In pro-drop languages such as Chinese, Japanese, and Korean,
repetitive elements are easily omitted when they are inferable from the context.
This phenomenon is commonly referred to as zero anaphora,
and the omitted element is called as a zero pron...
In pro-drop languages such as Chinese, Japanese, and Korean,
repetitive elements are easily omitted when they are inferable from the context.
This phenomenon is commonly referred to as zero anaphora,
and the omitted element is called as a zero pronoun (or zero anaphor).
The resolution of zero pronouns is essential for natural language understanding,
and thus is important for many NLP applications such as machine translation, information extraction, text summarization, question answering, and so on.
One way to resolve zero pronouns is to pair each zero pronoun with the most plausible antecedent chosen from the antecedent candidates.
That is, if a number of noun phrases precede a zero pronoun,
the most plausible antecedent should be chosen from these noun phrases.
Various kinds of features have been proposed to choose the most plausible antecedent, such as agreement feature, syntactic feature and structured feature.
However, zero pronoun resolution is still difficult,
no matter how nice are the features defined to choose the most plausible antecedent.
This is because the number of antecedent candidates is generally very large,
since a sentence has many noun phrases and all noun phrases preceding a zero pronoun can be its antecedent.
This high cardinality of antecedent candidates is one of the major reasons which make zero pronoun resolution difficult.
Therefore, a method to reduce this cardinality should be considered before defining the features to choose the most plausible antecedent.
Languages are believed to have a hierarchy of linguistic units.
On the basis of the hierarchy structure,
candidate antecedents can be grouped into a larger unit such as phrase, clause, sentence, and paragraph.
Since the number of clauses in a sentence is far less than that of phrases,
it is easier, from the standpoint of the number of targets, to choose the clause with the antecedent of a zero pronoun than to find the antecedent directly.
Therefore, this paper proposes a method for intra-sentential zero pronoun resolution by using clause-level antecedent identification.
The proposed method first identifies the clause which may contain the antecedent for a given zero pronoun using structural information of clause.
A parse tree kernel is adopted for modeling the structural information of clauses.
Then, the antecedent of the zero pronoun is chosen from the noun phrases within the identified clause.
If the first step is trustworthy enough,
the proposed method outperforms other methods without the first step since the second step chooses the antecedent from far less number of candidates.
Our experimental results show that the proposed method outperforms other methods without the first step,
no matter what features are used to identify antecedents.
This implies that the large norm of the candidate set is a major factor affecting the performance of zero pronoun resolution,
and thus our two-step approach is more effective in resolving zero pronouns.
In our future work, we intend to improve the first step of the identification of clauses further.
The performance of zero anaphora resolution will be much better,
if additional semantic information or a combination of structural and semantic features is used to identify the clause with antecedent.
In addition, we plan to extend this approach to resolve zero pronouns which occur in various grammatical positions as well as to resolve inter-sentential zero pronouns.