The DCU Discourse Parser for Connective, Argument Identification and Explicit Sense Classification
Summary
This paper presents Dublin City University's entry to the CoNLL-2015 shared task on shallow discourse parsing. The system follows the pipeline design of Lin et al. (2014) and splits the problem into seven components. These are connective detection, argument position classification, three argument extractors, explicit sense classification, and a non-explicit sense step. The team concentrated on explicit relations. Connective detection, argument position, and explicit sense use maximum entropy classifiers with hand-tuned lexical and syntactic features, with parses produced by the Berkeley Parser. Argument span extraction is cast as token-level sequence labelling with recurrent neural networks over word embeddings, which also helps with sparse training data and unseen words. Non-explicit relations are handled only by a majority-class baseline. Overall F1 is modest: 16.51% on dev, 12.73% on the WSJ test set and 11.15% on the blind test. Connective detection reached about 80%, and sense classification was the weakest step. The paper is useful as an early example of combining feature-based classifiers with RNN span labelling for PDTB-style parsing. It also shows clearly how errors accumulate across a pipeline.
Key points
- Data: CoNLL-2015 shared task release (LDC2015E21, PDTB-based). Connective candidates come from the 99 connective types found in the PDTB training corpus.
- Argument position classifier (same-sentence vs. other) uses 11 of 17 candidate features and reaches 97.78% accuracy on dev. Same-sentence and previous-sentence cases together cover 90.20% of explicit relations.
- Argument extraction uses Theano RNNs with IOBE labels and Skip-gram (word2vec) embeddings. Tuned settings are a context window of 5, 300 hidden units, 10 iterations and 100-dimensional embeddings. Inter-sentential spans remained hard to extract.
- All non-explicit candidates (Implicit, AltLex, EntRel, NoRel) are labelled EntRel. This, together with propagated pipeline errors, explains the roughly 8% sense F1.
- Official overall F1: 16.51% (dev), 12.73% (WSJ test), 11.15% (blind test). Decoding ran on the TIRA platform.
Cited by
- Contextualized Embeddings for Connective Disambiguation in Shallow Discourse Parsing
- The CoNLL-2015 Shared Task on Shallow Discourse Parsing
Cite
@inproceedings{wang2015dcu,
title = {{The DCU Discourse Parser for Connective, Argument Identification and Explicit Sense Classification}},
author = {Wang, Longyue and Hokamp, Chris and Okita, Tsuyoshi and Zhang, Xiaojun and Liu, Qun},
booktitle = {Proceedings of the Nineteenth Conference on Computational Natural Language Learning - Shared Task},
year = {2015},
doi = {10.18653/v1/k15-2014},
url = {https://doi.org/10.18653/v1/k15-2014}
}References
- Ziheng Lin, Hwee Tou Ng and Min-Yen Kan. 2014. A PDTB-styled End-to-End Discourse Parser, volume 1-34. Natural Language Engineering. find
- Sucheta Ghosh, Richard Johansson and Sara Tonelli. 2011. Shallow Discourse Parsing with Conditional Random Fields. In Proceedings of the 5th International Joint Conference on Natural Language Processing. find our card
- Features Connective Word Connective POS Connective + 1st Previous Word Connective + 2st Previous Word Connective + 1st Previous POS Low-Cased Connective Sucheta Ghosh, Giuseppe Riccardi and Richard Johansson. 2012. Global Features for Shallow Discourse Parsing, 150-159. In Proceedings of the 13th An… find
- Attapol T. Rutherford and Nianwen Xue. 2014. Discovering Implicit Discourse Relations Through Brown Cluster Pair Representation and Coreference Patterns, 645. In Proceedings of EACL 2014. find our card
- Rashmi Prasad, Nikhil Dinesh, Alan Lee, Eleni Miltsakaki, Livio Robaldo, Aravind Joshi, Bonnie Webber. 2008. The Penn Discourse TreeBank 2.0. In Proceedings of the LREC. find our card
- Grgoire Mesnil, Xiaodong He, Li Deng and Yoshua Bengio. 2013. Investigation of Recurrent-Neural-Network Architectures and Learning Methods for Spoken Language Understanding. In Proceedings of the Interspeech 2013. find
- Christian Raymond and Giuseppe Riccardi. 2007. Generative and discriminative algorithms for spoken language understanding. In Proceedings of the Interspeech 2007. find
- Frdric Bastien, Pascal Lamblin, Razvan Pascanu, James Bergstra, Ian Goodfellow, Arnaud Bergeron, Nicolas Bouchard, David Warde-Farley, Yoshua Bengio. 2012. Theano: new features and speed improvements. In Proceedings of the Python for Scientific Computing Conference (SciPy). find
- James Bergstra, Olivier Breuleux, Frdric Bastien, Pascal Lamblin, Razvan Pascanu, Guillaume Desjardins, Joseph Turian, David Warde-Farley, Yoshua Bengio. 2010. Theano: a CPU and GPU math expression compiler. In Proceedings of the Python for Scientific Computing Conference (SciPy). find
- Evgeny A. Stepanov and Giuseppe Riccardi. 2013. Comparative evaluation of argument extraction algorithms in discourse relation parsing. In Proceedings of 13th International Conference on Parsing Technologies (IWPT 2013). find our card
- Components Connectives Arg1 Arg2 Arg1 & Arg2 Sense Overall Precision 0.9010 0.3437 0.3778 0.2559 0.3194 0.1420 F1 0.8565 0.3995 0.4392 0.2975 0.0938 0.1651 Table 5: Official Results. find
- Xu Ming, Zhu Qiao and Zhou Guo Dong. 2012. A Unified Framework for Discourse Argument Identification via Shallow Semantic Parsing. In Proceedings of 24th International Conference on Computational Linguistics. find
- Test Set Recall 0.8570 0.4384 0.5034 0.3074 0.1124 0.1537 Blind Set Recall 0.7464 0.3755 0.4690 0.2589 0.0701 0.1307 Precision 0.9040 0.3100 0.3559 0.2174 0.2257 0.1087 F1 0.8799 0.3632 0.4170 0.2546 0.0849 0.1273 Precision 0.8487 0.2794 0.3489 0.1926 0.0905 0.0972 find
This page in Markdown → https://md.research.elamur.ai/free/en/papers/wang-2015-dcu-discourse-parser-connective-argument/