[Paper Review] MLBiNet: A Cross-Sentence Collective Event Detection Network

Dongfang Lou, Zhilin Liao+3 more|arXiv (Cornell University)|May 20, 2021

Topic Modeling42 references4 citations

TL;DR

This paper proposes MLBiNet, a multi-layer bidirectional neural network that jointly models document-level semantic information and cross-sentence event inter-dependency for collective event detection. By employing bidirectional decoders, an information aggregation module, and stacked decoding layers, MLBiNet captures intra- and inter-sentence dependencies, achieving a new state-of-the-art F1 score of 85.7 on the ACE 2005 benchmark, outperforming prior methods by 1.9 points.

ABSTRACT

We consider the problem of collectively detecting multiple events, particularly in cross-sentence settings. The key to dealing with the problem is to encode semantic information and model event inter-dependency at a document-level. In this paper, we reformulate it as a Seq2Seq task and propose a Multi-Layer Bidirectional Network (MLBiNet) to capture the document-level association of events and semantic information simultaneously. Specifically, a bidirectional decoder is firstly devised to model event inter-dependency within a sentence when decoding the event tag vector sequence. Secondly, an information aggregation module is employed to aggregate sentence-level semantic and event tag information. Finally, we stack multiple bidirectional decoders and feed cross-sentence information, forming a multi-layer bidirectional tagging architecture to iteratively propagate information across sentences. We show that our approach provides significant improvement in performance compared to the current state-of-the-art results.

Motivation & Objective

• To address the limitations of existing event detection models that neglect document-level semantic and event inter-dependency information.
• To model both intra-sentence and inter-sentence event dependencies explicitly, especially in cases where event triggers are scattered across multiple sentences.
• To unify document-level semantic aggregation and event inter-dependency modeling into a single, end-to-end trainable framework.
• To improve collective event detection performance by iteratively propagating information across sentences through a stacked decoding architecture.

Proposed method

• A bidirectional decoder is designed to model event inter-dependency within each sentence by capturing forward and backward dependencies in the event tag sequence.
• An information aggregation module summarizes sentence-level semantic and event tag information using pooling operations (e.g., average, max, or LSTM-based aggregation).
• Multiple bidirectional decoder layers are stacked to propagate information across adjacent sentences, enabling long-range context modeling through iterative refinement.
• The model treats event detection as a Seq2Seq task, where the input is a document and the output is a sequence of event tags, allowing for joint prediction of multiple events.
• Cross-sentence information propagation is achieved by feeding the output of one layer to the next, enabling contextual signals to flow across sentences in a hierarchical manner.
• The architecture is trained end-to-end using a cross-entropy loss on the predicted event tag sequences.

Experimental results

Research questions

• RQ1 Can a unified deep learning framework effectively model both document-level semantic information and event inter-dependency for collective event detection?
• RQ2 How does bidirectional decoding within sentences improve the modeling of event dependencies compared to unidirectional or non-recurrent tagging structures?
• RQ3 To what extent can stacked bidirectional decoding layers enhance cross-sentence information propagation and improve detection accuracy?
• RQ4 Does the proposed information aggregation module effectively summarize relevant semantic and event signals from each sentence for downstream layers?
• RQ5 How does MLBiNet compare to state-of-the-art models in terms of F1 score and robustness on the ACE 2005 benchmark?

Key findings

• MLBiNet achieves a new state-of-the-art F1 score of 85.7 on the ACE 2005 benchmark, surpassing the previous best result by 1.9 points.
• The ablation study confirms that both the bidirectional decoder and the multi-layer stacking mechanism significantly contribute to performance gains.
• The information aggregation module using LSTM outperforms average and max pooling, indicating that sequential modeling of event signals enhances detection.
• The model effectively captures cross-sentence dependencies, as evidenced by improved detection of event triggers like 'firing' and 'death' when context from neighboring sentences is available.
• The performance gain is particularly notable in cases involving co-occurring events, where 44.4% of triggers in the ACE 2005 corpus appear in such configurations.
• The model demonstrates robustness in handling ambiguous triggers—such as 'firing' which can denote either 'Attack' or 'End_Position'—by leveraging cross-sentence context.