Anaphora Resolution and its Importance in MT

As noted earlier, anaphora resolution is of crucial importance in order to translate anaphoric expressions correctly into a target language. Let us consider the sentences [Hutchins & Somers, 1992]:

1. [The monkey]$_i$ ate the banana because it$_i$ was hungry.
   
   
2. The monkey ate [the banana]$_i$ because it$_i$ was ripe.
   
   
3. The monkey ate the banana because it was tea-time.
   
   

In each sentence the pronoun it refers to something different: in sentence (1), it refers to the monkey, in sentence (2) to the banana, and in sentence (3), to the abstract notion of time. If we wish to translate these sentences into Spanish or German (languages which mark the gender of pronouns), anaphora resolution will be absolutely essential since, in these languages, pronouns take the gender of their antecedents. Therefore, in Spanish, we would obtain the following pronouns: (1) éste (in the masculine form since the antecedent--the monkey--is masculine), (2) ésta (feminine--the banana), and (3) an omitted pronoun (since the second clause of the sentence is impersonal in Spanish and does not need any subject). On the other hand, in German we would obtain: (1) er (masculine antecedent), (2) sie (feminine antecedent), and (3) es (neutral).

Besides these problems, originated by the gender of anaphoric expressions in different languages, there are other differences (that we have named discrepancies) which influence the process of translation of these expressions. These discrepancies have been previously studied by other authors. Mitkov & Schmidt [Mitkov & Schmidt, 1998] present several problems to be taken into account in the translation of pronominal anaphors between different languages (German, French, English, Malay, and Korean); Nakaiwa & Ikeara [Nakaiwa & Ikehara, 1992] treat the problem of the translation of elliptical constructions in a Japanese-English MT system; and Mitkov et al. [Mitkov et al., 1994] and Geldbach [Geldbach, 1999] present the discrepancies in an English-Korean MT system and a Russian-German MT system, respectively.

Another difference between languages is that of number discrepancies, in which certain nouns are referred to by a singular pronoun in one language and by a plural noun in the other. For example, the word people is plural in English, whereas in Spanish or German it is singular. Hence, in translations from English to Spanish, or from English to German, the plural pronoun will become a singular pronoun.

On the other hand, although in the majority of cases language-pairs pronouns in the source language are translated by target-language pronouns that correspond to the antecedent of the anaphor, there are some exceptions. In most of these cases, pronominal anaphors are simply omitted in the target language. For instance, in translations from English to Spanish, pronouns are frequently not translated because of the typical Spanish elliptical zero-subject construction. Other languages with typical zero constructions are Japanese, Italian, Thai, or Chinese.

In some languages, however, the pronoun is directly translated by its antecedent. For example, in English-Malay translations there is a tendency to replace the it pronoun with its antecedent, in which case the translator must first be able to identify the antecedent.

Some languages translate pronouns into different expressions, depending on the syntactic and semantic information of the antecedent. For example, in English-Korean translation pronouns can be elliptically omitted, or they can be translated into definite noun phrases, into their antecedent, or into different Korean pronouns.

All the above-mentioned problems in the translation of anaphoric expressions into a target language show that it is very important to carry out a detailed analysis of these expressions (including their resolution and the identification of the antecedent).

Because the majority of MT systems only handle one-sentence input, they usually cannot deal with anaphora resolution, and if they do, their successful operation usually does not go beyond the sentence level. In order to assess the deficiencies of MT systems, we analyzed the characteristics of different MT systems, with an emphasis on those characteristics related to anaphora resolution and translation into a target language. An overview of our analysis can be seen in Table 1.

Table 1: Characteristics of main MT systems

MT system	Strategy1	Restrict2	Partial3	Anaphor4	Corefer5	Zero6
Systran	Direct	No	Yes	Yes	No	Yes
Météo	Direct	Yes	No	No	No	No
SUSY	Transfer	No	No	Yes	No	No
Ariane	Transfer	No	No	Yes	No	Yes
Eurotra	Transfer	No	No	No	No	No
METAL	Transfer	No	Yes	Yes	No	Yes
Candide	Transfer	Yes	No	No	No	No
Inter-Nostrum	Transfer	Yes	Yes	No	No	No
IXA	Transfer	No	Yes	No	No	No
Episteme	Transfer	No	No	No	No	No
KANT	Interlingua	Yes	No	Yes	No	Yes
DLT	Interlingua	No	No	No	No	No
DLT (BKB)	Interlingua	No	No	Yes	No	No
Rosetta	Interlingua	No	No	No	No	No
CREST	Interlingua	Yes	No	Yes	Yes	Yes
$\mu$kosmos	Interlingua	Yes	No	No	No	No

¹ Strategy of translation: direct, transfer, or interlingua.

² Restricted domain.

³ Partial parsing.

⁴ Resolution of intersentential anaphora.

⁵ Identification of co-reference chains.

⁶ Translation of zero pronouns into the target language.

The table reflects a number of different system characteristics.

1. MT system. The MT systems studied included Systran [Toma, 1977,Wheeler, 1987]; Météo [Chandioux, 1976,Chandioux, 1989]; SUSY [Maas, 1977,Maas, 1987]; Ariane [Boitet & Nédobejkine, 1981,Boitet, 1989]; Eurotra [Varile & Lau, 1988,Allegranza et al., 1991]; METAL [Bennet & Slocum, 1985,Thurmair, 1990]; Candide [Berger et al., 1994]; Inter-Nostrum [Canals-Marote et al., 2001a,Canals-Marote et al., 2001b]; IXA [Díaz-Ilarraza et al., 2000,Díaz-Ilarraza et al., 2001]; Episteme [Amores & Quesada, 1997,Quesada & Amores, 2000]; KANT [Goodman, 1989,Nirenburg, 1989,Mitamura et al., 1991]; DLT [Witkam, 1983,Schubert, 1986]; DLT with Bilingual Knowledge Bank (BKB) [Sadler, 1989]; Rosetta [Appelo & Landsbergen, 1986,Landsbergen, 1987]; CREST [Farwell & Helmreich, 2000]; and $\mu$kosmos [Mahesh & Nirenburg, 1995a,Mahesh & Nirenburg, 1995b].
   
