An Estimate of Referent of Noun Phrases in Japanese Sentences

Masaki Murata, Makoto NagaoACL 1998

TL;DR

This paper tackles the challenge of determining what a noun phrase refers to in Japanese, a language lacking articles that make this task difficult in English. They propose a method that combines a noun phrase's "referential property" (whether it's generic, definite, or indefinite) with its modifiers and possessors to accurately link noun phrases to their antecedents, crucial for machine translation and dialogue systems.

What Problem Does This Paper Address?

In natural language processing, especially for tasks like machine translation or conversational AI, understanding what a noun phrase (like "the old man" or "a dog") refers to is fundamental. This is known as anaphora resolution. In English, articles like "the" and "a" provide strong clues. For instance, "the" often signals that the noun phrase refers to something already mentioned or uniquely identifiable. However, Japanese, like many other languages, does not use articles. This absence makes it significantly harder to determine if a Japanese noun phrase refers to an entity already introduced in the discourse (an antecedent) or if it's introducing a new one. Without this disambiguation, a machine translation system might incorrectly translate repeated mentions of "old man" as distinct individuals when they actually refer to the same person, leading to unnatural or incorrect translations. This paper aims to solve this ambiguity problem for Japanese noun phrases.

Methods

The authors propose a three-pronged approach to determine noun phrase referents:

  1. Referential Property Constraint: This is the foundational element. They leverage a previously developed method (Murata and Nagao, 1993) to classify noun phrases into three categories:
  2. Generic: Refers to all members of a class or the class itself (e.g., "Dogs are useful").
  3. Definite: Refers to a contextually non-ambiguous member of a class (e.g., "The dog went away"). This type is most likely to have an antecedent.
  4. Indefinite: Refers to an arbitrary member of a class (e.g., "There are three dogs"). This type is unlikely to have an antecedent. This classification is based on surface linguistic cues within the sentence.

  5. Modifier Constraint: This rule states that if two noun phrases have different modifiers, they generally do not refer to the same entity. For example, "the lump on his right cheek" and "the lump on his left cheek" refer to different lumps. Conversely, a noun phrase without a modifier can refer to an entity with a modifier.

  6. Possessor Constraint: This constraint is applied when a noun phrase denotes a part of something (e.g., a "cheek"). The system attempts to identify the "possessor" of that part, typically a human or animal entity mentioned earlier in the text. If two noun phrases have the same possessor, they are more likely to refer to the same entity. For instance, if "the old man's cheek" is mentioned, and later "his cheek" appears, the system can infer they refer to the same cheek because the possessor ("the old man") is consistent.

When two noun phrases share the same head noun and satisfy these three constraints, the system judges them to have the same referent.

Datasets

The paper describes experiments using:

  • Training sentences: Used to hand-craft heuristic rules for the anaphora resolution system.
  • Test sentences: Used to evaluate the effectiveness of the developed rules and methods.

Specific examples of sentence sources include:

  • Example sentences (43)
  • Folk tale "KOBUTORI JIISAN" (93 sentences)
  • Essay in "TENSEIJINGO" (26 sentences)
  • Editorial (26 sentences)
  • Article in "Scientific American (in Japanese)" (16 sentences)

For test sentences, they used:

  • Folk tale "TSURU NO ONGAESHI" (91 sentences)
  • Two essays in "TENSEIJINGO" (50 sentences)
  • Editorial (30 sentences)
  • "Scientific American (in Japanese)" (13 sentences)

The paper does not mention the creation of new, specific benchmarks but rather uses existing text corpora.

Key Insights and Ideas

  • The absence of articles in Japanese is a significant hurdle: This is the core problem they address, highlighting the difficulty compared to article-rich languages like English.
  • Referential property is a crucial starting point: Classifying noun phrases as generic, definite, or indefinite provides a strong initial signal for potential antecedents.
  • Modifiers and possessors are essential for disambiguation: Relying solely on referential property is insufficient. Modifiers and possessors act as fine-grained filters, preventing incorrect co-reference assignments when these attributes differ.
  • Possessor information can be language-independent: The idea of identifying possessors based on semantic roles (like human/animal subjects) is presented as a potentially universal mechanism for anaphora resolution, applicable beyond Japanese.
  • A combined approach yields the best results: The experiments demonstrate that integrating all three constraints (referential property, modifier, and possessor) leads to significantly higher precision and recall than using subsets of these constraints.

Key Results

The paper reports the following performance metrics:

  • On training sentences:
    • Precision: 82% (130/159)
    • Recall: 85% (130/153)
  • On test sentences:
    • Precision: 79% (89/113)
    • Recall: 77% (89/115)

Table 2 compares their proposed method (Method 1) against variations:

  • Method 1 (Full): Uses referential property, modifiers, and possessors. Achieved the highest precision and recall on both training and test sets.
  • Method 2 (Definite only): Only allows co-reference if the noun phrase is estimated as definite. Lower recall due to misclassifications of definite NPs as indefinite/generic.
  • Method 3 (No referential property): Relies on other cues like distance, topic, modifiers, and possessors. Lower precision and recall than Method 1.
  • Method 4 (No modifier/possessor): Relies only on referential property. Significantly lower precision, indicating the critical role of modifier and possessor constraints.

Limitations and Future Work

  • Limitations:
    • The modifier constraint is acknowledged as not being "truly applicable to all cases," with exceptions where noun phrases might refer to entities with different modifiers.
  • The system's performance can be negatively impacted if the possessor of a noun is estimated incorrectly.
  • The paper notes that determining if two different expressions are "equal in meaning" (e.g., different descriptions of the same hole) is a challenge that requires further analysis beyond simple modifier matching.
  • Future Work:
    • The authors suggest that the possessor constraint, based on semantic roles, is a language-independent knowledge that can be applied to other languages.
  • The need to decide whether different expressions are semantically equivalent, even with different modifiers, points towards future work in deeper semantic understanding.

Summary updated

BibTeX
@inproceedings{murata1998estimate,
  author = {Masaki Murata and Makoto Nagao},
  title = {{An Estimate of Referent of Noun Phrases in Japanese Sentences}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1998},
}
Figure