File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/01/j01-4005_metho.xml
Size: 49,502 bytes
Last Modified: 2025-10-06 14:07:31
<?xml version="1.0" standalone="yes"?> <Paper uid="J01-4005"> <Title>An Algorithm for Anaphora Resolution in Spanish Texts</Title> <Section position="4" start_page="546" end_page="548" type="metho"> <SectionTitle> 3. Anaphora Resolution Algorithm </SectionTitle> <Paragraph position="0"> In the algorithm, all the types of anaphora are identified from left to right as they appear in the sentence. The most important proposals for anaphora resolution--such as those of Baldwin (1997), Lappin and Leass (1994), Hobbs (1978), or Kennedy and Boguraev (1996)--are based on a separation between constraints and preferences.</Paragraph> <Paragraph position="1"> Computational Linguistics Volume 27, Number 4 Constraints discard some of the candidates, whereas preferences simply sort the remaining candidates. A constraint defines a property that must be satisfied in order for any candidate to be considered as a possible solution of the anaphor. For example, pronominal anaphors and antecedents must agree in person, gender, and number. 3 Otherwise, the candidate is discarded as a possible solution. A preference is a characteristic that is not always satisfied by the solution of an anaphor. The application of preferences usually involves the use of heuristic rules in order to obtain a ranked list of candidates.</Paragraph> <Paragraph position="2"> Each type of anaphora has its own set of constraints and preferences, although they all follow the same general algorithm: constraints are applied first, followed by preferences.</Paragraph> <Paragraph position="3"> Based on the preceding description, our algorithm contains the following main components: * identification of the type of pronoun * constraints -- morphological agreement (person, gender, and number) -- syntactic conditions on NP-pronoun noncoreference * preferences In order to apply this algorithm to unrestricted texts, it has been necessary to use partial parsing. In our partial-parsing scheme, as presented in Ferr~ndez, Palomar, and Moreno (1999), we only parse coordinated NPs and PPs, verbal chunks, pronouns, and what we have called free conjunctions (i.e., conjunctions that do not join coordinated NPs or PPs). Words that do not appear within these constituents are simply ignored. The NP constituents include coordinated adjectives, relative clauses, coordinated PPs, and appositives as modifiers.</Paragraph> <Paragraph position="4"> With this partial-parsing scheme, we divide a sentence into clauses by parsing first the free conjunction and then the verbs, as in the following example: (8) Pedro compr6 un regalo y se lo dio a Ana.</Paragraph> <Paragraph position="5"> Pedro bought a gift and her it gave to Ana 'Pedro bought a gift and gave it to Ana.' In this example, we have parsed the following constituents: np(Pedro), v(comprO), np(un regalo),freeconj(y), pron(se), pron(lo), v(dio), pp(a Ana). We are able to divide this sentence into two clauses because it contains the free conjunction y 'and' and the two verbs compr6 'bought' and clio 'gave'.</Paragraph> <Section position="1" start_page="547" end_page="548" type="sub_section"> <SectionTitle> 3.1 Identification of the Kind of Pronoun </SectionTitle> <Paragraph position="0"> The algorithm uses partial-parse trees to automatically identify omitted pronouns by employing the following steps: * The sentence is divided into clauses (by parsing the free conjunction followed by the verbs).</Paragraph> <Paragraph position="1"> Palomar et al. Anaphora Resolution in Spanish Texts An NP or pronoun is sought for each clause by analyzing the clause constituents on the left-hand side of the verb, unless the verb is imperative or impersonal. The chosen NP or pronoun must agree in person and number with the clausal verb. (In evaluating this algorithm, Ferr~ndez and Peral \[2000\] achieved a success rate of 88% for detecting omitted pronouns.) The remaining pronouns are identified based on part-of-speech (POS) tagger outputs. null</Paragraph> </Section> <Section position="2" start_page="548" end_page="548" type="sub_section"> <SectionTitle> 3.2 Morphological Agreement </SectionTitle> <Paragraph position="0"> Person, gender, and number agreement are checked in order to discard potential antecedents. For example, in the sentence (9) Juanj vio a Rosa/. Ella/ estaba muy feliz.</Paragraph> <Paragraph position="1"> Juanj saw to Rosa/ Shei was very happy 'Juan saw Rosa. She was very happy.' there are two possible antecedents for ella 'she', whose slot structures 4 are np (conc (sing, masc), X, Juan) np (conc (sing, fem), Y, Rosa) whereas the slot structure of the pronoun is pron (conc (sing, fem), Z, ella).</Paragraph> <Paragraph position="2"> In order to decide between the two antecedents, the unification of both slot structures (pronoun and candidate) is carried out by the slot unification parser (Ferr~ndez, Palomar, and Moreno 1999). In this example, the candidate Juan is rejected by this morphological agreement constraint.</Paragraph> </Section> <Section position="3" start_page="548" end_page="548" type="sub_section"> <SectionTitle> 3.3 Syntactic Conditions on NP-Pronoun Noncoreference </SectionTitle> <Paragraph position="0"> These conditions are based on c-command and minimal-governing-category constraints as formulated by Reinhart (1983) and on the noncoreference conditions of Lappin and Leass (1994). They are of great importance in any anaphora resolution system that does not use semantic information, as is the case with our proposal. In such systems, recency is important in selecting the antecedent of an anaphor. That is to say, the closest NP to the anaphor has a better chance of being selected as the solution. One problem, however, is that such constraints are formulated using full parsing, whereas if we want to work with unrestricted texts we should be using partial parsing, as previously defined.