   
2. Strategy of translation. This characteristic indicates the strategy used by the MT system in accordance with the existence of intermediate representations: direct, transfer, or interlingua.
   
   
3. Restricted domain. This characteristic tells if the texts of the source language are of a specific domain (restricted domain).
   
   
4. Partial parsing. This characteristic indicates if the MT system carries out a partial parsing of the source text by identifying only some constituents (noun phrases, prepositional phrases, etc.) and some relations between them.
   
   
5. Resolution of intersentential anaphora. This characteristic indicates whether the MT system resolves intersentential anaphora. If it does not, then the anaphoric expressions that have their antecedents in previous sentences will be incorrectly translated into the target language, in most of cases.
   
   
6. Identification of co-reference chains. This characteristic tells us if the co-reference chains of the source text are identified after resolving intersentential anaphora.
   
   
7. Translation of zero pronouns. This characteristic indicates if the MT system detects and resolves omitted pronouns (zero pronouns¹) in the source language that are compulsory in the target language.
   
   

After analyzing the characteristics of the primary commercial MT systems, we concluded that there is no MT system that can work on unrestricted texts, resolve intersentential anaphora, identify the co-reference chains of the text, and translate zero pronouns into the target language after carrying out a partial parsing of the source text.

Unlike other systems, such as the KANT interlingua system, the Météo system, and the Candide system, among others, that are designed for well-defined domains, our interlingua MT approach, called AGIR (Anaphora Generation with an Interlingua Representation), works on unrestricted texts. Although we could have applied full parsing to these texts, we have instead utilized partial parsing, due to the unavoidable incompleteness of the grammar. This is a main difference between our system and other interlingua systems, such as the DLT system (which is based on a modification of Esperanto), the Rosetta system (which experiments with Montague semantics as the basis for an interlingua), the KANT system, and others.

After parsing and solving pronominal anaphora, an interlingua representation of the entire text is obtained. To do this, sentences are split into clauses, and a complex feature structure based on semantic roles (agent, theme, etc.) is generated for each one. For each clause, the different semantic roles that appear will be identified and linked with one entity of the text. If there is an anaphor in the text, it will be linked with the entity that represents its antecedent. The AGIR's interlingua representation has been presented in more detail in [Peral et al., 1999,Peral & Ferrández, 2000a].

From the interlingua representation, the translation of the anaphor (including the intersentential anaphor), the detection of co-reference chains of the whole text, and the translation of Spanish zero pronouns into English have been carried out. AGIR has been designed to deal with all these issues which are hardly considered by most of the real MT systems. Furthermore, our approach can be used for other applications, for example, for cross-language information retrieval, summarization, etc.

It is important to note that although some of the above-mentioned MT systems resolve different problems, such as zero pronouns or pronominal anaphora, their results are not very satisfactory. Furthermore, we present some examples (extracted from the corpora used in the evaluation of our approach--see Section 4) of incorrect Spanish-English-Spanish translations of pronouns done by Systran² that AGIR does correctly³:

• (S) Siempre creí que a lo que yo aspiraba era a la comunicación perfecta con un hombre, o, mejor dicho, con el hombre, con ese príncipe azul de los sueños de infancia, un ser que sabría adivinarme hasta en los más menudos pliegues interiores. Ahora he aprendido no sólo que [esa fusión]$_i$ es imposible, sino además que Ø$_i$ es probablemente indeseable.
  
  
• (E) I always thought that to which I aspired I was to the perfect communication with a man, or, rather, with the man, that blue prince of the childhood dreams, a being who would know to guess to me until in slightest you fold interiors. Now I have learned not only that that fusion is impossible, but in addition that he is probably undesirable.
  
  

In this example, Systran incorrectly translates into English the zero pronoun of the last sentence of the paragraph, proposing the pronoun he instead of the pronoun it (the antecedent is the noun phrase esa fusión--that fusion). Our system proposed the correct pronoun. It is important to note that although the zero pronoun is identified by Systran, it is incorrectly solved and subsequently incorrectly translated.

• (S) Al pasar de [la luminosidad]$_i$ de la calle, y tal vez por contraste con ella$_i$, impresionaba la oscuridad interior y el vacío de la Catedral, en la que apenas había a la vista cuatro o cinco personas.
  
  
• (E) When happening of the luminosity of the street, and perhaps in contrast with her, it impressed the inner dark and the emptiness of the Cathedral, in which as soon as there were at sight four or five people.
  
  

In this case, Systran incorrectly translates into English the pronoun ella (with antecedent la luminosidad--the luminosity), by proposing the pronoun her instead of it.

• (E) If you have not already done so, unpack [your printer]$_i$ and the accessory kit that came with it$_i$.
  
  
• (S) Si usted no ha hecho ya así pues, desempaquete su impresora y el kit de accesorios eso vino con él.
  
  

This example shows an incorrect English-Spanish translation of the pronoun it. In this case, the pronoun is incorrectly solved (the antecedent is the noun phrase your printer, feminine) and then it is incorrectly translated (pronoun él--masculine--instead of pronoun ésta--feminine).

All the above examples illustrate that the translation of pronouns could be notably improved if their antecedents were correctly identified and, subsequently, pronouns were translated into the target language.