Description of the Odin Event Extraction Framework and Rule Language

Marco Antonio Valenzuela-Escárcega, Gus Hahn-Powell, Mihai SurdeanuarXiv 2015 · arXiv

Introduction

Information Extraction (IE) has seen a significant shift toward machine learning (ML) models. While these models are powerful, they often function as "black boxes," requiring extensive labeled datasets and presenting challenges for domain adaptation. Rule-based systems, conversely, offer high interpretability and easier domain transfer but have historically suffered from a lack of standardized, powerful languages and high entry costs.

Odin (Open Domain INformer) is a domain-independent framework designed to bridge this gap. It provides a declarative, YAML-based rule language that allows researchers to build robust extraction models quickly. Odin is engineered to be simple to use, powerful enough for complex syntactic structures, robust against parsing errors, and efficient enough for large-scale applications, processing over 100 sentences per second.

The Odin Rule Language

At the heart of Odin is its declarative rule language. Rules are defined using a subset of YAML, making them human-readable and easy to modify. This transparency allows domain experts, such as biologists or clinicians, to directly inspect and refine the extraction logic without deep programming knowledge.

Table 1: Fields and default values for Odin rules.

As shown in Table 1, each rule consists of several fields that define its behavior. Key fields include name, label, and pattern. The priority field allows for iterative execution, while action enables the integration of custom code for complex tasks like coreference resolution.

Methodology

Odin employs a hybrid approach to pattern matching, combining surface-level and syntactic analysis:

Token and Dependency Patterns

  • Token Patterns: These surface-level patterns match sequences of tokens based on text, lemmas, part-of-speech (POS) tags, or named entity labels. They function similarly to regular expressions but operate on structured token data.
  • Dependency Patterns: To capture complex relationships, Odin traverses syntactic dependency graphs (e.g., Stanford Dependencies). It introduces specialized operators like > (outgoing edge) and < (incoming edge) to navigate these trees efficiently.

Execution Strategy

To maintain high throughput, Odin utilizes Trigger-Based Execution. It uses "event triggers" (e.g., specific verbs like "phosphorylates") as shallow filters. The system only invokes expensive syntactic matching once a relevant trigger is identified in the text.

Furthermore, Odin supports Recursive Cascades. Rules are applied in numbered iterations. The output of one rule—a "Mention"—can serve as an argument for subsequent rules. This allows for the extraction of deeply nested or recursive events, such as a phosphorylation event that serves as a cause for a larger regulatory process.

Performance and Robustness

Odin is designed for real-world utility. In benchmarks using a biomedical grammar of 211 rules, the framework achieved a throughput of approximately 110 sentences per second on standard consumer hardware.

A critical design philosophy in Odin is Robustness through Hybridization. Because syntactic parsers are not infallible, Odin allows developers to mix dependency patterns with surface patterns. If a parser fails to identify a complex dependency, a fallback token-based rule can still capture the necessary information. Additionally, Odin integrates a taxonomy system, where labels can be hierarchical (e.g., an "Animalia" mention automatically matches a rule looking for "Organism"), enhancing the flexibility of the extraction process.

Conclusion

Odin provides a powerful alternative to black-box ML models for event extraction. By prioritizing interpretability, speed, and ease of use, it lowers the barrier for developing high-quality IE systems in specialized domains. While currently dependent on external NLP pre-processing pipelines and manual rule engineering, Odin represents a significant step forward in making rule-based extraction scalable and accessible. Future work may explore automated rule induction to further reduce the manual effort required for grammar development.

Summary updated

Abstract

This document describes the Odin framework, which is a domain-independent platform for developing rule-based event extraction models. Odin aims to be powerful (the rule language allows the modeling of complex syntactic structures) and robust (to recover from syntactic parsing errors, syntactic patterns can be freely mixed with surface, token-based patterns), while remaining simple (some domain grammars can be up and running in minutes), and fast (Odin processes over 100 sentences/second in a real-world domain with over 200 rules). Here we include a thorough definition of the Odin rule language, together with a description of the Odin API in the Scala language, which allows one to apply these rules to arbitrary texts.
BibTeX
@misc{valenzuelaescarcega2015description,
  author = {Marco Antonio Valenzuela-Escárcega and Gus Hahn-Powell and Mihai Surdeanu},
  title = {{Description of the Odin Event Extraction Framework and Rule Language}},
  year = {2015},
  eprint = {1509.07513},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1509.07513},
}
Figure