Source-linked AI summary
Text2Event: Controllable Sequence-to-Structure Generation for End-to-end Event Extraction
Yaojie Lu, Hongyu Lin, Jin Xu, Xianpei Han, Jialong Tang, Annan Li, Le Sun, Meng Liao, Shaoyi Chen
TL;DR
Event extraction involves complex event records and a semantic gap between text and event, while traditional decomposition requires multiple subtasks and fine-grained annotations. TEXT2EVENT instead uses unified sequence-to-structure generation with schema-constrained decoding and curriculum learning. It achieves competitive performance with coarse text-record annotations in supervised and transfer learning, including improvements over sequence-generation baselines on Arg-C F1.
Problem
Event extraction is difficult because event records have complex structures and varied textual expressions, while decomposition-based methods require fine-grained annotations and manually composed subtasks.
Method
TEXT2EVENT uniformly models event extraction as end-to-end sequence-to-structure generation, enhanced with constrained decoding for schema knowledge and curriculum learning.
Results
TEXT2EVENT achieves competitive performance with coarse text-record annotations in supervised and transfer learning, improving Arg-C F1 by 4.6% and 2.7% over two sequence-generation baselines.
Takeaways & Limitations
Uniform sequence-to-structure modeling supports event extraction using only record-level annotations and can motivate structure-prediction approaches for other information extraction tasks.
Takeaways & Limitations
Greedy decoding cannot guarantee valid event structures and may produce invalid event types, argument-type mismatches, or incomplete structures without schema-guided constraints.
Abstract
from arXiv · showhide
Event extraction is challenging due to the complex structure of event records and the semantic gap between text and event. Traditional methods usually extract event records by decomposing the complex structure prediction task into multiple subtasks. In this paper, we propose Text2Event, a sequence-to-structure generation paradigm that can directly extract events from the text in an end-to-end manner. Specifically, we design a sequence-to-structure network for unified event extraction, a constrained decoding algorithm for event knowledge injection during inference, and a curriculum learning algorithm for efficient model learning. Experimental results show that, by uniformly modeling all tasks in a single model and universally predicting different labels, our method can achieve competitive performance using only record-level annotations in both supervised learning and transfer learning settings.
1 Introduction
Event extraction must map text to structurally complex event records despite varied expressions, while decomposition-based methods require fine-grained annotations and manually designed subtask composition. TEXT2EVENT addresses these challenges with end-to-end sequence-to-structure generation, constrained decoding, and curriculum learning.
- Challenges: Event records combine event types, triggers, and arguments in table-like structures, with different event types requiring different structures.Transport and Arrest-Jail events illustrate this structural variation.
- Challenges: The same event can appear through diversified trigger words and heterogeneous syntactic structures.The paper gives “the dismission of the man” and “the man departed his job” as equivalent End-Position expressions.
- Prior methods: Decomposition-based methods split event extraction into entity recognition, trigger detection, and argument classification before composing the predicted structure.The composition may use pipeline modeling, joint modeling, or joint inference.
- Prior methods: These methods need massive fine-grained annotations and can suffer error propagation or suboptimal, inflexible information-sharing designs.The paper contrasts pipeline error propagation with heuristic dependencies in joint models.
- TEXT2EVENT: TEXT2EVENT directly extracts events through a unified sequence-to-structure architecture that universally generates triggers, arguments, and labels as natural-language words.It can be learned from coarse sentence–event-record pairs rather than token-level annotations.
- TEXT2EVENT: The framework adds constrained decoding for schema-guided inference and curriculum learning that progresses from simple substructures to full event generation.Training begins with trigger and independent argument generation before full event-structure generation.
- Results: Experiments on ACE and ERE verify TEXT2EVENT in supervised and transfer learning settings, with the method described as effective, data-efficient, and easy to implement.The paper presents these properties as consequences of uniformly modeling tasks and predicting different labels in one model.
- Broader relevance: The sequence-to-structure formulation is presented as applicable beyond event extraction because many information extraction tasks can be formulated as structure prediction.The authors state that the method can motivate other information extraction models.
2 TEXT2EVENT: End-to-end Event Extraction as Controllable Generation
TEXT2EVENT reformulates event extraction as end-to-end sequence-to-structure generation, using linearized event trees, a transformer encoder-decoder, and schema-constrained decoding. The model generates event labels and text spans directly while restricting outputs to valid structures.
- 2.2 Sequence-to-Structure Network: TEXT2EVENT directly generates complete event structures from input text with a transformer-based encoder-decoder.The model outputs a linearized event representation token by token and transforms it into final event records.
- 2.1 Event Extraction as Structure Generation: Event records become labeled trees whose roots are event types, internal nodes are argument roles, and leaves are linked text spans.The tree is linearized through depth-first traversal, preserving text-span order and supporting multiple event records through a virtual Root.
- 2.1 Event Extraction as Structure Generation: The linearization uses parentheses as structure indicators, with each event connected to a virtual Root and empty-event inputs represented as ().Depth-first traversal produces a deterministic token sequence that can be reversibly converted into event records.