</Paragraph> <Paragraph position="1"> We have therefore proposed a set of noncoreference conditions for Spanish, using partial parsing, although they could easily be extended to other languages such as English. In our system, the following types of pronouns are noncoreferential with a noun phrase (NP) under the conditions noted (noncoindexing indicates that a candidate is rejected by these conditions).</Paragraph> </Section> </Section> <Section position="5" start_page="548" end_page="554" type="metho"> <SectionTitle> 4 The term slot structure is defined in Ferr~ndez, Palomar, and Moreno (1998b). The slot structure stores </SectionTitle> <Paragraph position="0"> morphological and syntactic information related to the different constituents of a sentence.</Paragraph> <Paragraph position="1"> the NP is included in another constituent (e.g., the NP is included in a PP) Ante Luisj sei frot6 con la toalla.</Paragraph> <Paragraph position="2"> in front of Luisj himself/ rubbed with the towel 'He rubbed himself with the towel in front of Luis.' In this sentence, we would have obtained the following sequence of constituents after our partial-parsing scheme: pp(prep(ante), np(Luis ) ) , pron(se) , v(frot6 ) , pp(prep( con) , np(la toalla) ). Following the above-stated condition, the NP Luis cannot corefer with the reflexive pronoun se since Luis is included in a PP (ante Luis). the NP is in a different clause or sentence (11) Anaj trajo un cuchillo y Eva/ sei cort6.</Paragraph> <Paragraph position="3"> Anaj brought a knife and Eva/ herself/ cut 'Ana brought a knife and Eva cut herself.' the NP appears after the verb and there is another NP in the same clause before the verb</Paragraph> <Paragraph position="5"> Juan/ sei cort6 con el cuchilloj.</Paragraph> <Paragraph position="6"> Juan/ himself/ cut with the knifej 'Juan cut himself with the knife.' Under these conditions, coreference is allowed between the NP and the reflexive pronoun, since both are in the same clause. For example: Juan/ queria verlo por s~ mismoi.</Paragraph> <Paragraph position="7"> Juan/ wanted see it for himself/ 'Juan wanted to see it for himself.' In this example, Juan and the reflexive pronoun si mismo 'himself' corefer since Juan is in the same clause as the anaphor, it is not included in another constituent, and it appears before the verb.</Paragraph> <Paragraph position="8"> Clitic pronouns are noncoreferential when: the NP is included in a PP (except those headed by the preposition a 'to') (14) Con Juan/ loj compr6.</Paragraph> <Paragraph position="9"> with Juan/ itj bought 'I bought it with Juan.' the NP is located more than three constituents before the clitic pronoun in the same clause (15) En casai \[el martillo\]j no se loj di. at home/ \[the hammer\]j not him itj gave 'I didn't give him the hammer at home.' In this example, the direct object el martillo 'the hammer' has been moved from its common position after the verb, and it is necessary to fill the resulting gap with the pronoun lo 'it' even though it does not appear in the English translation. This phenomenon 5 can be considered an exception to the c-command constraints as formulated by Reinhart when applied to Spanish clitic pronouns.</Paragraph> <Paragraph position="10"> Moreover, if the last two conditions are not fulfilled by the NP and the verb is in the first or second person, then this NP will necessarily be the solution of the pronoun: (16) \[El boligrafo\]i 1Oi comprar~s en esa tienda. \[The pen\]/ iti will buy in that shop 'You will buy the pen in that shop.' Personal and demonstrative (nonclitic) pronouns are noncoreferential when the NP is in the same clause as the anaphor, and: the pronoun comes before the verb (in full parsing, this would mean that it is the subject of its clause) Ante Luisi 61j salud6 a Pedrok.</Paragraph> <Paragraph position="11"> in front of Luisi hey greeted to Pedrok 'He greeted Pedro in front of Luis.' the pronoun comes after the verb (in full parsing, this would mean that it is the object of the verb) and the NP is not included in another NP (18) \[El padre de Juanj\]i le venci6 a 41j.</Paragraph> <Paragraph position="12"> \[Juanj's father\]/ him beat to himj 'Juan's father beat him.' In this example, the pronoun ~I 'him' cannot corefer with the NP el padre de Juan 'Juan's father', but it can corefer with Juan since it is a modifier of the NP el padre de Juan.</Paragraph> <Paragraph position="13"> It should be mentioned that the clitic pronoun le is another form of the pronoun dl 'him'. This is a typical phenomenon in Spanish, where clitic pronouns occupy the object position. Sometimes both the clitic pronoun and the object appear in the same clause, as occurs in the previous example and in the following one: (19) A Pedro/ yo lei vi ayer.</Paragraph> <Paragraph position="14"> to Pedroj I himi saw yesterday</Paragraph> <Paragraph position="16"> This example also illustrates the previously mentioned exception of c-command constraints for Spanish clitic pronouns. In this case, the direct object a Pedro 'to Pedro' has been moved before the verb, and the clitic pronoun le 'him' has been added. It should also be remarked that, as noted earlier, the clitic pronoun does not appear in the English translation.</Paragraph> <Paragraph position="17"> 5 Mathews (1997) calls this phenomenon &quot;clitic doubling&quot; and defines it as the use of a clitic pronoun with the same referent and in the same syntactic function as another element in the same clause. Computational Linguistics Volume 27, Number 4 (c) the pronoun is included in a PP that is not included in another constituent and the NP is not included in another constituent (NP or PP) (20) \[El padre de Luisj\]i juega con 61j.</Paragraph> <Paragraph position="18"> \[Luisj's father\]/ plays with himj 'Luis's father plays with him.' In this example, the pronoun ~I 'him' is included in a PP (which is not included in another constituent) and the NP el padre de Luis is not included in another NP or PP. Therefore, the NP cannot corefer with the pronoun. However, the NP Luis can corefer because it is included in the NP el padre de Luis. (d) the pronoun is included in an NP, so that the NP in which the pronoun is included cannot corefer with the pronoun (21) Pedro/ vio \[al hermano de ~li\] j.