 
The clitic doesn't reduce the subcategorization requirements of the verb, but it is just a marker of the 
ergative verb form. 
6. Inherent reflexive interpretation 
Example: 
Gino si e' arrabbiato. / John got angry 
In this case the clitic doesn't represent an argument of the verb; ``si'' is simply a marker of 
reflexivization. 
An interesting case is the following: 
``Si sono lavati due bambini'' / Two children were being washed ­ Two children were washing/each other 
When the phrase begins with ``si'' there is an ambiguous situation. This clitic can have an impersonal 
interpretation, but also can have a reflexive/reciprocal interpretation. To solve the ambiguity we can prepose the 
subject NP: 
``Due bambini si sono lavati'' 
the ambiguity is still present, ``si'' can have a reciprocal interpretation -- one child washed the other, and a non -- 
reciprocal interpretation -- every child washed himself. 

3. Anaphora and Database creation 

3.1 The structure of our database 
Our database is made up of the following relations: 
Tokens -- contains the words of our text in the order of their appearance and information about the place of 
the word in the text and the place of the word in the sentence; 
Sentences -- contains the sentences of our text; 
Anaphora -- is the relation where the anaphora links are stored; 
SnSentences -- contains the syntactic analyses of the sentences of our text ; 
Feats -- contains all the possible morphological features of every word of our text, 
Another two sub­relations have been defined: 
Clit -- contains only those tokens that have the morphological characteristic clit; 
Clitvt -- contains only those tokens that have the morphological tag vt, vsup, vc, vin or vcl, i.e. that of the 
verb with the morphological features to be inherited by the empty subject clitic. 
We envisage to port the database under XML as soon as enough data are available to represent syntactic 
constituency at sentence level. 
Here below we shall present the main relations and joins between relations. 
3.1.1. The relation ``Tokens'' 
The relation ``Tokens'' contains all words of the text. It has the following structure: 
Nfrase represents the number of the sentence in which the token occurs; 
PosSentence represents the token's position in the corresponding sentence, as a number; 
Ntoken is an unique number and represents the token's position in the whole text ; 
Token is a text field and contains the token as a string: 
Tag is the field that contains the morphological characteristic of the token: 
CarSin contains the syntactic characteristic of the token; 

Sentence is a calculated field from the relation ``Sentence'' that gives the whole sentence in which the token 
occurs, as a string; 
Feats contains the morphological features of the token, such as: number, genre, etc. This is obtained, using the 
information from ``Tag'', from the relation ``Feats'' where there are all the possible morphological features for 
this token; 
ConstSint gives the token's syntactic constituent. 
Another important aspect is constituted by the joins between this relation and other relations in the database. So 
there are relationships between ``Tokens'' and ``Serntence'', ''Anaphora'' and ``Feats''. The fields that holds these 
relationships are ``NSentence'', ''Ntoken'' and ``Token''. 
3.1.2. The relation ``Sentence'' 
This relation contains the sentences from our text. 
NSentence contains the unique number of the sentence 
CountWSentence represents the number of tokens that occur in the sentence 
Sentence contains the whole sentence as a string 
3.1.3. The relation ``Anaphora'' 
This relation was projected for the anaphoric relations. 
The meaning of the fields is as follows : 
NAntecedent represents the unique number of the token which is the antecedent for the clitic; 
NAnaphor represents the unique number of the token which is the clitic (anaphor); 
Dist is a calculated field (NAnahor -- NAntecedent) and it gives the distance between the antecedent and the 
anaphor in number of words; 

FDist is also a calculated field using the ``Anaphora'' 's join with ``Sentence'', and it gives the distance between 
the anaphor and its antecedent in number of sentences. 
This relation has two joins with ``Tokens'' . Those joins are : 
3.2. The database scheme 
DataBase { TOKENS (NSentence , Pos Sentence, Ntoken ,Token, Tag, CarSin,Sentence , Feats) 
SENTENCES (NSentence, CountWSentence,Sentence) 
ANAPHORA (NAntecedent, NAnaphor, Dist, FDist ) 
FEATS (Nr, Token, Feat) 
SNSENTENCES (AnalyseFrasi, NSentence) } 
There are another two relations (sub­relation) ``Clit'' and ``Clitv'', which are two projections for ``Tokens'' 
whenever an empty subject clitic has been individuated. They have the same structure, but the criteria of the 
selections were ``Tag = Clit'' and ``(Tag = vcl) or (Tag =vt) or (Tag = vin) or (Tag = vsup) or (Tag = vc) or (Tag 
= ause) or (Tag = ausa)'' where the corresponding categories are the verbal tags which contained the 
morphological features to be associated to the empty clitic. 
Our database has primary keys and external keys: 
Relation ``Tokens'' : the primary key is ``Ntoken'' and it is also an external key, used for the join with 
``Anaphora''; 
Relation `` Sentence '' : the primary key is ``NSentence ''. ``N Sentence '' is also an external key for the joins 
with ``Tokens'' and ``Sn Sentence ''; 
Relation ``Feats'' :''Nr'' is the primary key and ``Token'' is the external key that serves for the join with 
``Tokens''; 

Relation ``SnSentence '' : ``NSentence '' is the primary key and the external key for the join with `` Sentence''. 

3.4 The E­R Model 
Between Tokens and Clit and Clitv there exists a relation and its type is part­of. 

4. Results 

5. CONCLUSIONS 

APPENDIX