- 2.2 Sequence-to-Structure Network: Because linearized representations use natural-language words, TEXT2EVENT adopts T5 so general text-generation knowledge can be reused.The architecture encodes the input with transformer blocks and predicts the output sequence autoregressively.
- 2.3 Constrained Decoding: Constrained decoding addresses greedy decoding’s inability to guarantee valid event types, compatible argument types, and complete structures.The schema is injected as decoder guidance during inference.
- 2.3 Constrained Decoding: A trie-based constrained decoder dynamically prunes the candidate vocabulary according to the generated state and event schema.Candidates include event types, argument roles, trigger or argument spans, and structure indicators; decoding starts at ⟨bos⟩ and ends at ⟨eos⟩.
3 Learning
TEXT2EVENT is trained end to end from sentence–event-record pairs, but its structural output differs from ordinary text because frequent parentheses carry little semantic information. Curriculum learning therefore progresses from simple substructure generation to full event-structure generation.
- 3 Learning: TEXT2EVENT learns from coarse sentence–event-record pairs without fine-grained token-level annotations.Training uses a negative log-likelihood objective over the parallel text-record dataset.
- 3 Learning: Sequence-to-structure learning is difficult because parentheses create an output gap from natural-language generation and appear frequently while carrying little semantic information.These non-semantic indicators can mislead the learning process.
- 3 Learning: Curriculum learning first trains simple event substructure tasks, then trains the model on the full event structure generation task.The initial tasks generate (label, span) pairs for event types with triggers and roles with arguments.
- 3 Learning: The curriculum strategy uses data annotation more efficiently and makes learning more smooth, according to the authors.Full-structure training follows the substructure-learning stage using the same loss objective.
4 Experiments
TEXT2EVENT is evaluated on multiple event-extraction benchmarks under supervised and transfer-learning settings, using only parallel text-record annotations. It achieves competitive supervised performance, improves over sequence-generation baselines, and transfers knowledge across labels, while constrained decoding and curriculum learning help particularly in low-resource settings.
- Experimental Setup: Experiments use ACE05-EN, ACE05-EN+, and ERE-EN, with evaluation procedures reconstructed for generated trigger and argument offsets.ACE05-EN contains 599 documents and 33 event types; ERE-EN contains 38 categories and 458 documents.
- Supervised Learning Results: TEXT2EVENT uses weaker parallel text-record supervision yet surpasses most token- and entity-annotated baselines with a simpler encoder-decoder architecture.The model also achieves competitive performance with state-of-the-art systems.
- Supervised Learning Results: TEXT2EVENT improves Arg-C F1 by 4.6% and 2.7% over TANL and Multi-task TANL, respectively, among sequence-generation baselines.Direct structure generation can use event schemas during inference and avoid generating irrelevant information.
- Supervised Learning Results: TEXT2EVENT remains nearly stable on the harder ACE05-EN+ dataset, whereas OneIE performance decreases.The authors attribute this robustness to universal modeling and information sharing across labels.
- Transfer Learning Results: 3.7 and 3.2 F1-point gains occur on Trig-C and Arg-C, respectively, when transferring from frequent to remaining event types.The other baselines do not obtain significant improvements on both metrics through transfer learning.
- Detailed Analysis: Constrained decoding helps especially with 1% and 5% training data, while substructure learning improves Trig-C F1 by 4.7% and Arg-C F1 by 5.8% on average.The analysis also reports that natural-language event labels enable transfer from pre-trained language models.
5 Related Work
TEXT2EVENT combines event extraction with neural structure-generation methods in a uniform sequence-to-structure framework. It directly generates event schemas and text spans, using constrained decoding to support varied event types and transfer to new types.
- Prior event-extraction research includes pipeline classification, multi-task joint models, semantic structure grounding, and question-answering approaches.
- TEXT2EVENT uniformly models all event-extraction subtasks in a sequence-to-structure framework, enabling decision interactions and information sharing.
- TEXT2EVENT directly generates event schemas and text spans to form event records through constrained decoding.
- Constrained decoding allows TEXT2EVENT to handle various event types and transfer to new types easily.
6 Conclusions
TEXT2EVENT is a sequence-to-structure generation paradigm that learns event extraction from parallel text-record annotations while integrating constrained decoding and curriculum learning. Experiments report competitive performance with previous state-of-the-art methods in supervised and transfer-learning settings using coarse text-record annotation, and future work targets other information-extraction tasks.
- TEXT2EVENT uniformly models event-extraction subtasks in a sequence-to-structure framework and learns directly from parallel text-record annotation.
- The model combines a sequence-to-structure network with constrained decoding for event-knowledge injection and curriculum learning for efficient training.
- Experiments show competitive performance with previous state-of-the-art methods in supervised and transfer-learning settings using only coarse text-record annotation.
- Future work plans to adapt TEXT2EVENT to other information-extraction tasks, including N-ary relation extraction.