Sense classification of shallow discourse relations with focused RNNs

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Artificial Intelligence

Background:

• Understanding discourse relations is crucial for text comprehension.
• Existing automated methods require external resources and complex pipelines.
• A need exists for a unified, resource-independent approach to discourse relation classification.

Purpose of the Study:

• To introduce a novel system for classifying shallow discourse relation senses.
• To develop a single, end-to-end trainable model for diverse discourse relations.
• To create a language-independent system applicable at word and character levels.

Main Methods:

• Utilized focused recurrent neural networks (RNNs) for classification.
• Introduced a novel multi-dimensional RNN-attention mechanism for text embeddings.
• Developed a filtering/gating RNN for focused attention on input sequences.
• Generated argument embeddings for sense classification.

Main Results:

• The FR system achieves competitive performance on English datasets.
• Outperformed existing best systems by 2.5% on the Chinese blind dataset.
• Demonstrated language independence and applicability at granular levels.

Conclusions:

• The FR system offers a unified and resource-efficient approach to discourse relation sense classification.
• The novel RNN-attention mechanism effectively generates informative text embeddings.
• The system shows strong potential for cross-lingual and fine-grained text analysis.