</Paragraph> <Paragraph position="19"> Pedro/ saw \[the brother of himi\]j 'Pedro saw his brother.' (e) the pronoun is coordinated with other NPs, so that the other coordinated NPs cannot corefer with the pronoun (22) Juan/, \[el tio de Ana\]j, y 61k fueron de pesca.</Paragraph> <Paragraph position="20"> Juan/, \[Ana's uncle\]j, and hek went fishing 'He, Juan, and Ana's uncle went fishing.' (f) the pronoun is included in a relative clause, and the following condition is met: .</Paragraph> <Paragraph position="21"> (24) i. the NP in which the relative clause is included does not corefer with the pronoun (23) Pedroj vio a \[un amigo que juega con 41j\]i.</Paragraph> <Paragraph position="22"> Pedroj saw to \[a friend that plays with himj\]i 'Pedro saw a friend that he plays with.' ii. the NPs that are included in the relative clause follow the previous conditions iii. the remaining NPs outside the relative clause could corefer with the pronoun Personal and demonstrative (nonclitic) pronouns are noncoreferential when the NP is not in the same clause as the pronoun. (In this case, the NP can corefer with the pronoun, except when this NP also appears in the same sentence and clause as the pronoun, in which case it will have been discarded by the previous noncoreference conditions.) Anaj y Evai son amigas. Evai lej ayuda mucho.</Paragraph> <Paragraph position="23"> Anay and Evai are friends Evai herj helps a lot 'Ana and Eva are friends. Eva helps her a lot.' It is important to note that the above-mentioned conditions refer to those coordinated NPs and PPs that have been partially parsed. Moreover, as previously mentioned, NPs can include relative clauses, appositives, coordinated PPs, and adjectives. Palomar et al. Anaphora Resolution in Spanish Texts We should also remark that we consider a constituent A to be included in a constituent B if A modifies the head of B. Let us consider the following NP: (25) \[el hombre que ama a \[una mujer que lei ama\]j\]i \[the man who loves to \[a woman who him/ loveslj\]i 'the man who loves a woman who loves him.' We consider that the pronoun le 'him' is included in the relative clause that modifies the NP una mujer que le ama 'a woman who loves him', which then cannot corefer with it due to noncoreference condition 3(f)i. Under condition 3(f)iii, however, the pronoun le 'him' could corefer with the entire NP el hombre que area a una mujer que le area 'the man who loves a woman who loves him'.</Paragraph> <Paragraph position="24"> Another example might be the following: (26) Eva/ tiene \[un tio que lei toma el pelo\]j.</Paragraph> <Paragraph position="25"> Evai has \[an uncle that heri teases\]j 'Eva has an uncle who teases her.' In this example, the pronoun is included within the relative clause that modifies un tio 'an uncle', and therefore cannot corefer with it. But, following condition 3(f)iii, it can corefer with Eva.</Paragraph> <Section position="1" start_page="552" end_page="554" type="sub_section"> <SectionTitle> 3.4 Preferences </SectionTitle> <Paragraph position="0"> To obtain the different sets of preferences, we utilized the training corpus to identify the importance of each kind of knowledge that is used by humans when tracking down the NP antecedent of a pronoun. Our results are shown in Table 1. For our analysis, the antecedents for each pronoun in the text were identified, along with their configurational characteristics with reference to the pronoun. Thus, the table shows how often each configurational characteristic is valid for the solution of a particular pronoun. For example, the solution of a reflexive pronoun is a proper noun 53% of the time. The total number of pronoun occurrences in the study was 575. Thus, we were able to define the different patterns of Spanish pronoun resolution and apply them in order to obtain the evaluation results that are presented in this paper. The order of importance was determined by first sorting the preferences according to the percentage of each configurational characteristic; that is, preferences with higher percentages were applied before those with lower percentages. After several experiments on the training corpus, an optimal order--the one that produced the best performance--was obtained.</Paragraph> <Paragraph position="1"> Since in this evaluation phase we processed texts from different genres and by different authors, we can state that the final set of preferences obtained and their order of application can be used with confidence on any Spanish text.</Paragraph> <Paragraph position="2"> Based on the results presented in Table 1, we have extracted a set of preferences for each type of anaphora (listed below). We have distinguished between those pronouns that are included within PPs and those that are not. That is because when a pronoun is included in a PP, the preposition of this PP sets a preference.</Paragraph> <Paragraph position="3"> Preferences of omitted pronouns (OPR): 1. NPs that are not of time, direction, quantity, or abstract type; that is to say, inanimate candidates are rejected (e.g., hal~past ten, Market Street, three pounds, or a thing) 2. NPs in the same sentence as the omitted pronotm a If the NP b If the NP c If the NP d If the NP e If the NP f If the NP g If the NP h If the NP i If the NP j If the NP k If the NP 1 If the NP m If the NP n If the NP is included in another NP is included in a PP with the preposition en 'in' is a proper noun is an indefinite NP has been repeated more than once in the text has appeared with the verb of the anaphor more than once in the text has appeared in a PP more than once in the text occupies the same position with reference to the verb as the anaphor (before or after) appears before its verb is not a time-type is not a quantity-type is not a direction-type is not an abstract-type is not a company-type 3. NPs that are in the same sentence as the anaphor and are also the solution for another omitted pronotm 4. NPs that are in the previous sentence 5. NPs that are not included in another NP (e.g., when they appear inside a relative clause or appositive) 6. NPs that are not included in a PP or are included in a PP when its preposition is a 'to' or de 'of' 7. NPs that appear before the verb 8. NPs