- (Mitkov, 1999) ⇒ Ruslan Mitkov. (1999). “Anaphora Resolution: The State of the Art.” In: Technical report. University of Wolverhampton.
Anaphora resolution is a complicated problem in Natural Language Processing and has attracted the attention of many researchers. The approaches developed - traditional (from purely syntactic ones to highly semantic and pragmatic ones), alternative (statistic, uncertainty-reasoning etc.) or knowledge-poor, offer only approximate solutions. The paper is an introduction to anaphora resolution offering a brief survey of the major works in the field.
1.1 Basic notions and terminology
- The etymology of the term "anaphora" goes back to Ancient Greek with “anaphora” (anajora) being a compound word consisting of the separate words ana - back, upstream, back in an upward direction and jora - the act of carrying and denoted the act of carrying back upstream. For Computational Linguists embarking upon research in the field of anaphor resolution, I strongly recommend as a primer Graham Hirst's book "Anaphora in natural language understanding" (Hirst 1981) which may seem a bit dated in that it does not include developments in the 80's and the 90's, but which provides an excellent survey of the theoretical work on anaphora and of the early computational approaches and is still very useful reading.
- Various definitions of anaphora have been put forward, but I am tempted to paraphrase the classical definition given by Halliday and Hasan (Halliday & Hasan 1976) which is based on the notion of cohesion: anaphora is cohesion (presupposition) which points back to some previous item. [1 We shall not discuss cataphora which is the case when the "anaphor" precedes the antecedent. For example "Because she was going to the post office, Julie was asked to post a small parcel"] expressions and having the same referent in the real world, they are termed coreferential. [2 The relation between the anaphor and the antecedent is not to be confused with that between the anaphor and its referent; in the example below the referent is "the Empress" as a person in the real word whereas the antecedent is "the Empress" as a linguistic form.]
- The "pointing back" (reference) is called an anaphor and the entity to which it refers is its antecedent. The process of determining the antecedent of an anaphor is called anaphora resolution. Usually, both the antecedent and the anaphor are used as referring.
- Example (Huddleston 1984). “The Empress hasn't arrived yet but she should be here any minute.
- In this example, the pronoun "she" is the anaphor (for classification of anaphors, see below) and "the Empress" is the antecedent. Please note that the antecedent is not the noun "Empress" but the noun phrase "the Empress".
- There may be cases when the anaphor and more than one of the preceding (or following) entities (usually noun phrases) have the same referent and are therefore pairwise coreferential, thus forming a coreferential chain. In such a case, we regard each of the preceding entities which are coreferential with the anaphor(s) as a legitimate antecedent. Therefore, in such cases the task of anaphora resolution is considered successful, if any of the preceding entities in the coreferential chain is identified as an antecedent. Our paper will discuss the task of anaphora resolution only and not coreference resolution (except for briefly mentioning it in section 4.2). For more on coreference resolution, I suggest the reader consult the MUC (Message Understating Conference) Proceedings in which coreference resolution is extensively covered.
1.2 Types of anaphora
- There are various types of anaphora (Hirst 1981), but we will shall briefly outline those that are thought to be the three most widespread types in the Computational Linguistics literature.
- The most widespread type of anaphora is the pronominal anaphora which is realised by anaphoric pronouns.
- Example: Computational Linguists from many different countries attended the tutorial. They took extensive notes.
- It should be pointed out that not all pronouns in English are anaphoric. For instance, "it" can often be non-anaphoric such as in the case of the previous sentence. Other examples of non-anaphoric “it” include expressions such as "It is important", "It is necessary", "It has to be taken into account". A non-anaphoric "it" is termed pleonastic (Lappin & Leass 1994).
Definite noun phrase anaphora
- Typical cases of definite noun phrase anaphora is when the antecedent is referred by a definite noun phrase representing either same concept (repetition) or semantically close concepts (e.g. synonyms, superordinates).
- Example: Computational Linguists from many different countries attended the tutorial. The participants found it hard to cope with the speed of the presentation.
- One-anaphora is the case when the anaphoric expression is realised by a "one" noun phrase. Example: If you cannot attend a tutorial in the morning, you can go for an afternoon one.
- Finally, we distinguish intrasentential anaphors (referring to an antecedent which is in the same sentence as the anaphor) and intersentential anaphors (referring to an antecedent which is in a different sentence from that of the anaphor).
