Abstract: Zixuan Zhang, Heng Ji. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies.
2021.
Introduction: Bridging the Gap Between Semantics and Extraction
Information Extraction (IE) is the critical task of transforming unstructured text into structured networks of entities, relations, and events. Traditionally, IE models have treated sentences as linear sequences or utilized basic syntactic dependency trees. However, these approaches often struggle with "non-local" dependencies—instances where two related elements are separated by many words in a sentence, making their connection difficult for a model to learn from scratch.
This paper introduces a novel framework that leverages Abstract Meaning Representation (AMR)—a rooted, directed graph representing "who is doing what to whom"—to guide the IE process. The authors argue that Semantic Parsing and Information Extraction are two sides of the same coin: both seek to summarize sentence meaning in a graph format. By utilizing a pre-trained AMR parser, the IE model is provided with a "semantic roadmap" that highlights how distant words are closely connected in meaning.
Methodology: AMR-Guided Encoding and Decoding
The authors propose the AMR-IE framework, which consists of two primary innovations designed to integrate semantic topology into the extraction pipeline:
1. AMR-Guided Graph Encoding
The model first identifies candidate entity mentions and event triggers, then aligns these candidates to nodes in an AMR graph. Using a Graph Attention Network (GAT), the model allows these nodes to "message" their neighbors within the AMR structure. For instance, a node representing a person can aggregate information from distant action nodes (like "murdering" or "faces") simultaneously, regardless of their linear distance in the sentence. To prevent overfitting on the hundreds of specific AMR edge types, the authors manually grouped them into 12 semantic clusters (e.g., Spatial, Temporal, Core Roles).
2. AMR-Conditioned Graph Decoding
While most joint IE models decode results in a simple left-to-right order based on word position, AMR-IE introduces Hierarchical Ordered Decoding. This method sorts candidate nodes based on their depth in the AMR tree (top-down). By decoding the "root" concepts first, the model builds the information network in a logically organized manner, improving the efficiency and accuracy of the beam search used to find the optimal global output.
Experimental Results and Analysis
The model was evaluated on several diverse benchmarks, including ACE05-E, ERE-EN, and the biomedical BioNLP Genia datasets. AMR-IE achieved new state-of-the-art (SOTA) results across all subtasks on ACE05-E and ERE-EN.
The gains were most significant in "edge" tasks such as Relation Extraction and Event Argument Role Labeling. On ACE05-E, the model improved the Event Argument F1-score by nearly 3% over the previous SOTA (OneIE). Qualitative analysis shows that AMR-IE is particularly adept at identifying arguments that linear models miss by following the semantic paths provided by the AMR parser.
Key Insights and Future Directions
- Semantic Overlap: The core success of the model stems from the fact that an IE network is essentially a subset of an AMR graph. If the AMR parser identifies a relationship between two concepts, there is a high probability they share an IE relation.
- Topology as a Shortcut: AMR provides a direct edge between distant but related words, removing the need for the model to learn complex linguistic patterns to link them.
- Cross-Domain Generalization: The model's success on Genia biomedical datasets proves that general-purpose AMR parsers provide useful signals even in highly technical, specialized domains.
Despite its success, the framework is susceptible to error propagation; if the initial AMR parser produces an incorrect graph, the IE model may be misled. Future work aims to extend this framework to document-level extraction by integrating entity coreference graphs to link information across multiple sentences.
Summary updated