that have been repeated more than once in the text Preferences of clitic personal pronouns (CPPR): 1. NPs that are not of time, direction, quantity, or abstract type 2. NPs that are in the same sentence as the anaphor Palomar et al. Anaphora Resolution in Spanish Texts 3. NPs that are in the previous sentence 4. NPs that are not included in another NP (e.g., when they appear inside a clause or appositive) 5. NPs that are not included in a PP or are included in a PP when its preposition is a 'to' or de 'of' 6. NPs that have appeared with the verb of the anaphor more than once Preferences of personal and demonstrative pronouns that are included in a PP (PPRinPP and DPRinPP): 1. NPs that are not of time, direction, quantity, or abstract type; moreover, in the case of personal pronouns, the NP cannot be a company type 2. NPs that are in the same sentence as the anaphor 3. NPs that are in the previous sentence 4. NPs that are not included in another NP (e.g., when they appear inside a relative clause or appositive) 5. NPs that have been repeated more than once in the text 6. NPs that are included in a PP 7. NPs that occupy the same position (before or after) with respect to the verb as the anaphor Preferences of personal and demonstrative pronouns that are not included in a PP and of reflexive pronouns (PPRnotPP, DPRnotPP, and RPR): 1. NPs that are not of time, direction, quantity, or abstract type; moreover, in the case of personal pronouns, the NP cannot be a company type 2. NPs that are in the same sentence as the anaphor 3. NPs that are in the previous sentence 4. NPs that are not included in another NP (e.g., when they appear inside a relative clause or appositive) 5. NPs that are not included in a PP or that are included in a PP when its preposition is a 'to' or de 'of' 6. For the case of personal pronouns (PPRnotPP), NPs that are not included in a PP with the preposition en 'in' 7. NPs that appear before their verbs (i.e., the verb of the sentence in which the NP appears)</Paragraph> </Section> </Section> <Section position="6" start_page="554" end_page="555" type="metho"> <SectionTitle> 3.5 Resolution Procedure </SectionTitle> <Paragraph position="0"> The resolution procedure consists of the following steps: 2. Identify the NP candidate antecedents of a pronoun in order to create a list L. The list created will depend on the type of anaphor and the anaphoric accessibility space (empirically obtained from a deep study of the training corpus) and will be developed according to the following criteria: * For pronominal anaphora, demonstrative anaphora, and omitted pronouns, NP candidates will appear in the same sentence as the anaphor and in the four previous sentences. * For reflexive anaphora, NP candidates will appear in the same sentence as the anaphor.</Paragraph> <Paragraph position="1"> 3. Apply constraints to L to obtain LI: (a) morphological agreement (b) syntactic conditions on NP-pronoun noncoreference 4. If the number of elements of L1 - 1, then the solution is that element. 5. If the number of elements of L1 = 0, then the solution is an exophor. 6. If the number of elements of L1 > 1, then apply preferences to L1 to obtain L2. Depending on the type of anaphora, a different set and order of preferences will be applied (see Section 3.4).</Paragraph> <Paragraph position="2"> 7. If the number of elements of L2 = 1, then the solution is that element. 8. If the number of elements of L2 > 1, then apply the following three basic preferences in the order shown until only one candidate remains (these three preferences are common to all the pronouns): * NPs most repeated in the text * NPs that have appeared most with the verb of the anaphor * the first candidate of the remaining list (the closest one to the anaphor) After applying these basic preferences, the antecedent is obtained. 4. Empirical Evaluation</Paragraph> <Section position="1" start_page="555" end_page="555" type="sub_section"> <SectionTitle> 4.1 Description of Corpora </SectionTitle> <Paragraph position="0"> We have tested the algorithm on both technical manuals and literary texts. In the first instance, we used a portion of the Spanish edition of the Blue Book corpus. 6 This corpus contains the handbook of the International Telecommunications Union CCITT, published in English, French, and Spanish; it is one of the most important collections of telecommunications texts available and contains 5,000,000 words automatically tagged by the Xerox tagger. In the second instance, the algorithm was tested on Lexesp, a corpus 7 that contains Spanish literary texts from different genres and by different</Paragraph> </Section> </Section> <Section position="7" start_page="555" end_page="563" type="metho"> <SectionTitle> 6 CRATER (Proyecto CRATER 1994-1995) Corpus Resources and Terminology Extraction Project. Project </SectionTitle> <Paragraph position="0"> supported by the European Community Commission (DG-XIII). Computational Linguistics Laboratory, Faculty of Philosophy and Fine Arts, Autonomous University of Madrid, Spain. 7 The Lexesp corpus belongs to the project of the same name carried out by the Psychology Department of the University of Oviedo and developed by the Computational Linguistics Group of the University of Barcelona, with the collaboration of the Language Processing Group of the Catalonia University of Technology, Spain.</Paragraph> <Paragraph position="1"> authors. These texts were mainly obtained from newspapers and were automatically tagged by a different tagger than the one used to tag the Blue Book. The portion of the Lexesp corpus that we processed contained various stories, related by a narrator, and written by different authors. As was the case for the Blue Book corpus, this corpus also contained 5,000,000 words. Since we worked on texts from different genres and by different authors, the applicability of our proposal to other kinds of texts is assured.</Paragraph> <Paragraph position="2"> We selected a subset of the Blue Book corpus and another subset of the Lexesp corpus, and both were annotated with respect to coreference. One portion of the coreferentially tagged corpus (training corpus) was used for improving the rules for anaphora resolution (constraints and preferences), and another portion was reserved for test data (Table 2).