TL;DR
This paper introduces a novel probabilistic model for analyzing Japanese anaphora, specifically focusing on "zero pronouns" (omitted words). Their method integrates zero pronoun detection and antecedent resolution into a single framework, achieving improved accuracy by leveraging both syntactic and semantic features, and importantly, utilizing unannotated corpora to mitigate data sparsity issues.
What Problem Does This Paper Address?
Natural Language Processing (NLP) systems often struggle with anaphora resolution, the task of identifying what a pronoun or other referring expression refers to. While English has well-studied coreference resolution methods, languages like Japanese present a unique challenge due to the frequent omission of obligatory grammatical cases, known as "zero pronouns." Unlike explicit pronouns, these omitted elements must first be detected before their antecedents can be identified. This two-step process (detection then resolution) is computationally demanding and prone to errors. Existing approaches often rely heavily on manually annotated corpora, which are expensive and time-consuming to create, leading to data sparsity issues. This paper aims to address these challenges by proposing a unified probabilistic model for both detecting and resolving Japanese zero pronouns, while also exploring methods to effectively utilize readily available unannotated data.
Methods
The core of this paper is a probabilistic model that integrates zero pronoun detection and antecedent resolution. The system first performs morphological and syntactic analysis on the input Japanese text.
Zero Pronoun Detection: The system identifies potential "zero pronoun candidates" by analyzing syntactic relations and utilizing a verb dictionary with case frames. A key probabilistic parameter, $P_{zero}(c|v)$, quantifies the likelihood that a specific case ($c$) governed by a verb ($v$) is omitted. This probability is estimated based on the co-occurrence frequency of verb-case pairs extracted from unannotated corpora.
Antecedent Resolution: For each detected zero pronoun candidate, the system identifies potential antecedents from preceding sentences. The resolution process is framed as computing the probability that a zero pronoun refers to a specific antecedent, $P(a_i|\phi_c)$. This probability is decomposed into two main components:
- Syntactic Model: This component utilizes features like post-positional particles, distance between the zero pronoun and candidate, and whether the candidate is within a relative clause. Probabilities for these features are estimated from annotated corpora.
- Semantic Model: This component incorporates semantic classes of antecedents. Crucially, to combat data sparsity, the paper proposes estimating semantic probabilities ($P(n_i|v, c)$) using co-occurrences of verbs and their case nouns extracted from unannotated corpora.
The overall probability for a zero pronoun referring to an antecedent is a combination of these syntactic and semantic probabilities.
Certainty Score: To improve the reliability of the system's output, a "certainty score" is introduced. This score combines the probabilities of the top two antecedent candidates, allowing the system to filter out low-confidence anaphoric relations.
Datasets
- Kyotodaigaku Text Corpus version 2.0: Used for manual annotation of anaphoric relations for zero pronouns. 30 general articles were selected and annotated, containing 449 zero pronouns.
- Mainichi Shimbun newspaper articles (1994-1999): A large corpus of six years of newspaper articles used to produce the semantic model based on co-occurrence statistics from unannotated data.
- Japanese Bunruigoihyou thesaurus: Used to define 544 semantic classes for nouns.
Key Insights and Ideas
- Unified Detection and Resolution: Integrating zero pronoun detection and antecedent resolution into a single probabilistic framework is more effective than treating them as separate steps.
- Leveraging Unannotated Corpora: The ability to estimate semantic model probabilities from unannotated corpora is a significant breakthrough, effectively mitigating the data sparsity problem that plagues many anaphora resolution systems.
- Syntactic Features are Crucial: Syntactic features alone (Syn model) outperformed individual semantic models, highlighting their importance in Japanese zero pronoun resolution.
- Certainty Score for Reliability: The proposed certainty score effectively filters out low-confidence results, improving the overall precision of the system.
- Zero Pronoun Detection Matters: The paper demonstrates that improving zero pronoun detection directly leads to better overall accuracy, even for the resolution component.
Key Results
- The proposed combined model (Both2) achieved the highest accuracy in zero pronoun resolution across various k-best candidate settings, outperforming individual semantic, syntactic, and rule-based models.
- For k=1 (the top candidate), the complete model achieved an accuracy of 50.7%.
- The semantic model trained on unannotated corpora (Sem2) significantly outperformed the one trained on annotated corpora (Sem1), validating the approach to data sparsity.
- The certainty score mechanism was shown to improve accuracy by reducing coverage (i.e., by discarding less certain predictions).
- The system achieved a higher accuracy for zero pronoun detection when compared to only considering antecedent identification ($P(a_i|\phi_c)$).
Limitations and Future Work
- Polysemous Verbs: The current system merges multiple case frames for polysemous verbs, which could lead to errors. Future work should explore word sense disambiguation for predicate verbs to select appropriate case frames during zero pronoun detection.
- Rule-Based Model Development: The rule-based model was developed with significant manual effort (two man-months), and its performance, while decent, was surpassed by the probabilistic approach.
- Future Work: The authors suggest exploring word sense disambiguation for polysemous verbs as a key area for future improvement.
Summary updated