1.3 The process of anaphora resolution
Most of the anaphora resolution systems deal with resolution of anaphors which have noun phrases as their antecedents because identifying anaphors which have verb phrases, clauses, sentences or even paragraphs/discourse segments as antecedents, is a more complicated task. Typically, all noun phrases (NPs) preceding an anaphor are initially regarded as potential candidates for antecedents. Usually, a search scope has to be identified: most approaches look for NPs in the current and preceding sentence. However, an "ideal" anaphora resolution system should extend its scope of search: antecedents which are 17 sentences away from the anaphor have already been reported (Mitkov 1995a)!
Assuming that the scope of search for a specific approach has already been specified, the NPs preceding the anaphor within that scope are identified as candidates for antecedents and a number of anaphora resolution factors are employed to track down the correct antecedent.
Approaches to anaphora resolution usually rely on a set of "anaphora resolution factors". Factors used frequently in the resolution process include gender and number agreement, c-command constraints, semantic consistency, syntactic parallelism, semantic parallelism, salience, proximity etc. [In English antecedents usually agree in gender and number with the anaphors, but this is not always the case in other languages.] These factors can be "eliminating" i.e. discounting certain noun phrases from the set of possible candidates (such as gender and number constraints3, c-command constraints, semantic consistency) or "preferential", giving more preference to certain candidates and less to others (such as parallelism, salience). Computational linguistics literature uses diverse terminology for these - for example E. Rich and S. LuperFoy (Rich & LuperFoy 1988) refer to the "eliminating"
1.3.3 Computational strategies
2. Early work on anaphora resolution
2.1 STUDENT (Bobrow 1964)
2.2 SHRDLU (Winograd 1972)
2.3 Preference semantics (Wilks 1973)
2.4 Taking syntax on board: Hobbs' algorithm
3. Recent developments in anaphora resolution
3.1 Traditional approaches
3.1.1 D. Carter's shallow processing approach
3.1.2 E. Rich and S. LuperFoy's distributed architecture
3.1.3 J. Carbonell and R. Brown's multi-strategy approach
3.1.4 C. Rico Pérez' scalar product coordinating approach
3.1.5 R. Mitkov: combination of linguistic and statistical methods
3.1.6 Lappin and Leass' syntax-based approach
3.2 Alternative approaches
3.2.1 Nasukawa's "knowledge-independent" approach
3.2.2 Statistical/corpus processing approach (Dagan & Itai)
3.2.3 Connolly et al.'s machine learning approach (Connolly, Burger and Day 1994)
2.4 Aone & Bennett's machine learning approach (Aone & Bennet 1996)
3.2.5 An uncertainty-reasoning approach (Mitkov 1995b)
3.2.6 Two-engine approach (Mitkov 1997a)
3.2.7 Situational semantics approach (Tin & Akman 1994)
3.2.8 Using punctuation (Say & Vakman 1996)
3.3 Latest trends: knowledge-poor anaphora resolution
3.3.1 Kennedy and Boguraev's approach without a parser
3.3.2 Robust, knowledge-poor approach (Mitkov 1996, 1998).
3.1.4 Breck Baldwin's COGNIAC
4.1. Anaphora resolution in Machine Translation
4.1.1 Additional problems
4.1.2 Research and development to date
126.96.36.199 English-to-Japanese MT program (Wada 1990)
188.8.131.52 English-to-Chinese Machine Translation (Chen 1992)
= 184.108.40.206 Resolution of Japanese zero pronouns (Nakaiwa et al. 1991, 1994, 1995)
220.127.116.11 Portuguese-to-English MT (Saggion & Carvalho 1994)
18.104.22.168 English-to-German MT system (Preuß et al. 1994)
22.214.171.124 English-to-Korean MT system (Mitkov et al. 1994)
126.96.36.199 Extension of CAT2 (Mitkov et al. 1995)
4.2 Information extraction
4.2.1 Kameyama's algorithm
4.2.2 Pronoun resolution in M-LaSIE (Azzam et al. 1998)
4.3 Other applications
4.3.1 Automatic abstracting
5. Some new developments and issues that need further attention
5.1 Evaluation in anaphora resolution
5.2 Multilingual anaphora resolution
5.2.1 Anaphora resolution and languages covered
5.2.2 Recent multilingual developments
5.3 Anaphora resolution in the world of growing resources
5.4 A few open questions
5.4.1 Dependence and mutual dependence of factors
5.4.2 Do the factors hold good for all genres?
5.4.3 Factors and multilinguality
5.4.4 Order of constraints and priority of preferences
|1999 AnaphoraResolution||Ruslan Mitkov||Anaphora Resolution: The State of the Art||Technical Report||http://clg.wlv.ac.uk/papers/mitkov-99a.pdf||1999|