</Paragraph> <Paragraph position="3"> The annotation phase was accomplished in the following manner: (1) two annotators were selected, (2) an agreement was reached between the annotators with regard to the annotation scheme, (3) each annotator annotated the corpus, and, finally, (4) a reliability test (Carletta et al. 1997) was done on the annotation in order to guarantee the results. The reliability test used the kappa statistic that measures agreement between the annotations of two annotators in making judgments about categories. In this way, the annotation is considered a classification task consisting of defining an adequate solution among the candidate list. According to Vieira (1998), the classification task when tagging anaphora resolution can be reduced to a decision about whether each candidate is the solution or not. Thus, two different categories are considered for each anaphor: one for the correct antecedent and another for nonantecedents. Our experimentation showed one correct antecedent among an average of 14.5 possible candidates per anaphor after applying constraints. For computing the kappa statistic (k), see Siegel and Castellan (1988).</Paragraph> <Paragraph position="4"> According to Carletta et al. (1997), a k measurement such as 0.68 < k < 0.8 allows us to draw encouraging conclusions, and a measurement k > 0.8 means there is total reliability between the results of the two annotators. In our tests, we obtained a kappa measurement of k = 0.81. We therefore consider the annotation obtained for the evaluation to be totally reliable.</Paragraph> <Section position="1" start_page="556" end_page="558" type="sub_section"> <SectionTitle> 4.2 Experimental Work </SectionTitle> <Paragraph position="0"> We conducted a blind test over the entire test corpus of unrestricted Spanish texts by applying the algorithm to the partial syntactic structure generated by the slot unification parser.</Paragraph> <Paragraph position="1"> Over these corpora, our algorithm attained a success rate for anaphora resolution of 76.8%. We define &quot;success rate&quot; as the number of pronouns successfully resolved, divided by the total number of resolved pronouns. The total number of resolved pronouns was 1,677, including personal, demonstrative, reflexive, and omitted pronouns. anaphor as their antecedent. Our algorithm's &quot;recall percentage,&quot; defined as the number of pronouns correctly resolved, divided by the total number of pronouns in the text, was therefore 76.8%. A breakdown of success rate results for each kind of pronoun is also shown in Table 3. The pronouns were classified so as to provide the option of applying different kinds of knowledge to resolve each category of pronoun. One of the factors that affected the results was the complexity of the Lexesp corpus, due mainly to its complex narratives. On average, 16 words per sentence and 27 candidates per anaphor were found in this corpus.</Paragraph> <Paragraph position="2"> In our experiment, a &quot;successful resolution&quot; occurred if the head of the solution offered by our algorithm was the same as that offered by two human experts. We adopted this definition of &quot;success&quot; because it allowed the system to be totally automatic: solutions given by the annotators were stored in a file and were later automatically compared with the solutions given by our system. Since semantic information was not used at all, PP attachments were not always correctly disambiguated. Hence, at times the differences simply corresponded to different subconstituents.</Paragraph> <Paragraph position="3"> After the evaluation process, we tested the results in order to identify the limitations of the algorithm with respect to the resolution process. We identified the following: * There were some mistakes in the POS tagging (causing an error rate of around 3%).</Paragraph> <Paragraph position="4"> * There were some mistakes in the partial parsing with respect to the identification of complex noun phrases (causing an error rate of around 7%) (Palomar et al. 1999).</Paragraph> <Paragraph position="5"> * Semantic information was not considered (causing an error rate of around 32%). An example of this type of error can be seen in the following text extracted from the Lexesp corpus: (27) Recuerdo, pot ejemplo, \[un pequefio claro en un bosque en medio de las montafias canadienses\]i, con tres lagunas diminutas que, a causa de los sedimentos del agua. tenfan distintos y chocantes colores. Esta rareza habia hecho del sitioi un espacio sagrado al que peregrinaron los indios durante siglos y seguramente antes los pobladores paleolfticos. Y eso se notaba.</Paragraph> <Paragraph position="6"> Palomar et al. Anaphora Resolution in Spanish Texts (28) Canad~i es un pals muy hermoso, y aqu41i no era, ni mucho rnenos, el lugar m~s bello: pero guardaba tranquilamente dentro de sf toda su arrnonfa, como los melocotones guardan dentro de sf el duro hueso.</Paragraph> <Paragraph position="7"> '1 remember, for example, \[a small clearing in the woods in the middle of the Canadian mountains\]/, with three tiny lagoons that, due to the water sediments, had different and astonishing colors.</Paragraph> <Paragraph position="8"> This peculiarity had made the place/into a sacred site, to which the Indians made pilgrimages over the centuries, and surely even the Paleolithic Indians before them. And you could feel it.</Paragraph> <Paragraph position="9"> 'Canada is a very beautiful country and that one/was by no means the most beautiful place: but it calmly kept within itself all of its harmony, like peaches that keep the hard seeds within.' In this text, the demonstrative pronoun aqudl 'that one' corefers with the antecedent un peque~o claro en un bosque en medio de las monta~as canadienses 'a small clearing in the woods in the middle of the Canadian mountains', which is also linked to the definite noun phrase el sitio 'the place'. Our algorithm identified the proper noun Canadd, which is in the same sentence, as the anaphor, since the proper noun could only have been discarded by means of semantic information.