TL;DR
This paper introduces a novel unsupervised probabilistic model for Chinese event coreference resolution, achieving performance comparable to state-of-the-art supervised methods. By jointly modeling anaphoricity and coreference without relying on language-specific lexical resources, this approach offers a promising direction for event coreference in low-resource languages.
What Problem Does This Paper Address?
Event coreference resolution, the task of identifying mentions of the same real-world event in text, is significantly less studied and more challenging than entity coreference. This is because not only the event's "trigger" word but also its arguments (like time, place, and participants) must be compatible. While previous work on English event coreference has leveraged lexical resources like WordNet and FrameNet, such resources are scarce for Chinese. This lack of resources makes building high-performance event coreference resolvers, especially unsupervised ones, difficult for Chinese. The paper tackles this challenge by proposing an unsupervised probabilistic model for Chinese event coreference resolution that aims to rival supervised methods.
Methods
The core of the paper is a generative probabilistic model that jointly handles event coreference resolution and anaphoricity determination. Anaphoricity determination, deciding if an event mention needs a coreferent antecedent, is crucial because not all event mentions refer to actual events. The model uses an Expectation-Maximization (EM) algorithm to estimate its parameters from unannotated data. It models the probability of a candidate antecedent being correct given the active event mention and its context. A key aspect is the inclusion of a "dummy" candidate antecedent, allowing the model to learn when an event mention is non-anaphoric (i.e., doesn't refer to a true event or is a singleton). The model also incorporates several context features, including trigger-based features (lexical identity, basic verb compatibility, word similarity) and argument-based features (semantic role and class compatibility, named entity type compatibility), as well as a distance feature.
Datasets
The experiments were conducted on the Chinese portion of the ACE 2005 corpus, which contains 633 documents. The paper uses this corpus for five-fold cross-validation.
Key Insights and Ideas
- Unsupervised Chinese Event Coreference is Feasible: The paper demonstrates that it's possible to achieve competitive performance in Chinese event coreference resolution without supervised training data or language-specific lexical resources.
- Joint Anaphoricity and Coreference: Modeling anaphoricity determination alongside coreference resolution is crucial and beneficial, especially given the difficulty of identifying anaphoric event mentions.
- Generative Model's Language Independence: The underlying generative process of the model is language-independent, suggesting its potential applicability to other languages lacking extensive lexical resources.
- Importance of Trigger-Based Features: Trigger-based features, particularly those related to lexical identity, basic verb compatibility, and word similarity, are shown to be highly influential for recall.
- The "Dummy" Antecedent is Key: Including a dummy antecedent is vital for jointly modeling anaphoricity and coreference, allowing the model to learn to resolve mentions as non-anaphoric.
Key Results
The proposed unsupervised probabilistic model achieved a CoNLL score of 41.5 on the ACE 2005 Chinese corpus. This significantly outperformed the best supervised baseline, SinoCoreferencer, which achieved a CoNLL score of 39.2. This 2.3-point improvement highlights the effectiveness of the unsupervised approach. Ablation studies showed that the trigger-based feature (Feature 1) was the most impactful, with its removal causing a significant drop in the CoNLL score (5.3 points). The feature related to candidate antecedent plausibility based on event subtypes (P(c|k)) was the second most useful, with its removal leading to a 1.6-point drop.
Limitations and Future Work
The authors acknowledge that the performance of their resolver is limited by the errors made by upstream components of the information extraction pipeline (e.g., entity coreference, event extraction). Future work is planned to address this by performing joint inference for entity coreference, event extraction, and event coreference. Additionally, the paper notes that certain latent attributes like modality, polarity, and tense are not exploited, which could be a source of error. The model also doesn't explicitly handle timestamps for events, which can lead to errors when events occur at different times.
Summary updated