Abstract
Stance detection aims to assign a stance label (for or against) to a post toward a specific target. Recently, there is a growing interest in using neural models to detect stance of documents. Most of these works model the sequence of words to learn document representation. However, much linguistic information, such as polarity and arguments of the document, is correlated with the stance of the document, and can inspire us to explore the stance. Hence, we present a neural model to fully employ various linguistic information to construct the document representation. In addition, since the influences of different linguistic information are different, we propose a hierarchical attention network to weigh the importance of various linguistic information, and learn the mutual attention between the document and the linguistic information. The experimental results on two datasets demonstrate the effectiveness of the proposed hierarchical attention neural model.
Introduction
Stance detection is the task of determining whether the author of a text is "for" or "against" a specific target, such as "Gay Rights" or "Climate Change." While it shares similarities with sentiment analysis, stance detection is inherently more complex. An author might use negative sentiment to express a positive stance—for instance, "I hate that people oppose equality" conveys a positive stance toward equality despite the negative tone.
Existing neural models often treat documents as simple sequences of words, overlooking rich linguistic cues that human readers naturally utilize. These cues include:
- Sentiment: The overall tone of the text.
- Dependency Relations: Grammatical links between words (e.g., "adopt" and "children").
- Arguments: The specific logical points or claims made by the author.
The challenge lies in the fact that the relevance of these cues varies across different posts. In some cases, a single strong argument defines the stance, while in others, the cumulative sentiment is more indicative.
Proposed Method: Hierarchical Attention Network (HAN)
To address these challenges, we propose a Hierarchical Attention Network (HAN) that explicitly incorporates and dynamically weights three types of linguistic information: sentiment, dependency relations, and core arguments.
Linguistic Feature Extraction
The model first extracts specialized representations for each linguistic feature:
- Sentiment: Sequences of sentimental words are extracted using a known lexicon.
- Dependency: A parser identifies word pairs with grammatical relationships (e.g., subject-verb-object).
- Argument: An SVM classifier identifies argumentative sentences within the post. Each feature set is processed through its own Long Short-Term Memory (LSTM) network to create a specialized representation.
Hierarchical Attention Mechanism
The core of the HAN is its two-layered attention architecture:
- Linguistic Attention (Level 1): This layer learns which words in the document are most relevant to each specific linguistic feature.
- Hyper Attention (Level 2): This "manager" layer evaluates the document representation alongside the three linguistic representations ($q_1, q_2, q_3$) to determine their relative importance. For example, in a logically dense post, the Hyper Attention layer might assign higher weight to the "Argument" representation.
Experimental Evaluation
The HAN model was evaluated on two distinct benchmarks:
- H&N14: Online debate forum posts (average 114 words) covering targets like Abortion and Gay Rights.
- SemEval16: English tweets (average 18 words) covering targets like Climate Change and Feminism, testing the model's robustness on noisy, brief text.
Results
The HAN model consistently outperformed strong baselines, including standard LSTMs and the Target Attention Network (TAN).
- H&N14: HAN achieved a Macro F1-score of 62.18, surpassing TAN (60.49).
- SemEval16: HAN reached a Macro F1-score of 61.00, significantly higher than the standard LSTM (52.03) and TAN (59.56).
Ablation studies confirmed that the Argument feature set provided the most significant performance boost, suggesting that stance is driven more by content than emotion. Furthermore, removing the "Hyper Attention" layer led to a performance drop across all categories, validating the hierarchical structure's effectiveness.
Conclusion
The Hierarchical Attention Network demonstrates that integrating diverse linguistic information through a dynamic weighting mechanism significantly improves stance detection. While the model currently relies on external tools for feature extraction, which may introduce error propagation, its ability to capture nuanced relationships—such as dependency pairs—sets a new standard. Future work will focus on end-to-end argument extraction and modeling relationships between different targets.
Summary updated