</Paragraph> <Paragraph position="10"> As an example of an anaphor that was correctly resolved by the algorithm, we present the following sentence extracted from the Blue Book corpus. In this case, the antecedent los sistemas de transmisidn analdgica 'the systems of analogue transmission' was correctly chosen for the personal pronoun ellos 'them': ' En las conexiones largas o de Iongitud media, es probable que la fuente principal de ruido de circuito estribe en \[los sistemas de transmisi6n anal6gica\]i, ya queen ellosi la potencia de ruido suele set proporcional a la Iongitud del circuito.</Paragraph> <Paragraph position="11"> 'In long or medium connections, it is probable that the main source of circuit noise comes from \[the systems of analogue transmission\]/, since in them/the noise capacity is usually proportional to the length of the circuit.' The remainder of the errors were due to split antecedents (10%), cataphora (2%), exophora (3%), or exceptions in the application of preferences (43%).</Paragraph> <Paragraph position="12"> 5. Comparison with Other Approaches to Anaphora Resolution</Paragraph> </Section> <Section position="2" start_page="558" end_page="560" type="sub_section"> <SectionTitle> 5.1 Anaphora Resolution Approaches </SectionTitle> <Paragraph position="0"> Common among all languages is the fact that the anaphora phenomenon requires similar strategies for its resolution (e.g., pronouns or definite descriptions). All languages employ different kinds of knowledge, but their strategies differ only in the manner by which this knowledge is coordinated. For example, in some strategies just one kind of knowledge becomes the main selector for identifying the antecedent, with other kinds of knowledge being used merely to confirm or reject the proposed antecedent.</Paragraph> <Paragraph position="1"> In such cases, the typical kind of knowledge used as the selector is that of discourse structure. Centering theory, as employed by Strube and Hahn (1999) or Okumura and Tamura (1996), uses this type of approach. Other approaches, however, give equal Computational Linguistics Volume 27, Number 4 importance to each kind of knowledge and generally distinguish between constraints and preferences (Baldwin 1997; Lappin and Leass 1994; Carbonell and Brown 1988).</Paragraph> <Paragraph position="2"> Whereas constraints tend to be absolute and therefore discard possible antecedents, preferences tend to be relative and require the use of additional criteria (e.g., the use of heuristics that are not always satisfied by all antecedents). Nakaiwa and Shirai (1996) use this sort of resolution model, which involves the use of semantic and pragmatic constraints, such as constraints based on modal expressions, or constraints based on verbal semantic attributes or conjunctions.</Paragraph> <Paragraph position="3"> Our approach to anaphora resolution belongs in the latter category, since it combines different kinds of knowledge and no knowledge based on discourse structure is included. We choose to ignore discourse structure because obtaining this kind of knowledge requires not only an understanding of semantics but also knowledge about world affairs and the ability to almost perfectly parse any text under discussion (Azzam, Humphreys, and Gaizauskas 1998).</Paragraph> <Paragraph position="4"> Still other approaches to anaphora resolution are based either on machine learning techniques (Connolly, Burger, and Day 1994; Yamamoto and Sumita 1998; Paul, Yamamato, and Sumita 1999) or on the principles of uncertainty reasoning (Mitkov 1995).</Paragraph> <Paragraph position="5"> Computational processing of semantic and domain information is relatively expensive when compared with other kinds of knowledge. Consequently, current anaphora resolution methods rely mainly on constraint and preference heuristics, which employ morpho-syntactic information or shallow semantic analysis (see, for example, Mitkov \[1998\]). Such approaches have performed notably well. Lappin and Leass (1994) describe an algorithm for pronominal anaphora resolution that achieves a high rate of correct analyses (85%). Their approach, however, operates almost exclusively on syntactic information. More recently, Kennedy and Boguraev (1996) proposed an algorithm for anaphora resolution that is actually a modified and extended version of the one developed by Lappin and Leass (1994). It works from the output of a POS tagger and achieves an accuracy rate of 75%.</Paragraph> <Paragraph position="6"> There are other approaches based on POS tagger outputs as well. For example, Mitkov and Stys (1997) propose a knowledge-poor approach to resolving pronouns in technical manuals in both English and Polish. The knowledge employed in these approaches is limited to a small noun phrase grammar, a list of terms, and a set of antecedent indicators (definiteness, term preference, lexical reiteration, etc.).</Paragraph> <Paragraph position="7"> Still other approaches are based on statistical information, including the work of Dagan and Itai (1990, 1991) and Ge, Hale, and Charniak (1998), all of whom present a probabilistic model for pronoun resolution.</Paragraph> <Paragraph position="8"> We have adopted their ideas and adapted their algorithms to partial parsing and to Spanish texts in order to compare our results with their approaches.</Paragraph> <Paragraph position="9"> With reference to the differences between English and Spanish anaphora resolution, we have made the following observations: Syntactic parallelism has played a more important role in English texts than in Spanish texts, since Spanish sentence structure is more flexible than English sentence structure. Spanish is a free-word-order language and has different syntactic conditions, which increases the difficulty of resolving Spanish pronouns (hence, the greater accuracy rate for English texts).</Paragraph> <Paragraph position="10"> * A greater number of possible antecedents was observed for Spanish pronouns than for English pronouns, due mainly to the greater average Palomar et al. Anaphora Resolution in Spanish Texts length of Spanish sentences (which also makes the resolution of Spanish pronouns more difficult).