Abstract
Information Status (IS) classification is a critical task in discourse analysis, identifying how entities are introduced and tracked within a text. Traditionally, this task has relied on pre-identified "gold" mentions or error-prone syntactic parsers. We present the first end-to-end neural system for IS classification that utilizes a boundary-representation approach for mention extraction and a context-aware classifier. Our model achieves state-of-the-art performance on the ISNotes benchmark and demonstrates robust generalization across diverse domains, including scientific literature.
Introduction
In linguistics, Information Status (IS) categorizes discourse entities based on their cognitive accessibility to the reader. Mentions are typically classified as "old" (previously mentioned), "new" (introduced for the first time), or "mediated" (inferable from context). Historically, IS classification has been bottlenecked by the assumption that exact text spans representing entities—"gold mentions"—are provided. In practical applications, an automated system must identify these mentions autonomously.
Furthermore, while coreference resolution systems excel at linking recurring entities, IS classification requires identifying all mentions, including "singletons" (entities mentioned only once). Previous methods often relied on syntactic parsers, which struggle with cross-domain generalization and error propagation. This paper addresses these challenges by proposing a fully neural, end-to-end pipeline.
Methodology
The proposed system consists of two primary components: a Mention Extraction model (MenExt) and an Information Status Assignment model (ISAssign).
Mention Extraction (MenExt)
Instead of relying on syntactic trees, MenExt treats mention extraction as a binary classification task for all possible spans within a sentence. Using a RoBERTa-Large encoder, the model represents a span by inserting special boundary tokens ([SEP1] and [SEP2]) and concatenating their hidden states. This approach allows the model to reason over spans of any length without arbitrary caps.
IS Assignment (ISAssign)
Once mentions are identified, ISAssign categorizes them into one of eight IS classes. It utilizes the same boundary-token representation but incorporates a "previous string match" indicator. This binary flag indicates if the mention's text has appeared earlier in the document, providing a strong signal for "old" status.
Experimental Setup
The system was evaluated on three distinct datasets:
- ISNotes: A benchmark of 50 Wall Street Journal articles containing nearly 11,000 annotated mentions.
- BASHI: 50 WSJ texts focused on "bridging" anaphora—mentions linked to previous entities via association.
- SciCorp: 14 scientific papers used to test cross-domain generalization from news to technical literature.
Results and Discussion
The neural pipeline significantly outperforms previous baselines across multiple metrics.
Mention Extraction Performance
The MenExt model achieved an F1-score of 92.1, a substantial improvement over the previous neural state-of-the-art of 87.7. Notably, the model maintains high performance even for longer mentions, where traditional syntactic methods often fail.
IS Classification and End-to-End Accuracy
On gold mentions, the system reached 85.1% accuracy on ISNotes. In the more challenging end-to-end setting—where the model must both find and classify mentions—it achieved an overall F1 of 78.3. Regarding bridging anaphora, the system achieved an F1 of 35.8 on BASHI and 41.0 on SciCorp, beating many specialized models.
Cross-Domain Generalization
A key finding is the model's ability to generalize. When trained on news (ISNotes) and tested on scientific papers (SciCorp), the system remained competitive. This suggests that the structural patterns of information introduction are relatively consistent across genres, despite significant differences in vocabulary.
Conclusion
This work demonstrates that neural boundary representations are more effective for IS classification than traditional syntactic approaches. By removing the dependency on gold mentions and parsers, the system provides a more realistic and robust solution for discourse analysis. Future research will focus on improving the detection of bridging anaphora, which remains the "final boss" of IS classification due to the complex relational reasoning required.
Summary updated