📄 Using Explicit Discourse Connectives in Translation for Implicit Discourse Relation Classification

Summary

Classifying implicit discourse relations, which are relations between text spans that no connective such as "because" or "but" signals, remains a weak point of discourse parsing. One main reason is that PDTB offers little labeled data. The authors propose a way to generate extra training examples from English–French parallel text. Human translators sometimes add a connective that the source lacked (explicitation), and this added connective reveals the intended relation.

The pipeline works in four steps:

1. A pre-trained MT system translates the French side back into English.
2. An end-to-end English discourse parser runs on both the original and the back-translated text.
3. The method keeps consecutive sentence pairs that the parser marks as implicit in the source but explicit in the back-translation.
4. Each kept source pair receives the label of the explicit relation.

This yields a training set several times larger than PDTB's. With this added data, a plain bidirectional LSTM beats prior state-of-the-art systems on the PDTB 2.0 test set and in cross-validation. The approach does not depend on any particular classifier, so other implicit relation models could reuse it.

Key points

• PDTB provides only 12763 implicit training instances and a 761-relation test set. Some second-level senses have about a dozen examples.
• Back-translating into English lets the method use an English parser and the PDTB label set instead of a target-language parser. MT quality matters mainly for translating the connectives correctly.
• Accuracy rises with more mined intra-sentential samples but levels off around 40,000. That setting gave the best averaged accuracy, 45.87%, which beat using all mined samples.
• Training on mined samples alone, without PDTB data, performed worse than using PDTB only.
• The mined data contains noise, for example "if" wrongly tagged as a connective, which inflates Contingency.Condition. It also captures relation patterns that PDTB does not annotate, such as temporal readings of gerund constructions.