</Paragraph> <Paragraph position="11"> Spanish pronouns usually bear more morphological information. One result is that this constraint tends to discard more candidates in Spanish than in English.</Paragraph> <Paragraph position="12"> For comparison purposes, we implemented the following approaches on the same Spanish texts that were tested and described in Section 4.1.</Paragraph> </Section> <Section position="3" start_page="560" end_page="560" type="sub_section"> <SectionTitle> 5.2 Hobbs's Algorithm </SectionTitle> <Paragraph position="0"> Hobbs's algorithm (Hobbs 1978) is applied to the surface parse trees of sentences in a text. A surface parse tree represents the grammatical structure of a sentence. By reading the leaves of the parse tree from left to right, the original English sentence is formed. The algorithm parses the tree in a predefined order and searches for a noun phrase of the correct gender and number. Hobbs tested his algorithm for the pronouns he, she, it, and they, using 100 examples taken from three different sources. Although the algorithm is very simple, it was successful 81.8% of the time.</Paragraph> <Paragraph position="1"> We implemented a version of Hobbs's algorithm for slot unification grammar for Spanish texts. Since full parsing was not done, our specifications for the algorithm were adjusted, as follows: * NPs were tested from left to right, as they were parsed in the sentence.</Paragraph> <Paragraph position="2"> * Afterward, the NPs that were included in an NP (breadth-first) were tested.</Paragraph> <Paragraph position="3"> * This test was interrupted when an NP agreed in gender and number with the anaphor.</Paragraph> <Paragraph position="4"> The problems we encountered in implementing Hobbs's algorithm are similar to those found in implementing other approaches: the adaptation to partial parsing, and the inherent difficulty of the Spanish language (i.e., its free-word-order characteristics). The results of our test of this version of Hobbs's algorithm on the test corpus appear in Table 4.</Paragraph> </Section> <Section position="4" start_page="560" end_page="560" type="sub_section"> <SectionTitle> 5.3 Approaches Based on Constraints and Proximity Preference </SectionTitle> <Paragraph position="0"> Our approach has also been compared with the typical baseline approach consisting of constraints and proximity preference; that is, the antecedent that appears closest to the anaphor is chosen from among those that satisfy the constraints. For this comparison, the same constraints that were used previously (i.e., morphological agreement and syntactic conditions) were applied here. Then the antecedent at the head of the list of antecedents was proposed as the solution of the anaphor. These results are also listed in Table 4. As can be seen from the table, success rates were lower than those obtained through the joint application of all the preferences.</Paragraph> </Section> <Section position="5" start_page="560" end_page="562" type="sub_section"> <SectionTitle> 5.4 Lappin and Leass's Algorithm </SectionTitle> <Paragraph position="0"> An algorithm for identifying the noun phrase antecedents of third person pronouns and lexical anaphors (reflexive and reciprocal) is presented in Lappin and Leass (1994); this algorithm has exhibited a high rate (85%) of correct analyses in English texts. It relies on measures of salience that are derived from syntactic structures and on simple dynamic models of attentional state to select the antecedent noun phrase of a pronoun from a list of candidates.</Paragraph> <Paragraph position="1"> Computational Linguistics Volume 27, Number 4 We have implemented a version of Lappin and Leass's algorithm for Spanish texts. The original formulation of the algorithm proposes a syntactic filter on NP-pronoun coreference. This filter consists of six conditions for NP-pronoun noncoreference within any sentence (Lappin and Leass 1994, page 537). In applying this algorithm to Spanish texts, we changed these conditions so as to capture the appropriate context. As mentioned previously, our algorithm does not have access to full syntactic knowledge. Accordingly, we employed partial parsing over the text in our application of Lappin and Leass's algorithm. The salience parameters were weighted (weight appears in parentheses) and applied in the following way: * Sentence recency (100): Applied when the NP appeared in the same sentence as the anaphor.</Paragraph> <Paragraph position="2"> * Subject emphasis (80): Applied when the NP was located before the verb of the clause in which it appeared. This heuristic was necessary because of our algorithm's lack of syntactic knowledge. It should be noted, however, that since Spanish is a nearly free-word-order language and the exchange of subject and object positions within Spanish sentences is common, the heuristic is often invalid. For example, the two Spanish sentences Pedro compr6 un regalo 'Pedro bought a present' and Un regalo compr6 Pedro 'A present bought Pedro' are equivalent to one another and to the English sentence Pedro bought a present.</Paragraph> <Paragraph position="3"> * Existential emphasis (70): In this instance, we applied the parameter in the same way as Lappin and Leass, since the entire NP was fully parsed, which allowed us to tell when it was a definite or an indefinite NP.</Paragraph> <Paragraph position="4"> * Accusative emphasis (50): Applied when the NP appeared after the verb of the clause in which it appeared and the NP did not appear inside another NP or PP. For example, in the sentence Pedro encontr6 el libro de Juana 'Pedro found Juana's book', a value was assigned to el libro de Juana 'Juana's book' but not to Juana. Once again, it should be noted that this heuristic was necessary because of our algorithm's lack of syntactic knowledge.</Paragraph> <Paragraph position="5"> * Indirect object and oblique complement emphasis (40): Applied when the NP appeared in a PP with the Spanish preposition a 'to', which usually preceded the indirect object of its sentence.