-- General Information
[Abstract] Previous studies on RST-style discourse segmentation have achieved impressive results.
However, recent neural works either require a complex joint training process or heavily rely on
powerful pre-trained word vectors. Under this condition, a simpler but more robust segmentation
method is needed. In this work, we take a deeper look into intra-sentence dependencies to
investigate if the syntax information is totally useless, or to what extent it can help improve
the discourse segmentation performance. To achieve this, we propose a sequence-to-sequence model
along with a GCN based encoder to well utilize intra-sentence dependencies and a multi-head
biaffine attention based decoder to predict EDU boundaries. Experimental results on two benchmark
corpora show that the syntax information we use is significantly useful and the resulting model
is competitive when compared with the state-of-the-art.
-- Model Architecture
The figure below illustrates the proposed approach.
Since we use multi-head attention in this work for robustness, we use the following loss objective
to encourage the divergence between each two attention heads.
-- Example Analysis
To study the correlation between the EDU segmentation process and the syntactic information we use,
we give another analysis about the randomly selected examples in the Figure below. In dependency
structure, a fake root is usually added and only one word is the dependent of the root, which we
refer to as the root-dep unit (e.g., the word "have" in Figure (a)). Intuitively, we draw partial
dependency structure between EDU boundaries and root-dep units for the two examples respectively.
And the partial dependency structures in both examples reveal an interesting language phenomenon
that those words identifying EDU boundaries are direct dependents of root-dep units. Scrupulously,
we further display the proportion of EDU boundaries related to root-dep units in Table 6, and the
results show that this language phenomenon is common in both corpora. Under the conduction of
explicit dependency structures, those text units serving as dependents of root-dep units are well
equipped with "hints" for EDU boundary determination. Hence, we have reason to believe that the
refining method we use is stable and useful for RST-style discourse segmentation for languages
like English and Chinese.
-- Required Packages
Coming soon
-- Usage
Coming soon.
-- Reference
Please read the following paper for more technical details
Zhang L., Kong F., Zhou G. (2020) Syntax-Guided Sequence to Sequence Modeling for Discourse Segmentation. In: Zhu X., Zhang M., Hong Y., He R. (eds) Natural Language Processing and Chinese Computing. NLPCC 2020. Lecture Notes in Computer Science, vol 12431. Springer, Cham. https://doi.org/10.1007/978-3-030-60457-8_8
-- Developer
Longyin Zhang
Natural Language Processing Lab, School of Computer Science and Technology, Soochow University, China
mail to: zzlynx@outlook.com, lyzhang9@stu.suda.edu.cn
-- License
Copyright (c) 2019, Soochow University NLP research group. All rights reserved.
Redistribution and use in source and binary forms, with or without modification, are permitted provided that
the following conditions are met:
1. Redistributions of source code must retain the above copyright notice, this list of conditions and the
following disclaimer.
2. Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the
following disclaimer in the documentation and/or other materials provided with the distribution.