Abstract
In conversation, we often ask one-word questions such as ‘Why?’ or ‘Who?’. Such questions are typically easy for humans to answer, but can be hard for computers, because their resolution requires retrieving both the right semantic frames and the right arguments from context. This paper introduces the novel ellipsis resolution task of resolving such one-word questions, referred to as sluices in linguistics. We present a crowd-sourced dataset containing annotations of sluices from over 4,000 dialogues collected from conversational QA datasets, as well as a series of strong baseline architectures.
Introduction
Human communication is remarkably efficient. In a dialogue, a single word like "Why?" or "When?" can carry the weight of a full sentence, provided the context is clear. In linguistics, this phenomenon is known as sluicing—a type of ellipsis where a "wh-word" stands in for an entire clause. While humans resolve these effortlessly, they pose a significant challenge for conversational AI. Resolving a sluice requires a system to identify the relevant semantic frames and arguments from the preceding dialogue history.
Historically, ellipsis resolution has been treated as a span prediction task. However, this paper argues that such an approach is often insufficient for conversational sluicing, where the resolution requires synthesizing information from both a previous question ($Q_1$) and its answer ($A_1$). Instead, the authors frame sluice resolution as a Natural Language Generation (NLG) task: generating the complete, intended question from the dialogue context.
Methodology
The task is modeled as a sequence-to-sequence problem. Given a context sequence—composed of an initial question, its answer, and the follow-up sluice—the model must generate the resolved question. The input is structured using special delimiter tokens: <s> Q1 <del> A1 <del> Q2 </s>.
The researchers evaluated several architectures, including standard LSTM-seq2seq models and Transformer-based encoder-decoders. The most successful approach involved fine-tuning a pre-trained GPT-2 model. This leverages the broad linguistic knowledge acquired during large-scale pre-training to handle the complex structural dependencies inherent in ellipsis.
The Sluice Resolution Dataset
To facilitate research in this area, the authors introduced a novel dataset of over 4,000 annotated sluices. These were extracted from the QuAC and CoQA conversational QA datasets, focusing on five common wh-words: Why, What, Where, Who, and When. Annotations were crowd-sourced via Amazon Mechanical Turk, with the test set featuring multiple human resolutions per instance to account for linguistic variability.
Results and Analysis
Fine-tuned GPT-2 emerged as the superior model, achieving a BLEU score of 0.391, significantly outperforming the LSTM baseline (0.304). Human evaluations further validated these results; in 20% of cases, judges preferred the model's generated resolution over the human "gold" standard, citing higher expressivity.
Attention Dynamics
Visualizing the attention mechanisms reveals how the models process context. The Transformer model typically begins by focusing on the sluice itself ($Q_2$) before "scanning" back through the preceding question and answer to fill in missing structural information.
The study also demonstrated a tangible downstream benefit: replacing sluices with their resolved versions in a standard QA system led to measurable improvements in final answer accuracy. This confirms that sluice resolution is a critical component for robust conversational agents.
Conclusion
This work highlights the importance of treating ellipsis resolution as a generative task in dialogue systems. While fine-tuned language models show near-human performance on the provided dataset, challenges remain. Most sluices in the current data are resolved within a short context window. Future research should explore "long-distance" sluices and more diverse, adversarial dialogue settings to further push the boundaries of conversational understanding.
Summary updated