DICP: Deep In-Context Prompt for Event Causality Identification

Lin Mu, Jun Shen, Li Ni, Lei Sang, Zhize Wu, Peiquan Jin, Yiwen ZhangEMNLP 2025

TL;DR

This paper introduces DICP, a novel method for event causality identification (ECI) that injects in-context examples into deeper layers of pre-trained language models (PLMs). By leveraging hierarchical semantic representations and a multi-layer injection strategy, DICP significantly improves the model's ability to capture complex and implicit causal relationships, outperforming existing state-of-the-art methods.

What Problem Does This Paper Address?

Event causality identification (ECI) is a crucial task in natural language processing (NLP) that aims to determine causal relationships between events mentioned in text. While explicit causal cues (like "because") make this easier, many causal relationships are implicit, requiring a deeper understanding of context and semantics. Existing prompt-learning methods for ECI typically concatenate examples only at the input layer of PLMs. This shallow integration limits the model's capacity to fully leverage the abstract semantic information present in these examples, especially for complex or implicit causalities. The paper addresses this limitation by proposing a method to inject these examples into deeper layers of the PLM, where more abstract semantic representations are formed.

Methods

The proposed DICP framework integrates three key components:

  • Encoding In-Context Module: This module takes a set of diverse in-context examples (each containing an event pair, sentence, and causal label) and encodes them into semantic embeddings.
  • Deep Prompt Learning Module: This is the core innovation. Instead of just prepending examples to the input, DICP injects these encoded in-context embeddings into multiple deeper layers of the PLM's encoder. This allows the model to learn from task-specific patterns at various levels of abstraction. The input to each layer is a concatenation of the injected contextual embedding and the output from the previous layer.
  • Semantic Structure Module: This module enhances the model by incorporating explicit semantic relationships extracted from the text. It uses an Abstract Meaning Representation (AMR) parser to build a semantic graph, from which an event-centric structure (enriching event representations) and an event-association structure (modeling semantic paths between events) are extracted. These structures are then combined to form a final semantic representation.

The final representation for the [MASK] token (used for prediction) is a concatenation of the output from the deep prompt learning module and the semantic structure module. The model is trained using a cross-entropy loss function.

Datasets

The paper evaluates DICP on two widely used datasets for event causality identification:

  • EventStoryLine Corpus version 0.9 (ESC): Contains 258 documents, 4316 sentences, 5334 event mentions, and 1770 causal event pairs.
  • Causal-TimeBank (CTB): Contains 184 documents, 6813 event mentions, and 318 causal event pairs.

The evaluation uses 5-fold cross-validation on ESC and 10-fold cross-validation on CTB, measuring performance using Precision (P), Recall (R), and F1-score (F1).

Key Insights and Ideas

  • Deeper is Better for In-Context Learning: Injecting in-context examples into deeper layers of a PLM significantly boosts performance compared to input-layer-only concatenation. This suggests that abstract semantic representations in deeper layers are more conducive to learning complex causal relationships.
  • Multi-Layer Injection is Key: Distributing diverse in-context examples across multiple deeper layers allows the model to capture a broader range of causal patterns and improves generalization.
  • Semantic Structure Complements Prompting: Integrating explicit semantic information (via AMR) alongside deep in-context learning further enhances the model's ability to identify causal relationships, especially implicit ones.
  • Optimal Depth for Injection: There's an optimal number of deep layers for injecting prompts (found to be 2 on ESC). Too few layers limit information retrieval, while too many can introduce noise and interfere with decision-making.

Key Results

DICP demonstrates significant improvements over existing methods:

  • On the ESC dataset, DICP achieves an F1-score of 68.2%, outperforming previous state-of-the-art methods.
  • On the CTB dataset, DICP achieves an F1-score of 61.4%.
  • DICP outperforms large language models (LLMs) like GPT-4, achieving a 41.5% improvement in F1 over GPT-4 on ESC, highlighting the benefit of fine-tuning for specialized tasks.
  • Ablation studies show that removing the semantic structure module leads to a 1.9% decrease in F1, while removing in-context examples (DICPw/o.icl) results in a 4.1% decrease. Omitting deep layer injection (DICPw/o.deep) leads to a 2.9% decrease in F1.

Limitations and Future Work

The authors acknowledge two main limitations:

  1. Uniform Treatment of Examples: The current method assigns equal importance to all injected in-context examples, which might limit the model's ability to leverage the most relevant information. Future work could explore assigning different levels of importance to examples.
  2. Risk of Overfitting: Injecting examples into deeper layers could increase the risk of overfitting, making the model overly reliant on the provided examples and potentially overlooking crucial information in the input text.

Future work could also explore more sophisticated strategies for selecting and weighting in-context examples and investigate the impact of different PLM architectures.

Summary updated

BibTeX
@inproceedings{mu2025dicp,
  author = {Lin Mu and Jun Shen and Li Ni and Lei Sang and Zhize Wu and Peiquan Jin and Yiwen Zhang},
  title = {{DICP: Deep In-Context Prompt for Event Causality Identification}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2025},
}
Figure