Introduction
Discourse parsing is the process of identifying structural and semantic relationships between text segments. While identifying explicit relationships—those signaled by connectives like "but" or "because"—is a well-understood task, implicit sense classification remains a significant challenge. In these cases, the relationship must be inferred entirely from the context of the arguments ($Arg1$ and $Arg2$). Researchers at Dublin City University (DCU) developed a parser designed to tackle this problem using a novel approach centered on relational phrase embeddings.
Methodology
The DCU team framed sense classification as a supervised learning task, employing off-the-shelf classifiers such as libSVM. The system utilizes four primary feature types to represent the relationship between text segments:
- Relational Phrase Embeddings: This core innovation uses Paragraph Vectors (Doc2Vec) to create distributed representations of each argument. Inspired by knowledge graph models, the team treated the discourse relation as a "translational jump" in vector space, where $Arg1 + Sense \approx Arg2$.
- Production Features: These features are extracted from constituent parse trees to identify syntactic patterns that might signal specific discourse senses.
- Word-Pair Features: This involves the Cartesian product of all words in both arguments, identifying specific word associations (e.g., "rain" and "wet") that hint at a causal or temporal relationship.
- Heuristic Features: Applied primarily to explicit connectives, these include Part-of-Speech (POS) tags and the immediate lexical context surrounding the connective.
The application of these features varies depending on whether the connective is explicit or implicit, as summarized in the table below:
Experimental Setup
The model was trained and evaluated using the CoNLL 2015 Shared Task Data (LDC2015E21). To enhance the robustness of the embeddings, the researchers incorporated external data from a Wikipedia dump and the WMT14 dataset. Google’s Skip-gram (word2vec) models served as the foundational word embeddings.
Key Insights and Results
A central finding of the research is the "Polarity Problem." Because Paragraph Vectors tend to average word meanings, they often fail to capture the contrast required for "Comparison" or "Contingency" senses. If $Arg1$ is positive and $Arg2$ is negative, the averaged vector representation may obscure the "but" relationship.
The performance gap between explicit and implicit tasks remains stark:
- Explicit Connectives: The system achieved a high F1-score of 0.873 for sense classification.
- Implicit Connectives: The F1-score dropped significantly to 0.112.
While the model demonstrated high precision (often exceeding 0.80) on implicit tasks, its recall was very low. This suggests the model is conservative, only assigning labels when it has high confidence, but missing many valid relations. Furthermore, while adding 30 million sentence pairs improved performance for "Expansion" senses, it did not provide a universal performance boost, indicating that data volume cannot fully compensate for architectural limitations in handling semantic nuances.
Conclusion and Future Work
The DCU Discourse Parser highlights the potential of translational embeddings but also exposes the limitations of current vector-averaging techniques. Future improvements may involve moving toward RNN-LSTM or Convolutional Sentence Models to better preserve the tree structure and internal logic of arguments. Addressing the polarity problem and improving argument identification—the process of determining where $Arg1$ and $Arg2$ begin and end—remain critical hurdles for the field.
Summary updated