Identifying Various Kinds of Event Mentions in K-Parser Output

Arpit Sharma, Nguyen Ha Vo, Somak Aditya, Chitta BaralNAACL 2015 · DOI · OpenReview

Abstract

K-Parser is a semantic parser designed to identify and represent complex event mentions in text as interconnected acyclic graphs. By mapping syntactic dependencies to a rich ontology and applying common-sense reasoning, it can distinguish between simple temporal sequences and recursive events where one action serves as an argument for another. This paper describes the K-Parser architecture and evaluates its performance on the Winograd Schema Challenge, demonstrating high precision in event and entity identification.

Introduction

In Natural Language Understanding (NLU), identifying "what" happened is as critical as identifying the participants involved. While many semantic parsers excel at extracting entities like names and dates, they often struggle with the nuanced relationships between events. For instance, in the sentence "The knife was used for killing the dog," there are two distinct yet hierarchically linked events: a "using" event and a "killing" event. Most existing parsers fail to capture this recursive nature or effectively categorize events into specific aspectual types. K-Parser addresses these gaps by providing a graphical output that captures deep semantics and event interactions.

The K-Parser Pipeline

K-Parser transforms raw text into a semantic graph through a five-module pipeline:

  1. Syntactic Parsing: Utilizes the Stanford Dependency Parser to generate a traditional dependency graph of the sentence.
  2. Semantic Mapping: Maps syntactic dependencies to ontological relations from the KM (Component Library). It uses hand-crafted rules, a Multi-Layer Perceptron (MLP) classifier for preposition disambiguation, and discourse connectives to label event-event relations.
  3. Conceptual Classification: Employs Word Sense Disambiguation (WSD) via WordNet to assign conceptual classes (e.g., identifying "Barack Obama" as a person).
  4. Common-Sense Correction: A novel step that uses conceptual classes to rectify mapping errors. For example, it corrects a "possession" relation between two humans to a related_to relation, as humans typically do not possess each other in the KM framework.
  5. Feature Enrichment: Adds tenses, semantic roles (via Propbank), and co-reference resolution to link multiple mentions of the same entity.

Representing Complex and Recursive Events

A core philosophy of K-Parser is that verbs drive events. Every verb becomes the root of a subgraph. The parser distinguishes between four aspectual types—achievements, accomplishments, processes, and states—based on verb arguments like destination or time frames.

Crucially, K-Parser handles "complex" events where one event serves as an argument for another. In the phrase "used for killing," the "killing" event is the objective of the "using" event. This is represented as a nested relationship rather than a flat list of actions.

Figure 3: A complex event graph generated by K-Parser for the sentence "The knife was used for killing the dog."

Evaluation and Results

The system was evaluated using 282 sentences from the Winograd Schema Challenge (WSC), a benchmark requiring deep semantic understanding and common-sense reasoning. Because the WSC requires resolving ambiguity through context, it serves as a rigorous test for event identification.

Figure 4: Evaluation results showing precision and recall across various categories.

The results indicate high precision in identifying basic building blocks. The precision of 0.91 for Event-Event relations is particularly noteworthy given the difficulty of the WSC. However, a lower recall of 0.79 suggests that some complex linguistic structures remain challenging for the current rule-set.

Conclusion

K-Parser demonstrates a robust approach to event identification by integrating symbolic reasoning with statistical parsing. While it achieves high accuracy in entity and event detection, future work will focus on improving recall for event-event relations and expanding the rule-set to handle more idiomatic language and diverse linguistic edge cases.

Summary updated

BibTeX
@inproceedings{sharma2015identifying,
  author = {Arpit Sharma and Nguyen Ha Vo and Somak Aditya and Chitta Baral},
  title = {{Identifying Various Kinds of Event Mentions in K-Parser Output}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2015},
  doi = {10.3115/v1/w15-0811},
}
Figure