[Paper Review] Coherent Comment Generation for Chinese Articles with a Graph-to-Sequence Model

Wei Li, Jingjing Xu+4 more|arXiv (Cornell University)|Jun 4, 2019

Topic Modeling29 references9 citations

TL;DR

This paper proposes a graph-to-sequence model that constructs a topic interaction graph from Chinese news articles to improve coherent and informative comment generation. By modeling topic clusters and their semantic relations, the model outperforms strong baselines in fluency, coherence, and informativeness on a newly collected 200,000-sample Chinese news-comment corpus.

ABSTRACT

Automatic article commenting is helpful in encouraging user engagement and interaction on online news platforms. However, the news documents are usually too long for traditional encoder-decoder based models, which often results in general and irrelevant comments. In this paper, we propose to generate comments with a graph-to-sequence model that models the input news as a topic interaction graph. By organizing the article into graph structure, our model can better understand the internal structure of the article and the connection between topics, which makes it better able to understand the story. We collect and release a large scale news-comment corpus from a popular Chinese online news platform Tencent Kuaibao. Extensive experiment results show that our model can generate much more coherent and informative comments compared with several strong baseline models.

Motivation & Objective

• Address the challenge of generating coherent and informative comments for long-form Chinese news articles, which traditional sequence-to-sequence models struggle with due to length and semantic complexity.
• Overcome the limitations of title-only or short-text inputs that fail to convey sufficient context for meaningful comment generation.
• Improve comment quality by modeling the internal topic structure and inter-topic relationships in news articles through a graph-based representation.
• Develop a joint modeling approach that incorporates both the title and content into a unified graph structure to better capture the main narrative point.
• Release a large-scale, real-world Chinese news-comment dataset (200,000 samples) to support future research in automatic comment generation.

Proposed method

• Construct a topic interaction graph by identifying topic-centered sentence clusters from the article content and title, where each cluster forms a vertex.
• Define edges between vertices based on semantic similarity between topic clusters, using pre-trained sentence embeddings to compute pairwise relations.
• Integrate the title as a special vertex in the graph to anchor the main theme and guide comment generation toward the central narrative.
• Employ a graph-to-sequence decoder that attends over the graph structure, allowing the model to attend to relevant topic clusters during comment generation.
• Use a pointer-generator network with copy mechanism to preserve named entities and key terms from the input, improving factual consistency.
• Train the model end-to-end using cross-entropy loss on the generated comment sequences, with beam search for inference.

Experimental results

Research questions

• RQ1 Can a graph-based representation of news articles better capture the internal topic structure and inter-topic relationships than sequential or hierarchical models?
• RQ2 Does integrating the title as a dedicated vertex in the topic interaction graph improve the model’s ability to identify the main narrative point?
• RQ3 Can a graph-to-sequence model generate comments that are more coherent and informative than sequence-to-sequence or hierarchical attention models?
• RQ4 How does the model perform in generating comments that are specific to key characters or plot points, rather than generic or irrelevant remarks?
• RQ5 To what extent does the graph structure help the model avoid generating comments that contradict world knowledge or misidentify entities?

Key findings

• The proposed graph2seq model generates comments with significantly higher informativeness, as measured by unique word count and reduced general comment rate, compared to all baselines.
• The model produces the most coherent and topic-specific comments, with 85% of generated comments mentioning key characters or plot points, compared to 50% for the best baseline.
• The model achieves a 12% improvement in human-rated coherence and a 20% reduction in general comments compared to the strongest baseline (self-attention model).
• Case studies confirm that only the graph2seq model correctly identifies and references the main character (Wang Likun) in the news, while others fail to do so.
• Fluency scores are comparable to baselines, indicating that the graph structure does not compromise grammatical quality, despite increased complexity.
• The model demonstrates robustness to input disorder, as it can still generate meaningful comments even when the article content is not in a logical sequence.