</Paragraph> <Paragraph position="6"> * Head noun emphasis (80): Applied when the NP was not contained in another NP.</Paragraph> <Paragraph position="7"> * Nonadverbial emphasis (50): Applied when the NP was not contained in an adverbial PP. In this case, its application depended on the kind of preposition in which the NP was included.</Paragraph> <Paragraph position="8"> * Parallelism reward (35): Applied when the NP occupied the same position as the anaphor with reference to the verb of the sentence (before or after the verb).</Paragraph> <Paragraph position="9"> Finally, we followed Lappin and Leass in assigning the additional salience value to NPs in the current sentence and in degrading the salience of NPs in preceding sentences.</Paragraph> <Paragraph position="10"> Our results exhibited some similarities with Lappin and Leass's experiments. For example, anaphora was strongly preferred over cataphora, and both approaches Palomar et al. Anaphora Resolution in Spanish Texts preferred intrasentential NPs to intersentential ones. These results can be seen in</Paragraph> </Section> <Section position="6" start_page="562" end_page="563" type="sub_section"> <SectionTitle> 5.5 Centering Approach </SectionTitle> <Paragraph position="0"> The centering model proposed by Grosz, Joshi, and Weinstein (1983, 1995) provides a framework for modeling the local coherence of discourse. The model has two constructs, a list of forward-looking centers and a backward-looking center, that can be assigned to each utterance Ui. The list of forward-looking centers Cf(Ui) ranks discourse entities within the utterance Ui. The backward-looking center Cb(Ui+l) constitutes the most highly ranked element of Cf(Ui) that is finally realized in the next utterance Ui+l. In this way, the ranking imposed over Cf(Ui) must reflect the fact that the preferred center Cp(U/) (i.e., the most highly ranked element of Cf(Ui)) is most likely to be Cb(Ui+l).</Paragraph> <Paragraph position="1"> The ranking criteria used by Grosz, Joshi, and Weinstein (1995) order items in the Cf list using grammatical roles. Thus, entities with a subject role are preferred to entities with an object role, and objects are preferred to others (adjuncts, etc.).</Paragraph> <Paragraph position="2"> Grosz, Joshi, and Weinstein (1995) state that if any element of Cf(Ui) is realized by a pronoun in Ui+l, then Cb(Ui+l) must also be realized by a pronoun.</Paragraph> <Paragraph position="3"> Brennan, Friedman, and Pollard (1987) applied the centering model to pronoun resolution. They based their algorithm on the fact that centering transition relations will hold across adjacent utterances.</Paragraph> <Paragraph position="4"> Moreover, one crucial point in centering is the ranking of the forward-looking centers. Grosz, Joshi, and Weinstein (1995) state that Cf may be ordered using different factors, but they only use information about grammatical roles. However, both Strube (1998) and Strube and Hahn (1999) point out that it is difficult to define grammatical roles in free-word-order languages like German or Spanish. For languages like these, they propose other ranking criteria dependent upon the information status of discourse entities. They claim that information about familiarity is crucial for the ranking of discourse entities, at least in free-word-order languages.</Paragraph> <Paragraph position="5"> According to Strube's ranking criteria, two different sets of expressions, hearer-old discourse entities (OLD) and hearer-new discourse entities (NEW), can be distinguished. OLD discourse entities consist of evoked entities---coreferring resolved expressions (pronominal and nominal anaphora, previously mentioned proper names, relative pronouns, appositives)--and unused entities (proper names and titles). The remaining entities are assigned to the NEW set. The basic ranking criteria for pronominal anaphora resolution prefer OLD entities over NEW entities. 8 Strube (1998) thus proposes the following adaptation to the centering model: The Cf list is replaced by the list of salient discourse entities (S-list) containing discourse entities that are realized in the current and previous utterance.</Paragraph> <Paragraph position="6"> * The elements of the S-list are ranked according to the basic ranking criteria and position information: If X E OLD and y C NEW, then x precedes y.</Paragraph> <Paragraph position="7"> If x, y ~ OLD or x, y E NEW, 8 To resolve functional anaphora, a third set, MED, which includes inferable information, must be added between the OLD and the NEW sets. However, this set is not needed to resolve pronominal anaphora (Strube and Hahn 1999).</Paragraph> <Paragraph position="8"> if utterance(y) = utterance(x) and pos(x) < pos(y), then x precedes y.</Paragraph> <Paragraph position="9"> Since there is not a clear definition of what an utterance is, the following criteria are assumed: tensed clauses are defined as utterances on their own and untensed clauses are processed with the main clause in order to constitute only one utterance.</Paragraph> <Paragraph position="10"> Incorporating these adaptations, Strube (1998) then proposes the following algorithm: null 1. If a referring expression is encountered, (a) if it is a pronoun, test the elements of the S-list in order until the test succeeds; (b) update the S-list using information about this referring expression.</Paragraph> <Paragraph position="11"> 2. If the analysis of utterance U is finished, remove all discourse entities from the S-list that are not realized in U.</Paragraph> <Paragraph position="12"> The evaluation of this algorithm was performed in Strube (1998) and obtained a precision of 85.4% for English, improving upon the results of the centering algorithm by Brennan, Friedman, and Pollard (1987), which achieved only 72.9% precision when it was applied to the same corpus.</Paragraph> <Paragraph position="13"> Consequently, in adapting the centering model to Spanish anaphora resolution, we followed Strube's indications. The success rate of the algorithm was not satisfactory, as can be seen in Table 4.</Paragraph> </Section> </Section> class="xml-element"></Paper>