Improving Zero-shot Sentence Decontextualisation with Content Selection and Planning

Zhenyun Deng, Yulong Chen, Andreas VlachosEMNLP 2025

TL;DR

This paper introduces ECSP, a novel framework for zero-shot sentence decontextualization that enhances understanding by selecting and planning relevant contextual information. By breaking sentences into Elementary Discourse Units (EDUs) and strategically enriching ambiguous ones, ECSP significantly improves the clarity and coherence of decontextualized sentences, outperforming existing methods.

What Problem Does This Paper Address?

Many Natural Language Processing (NLP) tasks, such as summarization and question answering, rely on extracting individual sentences from larger documents. However, these extracted sentences often lack the necessary context (like coreferences or background information) to be fully understood on their own. This problem, known as sentence decontextualization, aims to rewrite sentences so they are self-contained and comprehensible without their original surrounding text. Previous approaches have struggled to effectively capture and integrate implicit discourse information, leaving some ambiguities unresolved. This paper tackles this challenge by proposing a more robust method for identifying and incorporating crucial contextual elements.

Methods

The core of the ECSP framework lies in its two-stage process: Content Selection and Content Planning.

  1. Content Selection:
    • EDU Segmentation: The input sentence and its context are first broken down into Elementary Discourse Units (EDUs) – basic, semantically independent spans of text. This is achieved by prompting a Large Language Model (LLM).
  2. Ambiguous EDU Identification: Within the segmented sentence, EDUs that are likely to be unclear without context (ambiguous EDUs) are identified.
  3. Relevant EDU Selection: For each ambiguous EDU, the system identifies and extracts "dominant" EDUs from the context that have discourse relations with the ambiguous EDU. These dominant EDUs serve as necessary context to clarify the ambiguity.
  4. Content Planning:
    • EDU Decontextualization: The identified ambiguous EDUs are then enriched with their selected relevant EDUs. This process aims to resolve ambiguities, determine references, and fill in implicit information. The goal is to rewrite the sentence while staying as close as possible to its original form.

This approach leverages LLMs for segmentation, identification, and selection, and then uses a structured planning process to generate the decontextualized sentence.

Datasets

The primary dataset used for evaluation is from Choi et al. (2021), a benchmark dataset for sentence decontextualization. This dataset consists of triplets: (sentence, context, decontextualized sentence), where the context is the paragraph containing the sentence, and the decontextualized sentence is the target output. The authors also mention evaluating on a fact-checking claim extraction dataset (Deng et al., 2024) and a multi-hop QA dataset (Yang et al., 2018b) to demonstrate the impact of their method on downstream tasks.

Key Insights and Ideas

  • EDU Granularity is Key: The paper argues that using Elementary Discourse Units (EDUs) as the fundamental building blocks for decontextualization is crucial. EDUs provide a richer semantic and finer-grained discourse structure than simply using entire sentences or phrases.
  • Discourse Relations Matter: Identifying and leveraging discourse relations between EDUs (e.g., Elaboration, Location) is essential for understanding and resolving ambiguities. The framework explicitly uses these relations to select relevant context.
  • Content Selection Improves Efficiency: By strategically selecting only the most relevant contextual EDUs, ECSP significantly reduces the amount of context needed for decontextualization, making the process more efficient without sacrificing quality.
  • Decontextualization Aids Downstream Tasks: The paper demonstrates that decontextualized sentences produced by ECSP lead to improved performance in downstream tasks like multi-hop evidence retrieval and reasoning, highlighting the practical utility of the method.

Key Results

  • State-of-the-Art Performance: ECSP outperforms existing zero-shot baselines (QADECONTEXT, DCE, Vanilla Prompt) across multiple metrics (SARI, BERTScore, ChrF, RougeL, BLEU, METEOR).
  • Improved EDU Segmentation: ECSP's EDU segmentation component achieves better semantic integrity and coherence compared to baselines like SEGBOT and NeuralSeg.
  • Effective Ambiguous EDU Identification: The method achieves high precision (up to 87.50% with GPT-40) in identifying ambiguous EDUs, significantly outperforming a coreference model.
  • Downstream Task Benefits: ECSP shows a 0.44 improvement in F1 score for evidence retrieval and a 1.58 improvement in F1 score for answer prediction in multi-hop QA tasks.
  • Efficiency Gains: While ECSP involves more LLM calls than vanilla prompting, it achieves significantly better decontextualization quality, demonstrating a favorable trade-off.

Limitations and Future Work

  • Limitations:
    • The method does not attempt to identify exact ambiguous text spans, focusing instead on identifying ambiguous EDUs.
  • It relies on LLMs for EDU segmentation, and while generally effective, improper segmentation can still impact performance.
  • For different types of ambiguous EDUs, multiple decontextualizations might be considered correct, suggesting room for more flexible content planning.
  • The method's performance is tied to the capacity of the underlying LLMs, though it shows universality across different LLMs.
  • Future Work:
    • Extending the framework's capability to more complex scenarios, including multimodal tasks.
  • Further exploration of more flexible content planning for different types of ambiguous EDUs.

Summary updated

BibTeX
@inproceedings{deng2025improving,
  author = {Zhenyun Deng and Yulong Chen and Andreas Vlachos},
  title = {{Improving Zero-shot Sentence Decontextualisation with Content Selection and Planning}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2025},
}
Figure