Abstract:Political perspective detection has become an increasingly important task that can help combat echo chambers and political polarization. Previous approaches generally focus on leveraging textual content to identify stances, while they fail to reason with background knowledge or leverage the rich semantic and syntactic textual labels in news articles. In light of these limitations, we propose KCD, a political perspective detection approach to enable multi-hop knowledge reasoning and incorporate textual cues as paragraph-level labels. Specifically, we firstly generate random walks on external knowledge graphs and infuse them with news text representations. We then construct a heterogeneous information network to jointly model news content as well as semantic, syntactic and entity cues in news articles. Finally, we adopt relational graph neural networks for graph-level representation learning and conduct political perspective detection. Extensive experiments demonstrate that our approach outperforms state-of-the-art methods on two benchmark datasets. We further examine the effect of knowledge walks and textual cues and how they contribute to our approach's data efficiency.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is **the insufficient background knowledge reasoning and under - utilization of text cues in Political Perspective Detection**. Specifically, existing methods mainly rely on text content to identify political stances, but fail to fully utilize external knowledge graphs for multi - hop reasoning, and also fail to effectively utilize the rich semantic and syntactic labels in news articles.
### Main Problems
1. **Lack of Background Knowledge Reasoning**: Existing methods fail to combine external knowledge graphs for multi - hop reasoning, and thus cannot fully utilize background knowledge.
2. **Insufficient Utilization of Text Cues**: Existing methods fail to fully mine semantic, syntactic, and entity cues in news articles, which are very important for identifying political stances.
### Solutions
To solve the above problems, the author proposes a framework named **KCD (Knowledge Walks and Textual Cues Enhanced Political Perspective Detection)**. The main contributions of KCD include:
1. **Introducing Multi - Hop Knowledge Reasoning**: By generating random walks (knowledge walks) on external knowledge graphs and combining these walks with news text representations, multi - hop knowledge reasoning is achieved.
2. **Constructing a Heterogeneous Information Network**: Construct a heterogeneous information network (HIN) to jointly model knowledge - enhanced news content as well as semantic, syntactic, and entity cues in news articles.
3. **Graph Neural Network Learning**: Adopt Relational Graph Neural Networks (R - GCN) for graph - level representation learning, and finally perform political perspective detection.
### Experimental Results
The experimental results show that KCD outperforms the existing state - of - the - art methods on two benchmark datasets (SemEval and Allsides). In particular, when using the Paragraph Averaging (PA) aggregation strategy, the performance of KCD is particularly prominent. In addition, removing text cues or knowledge walks will lead to a significant decline in performance, further proving the effectiveness of these components.
### Formula Summary
- Knowledge Walk Generation Formula:
\[
kwi=\{e^{(0)}, r_0, e^{(1)},..., r_{K - 1}, e^{(K)}\}
\]
where \(e^{(i - 1)}\) and \(r_{i - 1,i}\) represent the head entity and the relationship of the \(i\) - th triple respectively.
- Conditional Probability Formula:
\[
P(e^{(i)}|e^{(i - 1)}, r_{i - 1,i})=\frac{\exp(p(r_{i - 1,i}))}{\sum_{j = 1}^{|N_r(i - 1)|}\exp(p(r_j))}
\]
- Attention Mechanism Formula:
\[
v_p^i=\sum_{j = 1}^m\frac{\exp(\alpha\cdot v_k^{i,j})}{\sum_{q = 1}^m\exp(\alpha\cdot v_k^{i,q})}v_k^{i,j}
\]
where \(\alpha=\phi(W_a v_s^i + b_a)\), and \(\phi\) is the Leaky - ReLU function.
Through these improvements, KCD can more accurately identify political stances in news articles, especially when combining background knowledge and text cues.