💬 NLP

Improving Neural Argumentative Stance Classification in Controversial Topics with Emotion-Lexicon Features

This paper proposes a neural stance classification model enhanced by an expanded, contextually-aware emotion lexicon (eNRC) derived from DistilBERT embeddings, which significantly improves performance across five diverse controversial topic datasets compared to baseline, original lexicon, and LLM-based approaches.

Original authors: Mohammad Yeghaneh Abkenar, Weixing Wang, Manfred Stede, Davide Picca, Mark A. Finlayson, Panagiotis Ioannidis

Published 2026-02-27

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Mohammad Yeghaneh Abkenar, Weixing Wang, Manfred Stede, Davide Picca, Mark A. Finlayson, Panagiotis Ioannidis

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are walking through a bustling town square where people are shouting about controversial topics like "Should we ban plastic bags?" or "Is climate change real?"

Some people are shouting "FOR" (supporting the idea), and others are shouting "AGAINST" (opposing it). Your job is to stand in the middle, listen to a specific sentence, and instantly know which side the speaker is on. This is what computers call Stance Classification.

However, humans don't just use logic to argue; we use emotions. We get angry, hopeful, scared, or disgusted. A computer that only looks at the logical words often gets confused. This paper is about teaching a computer to "feel" the argument so it can guess the stance much better.

Here is the story of how the authors did it, using some simple analogies.

1. The Problem: The Computer's "Emotion Dictionary" is Outdated

Imagine the computer has a dictionary of emotional words (like "angry," "joyful," "fearful"). This is called the NRC Emotion Lexicon.

But here's the catch: This dictionary was written a while ago. It's like having a dictionary of slang from the 1990s. If a modern person says, "This policy is a nightmare," the old dictionary might not know that "nightmare" is a strong negative emotion. It might think it's just a word about sleeping.

Because the dictionary is missing these modern, charged words, the computer misses the emotional cues that reveal whether someone is for or against a topic.

2. The Solution: The "Smart Expander" (eNRC)

The authors decided to fix this dictionary. They didn't just ask humans to add more words (which takes forever); they built a Smart Expander using a tool called DistilBERT (a type of AI that understands how words relate to each other).

Think of the original dictionary as a small, well-organized garden with specific flowers (emotions).

The Old Way: You walk around and manually pick a few more flowers to add.
The New Way (eNRC): You use a robot gardener. The robot looks at every flower in the garden and asks, "What other plants in the world look or smell like this one?"
- If the garden has "anger," the robot finds "fury," "rage," and even modern slang like "pissed off" or "nightmare" because they live in the same "neighborhood" of meaning.
- The robot then adds these new words to the garden, but it's very careful. It uses a clustering method (grouping similar plants together) to make sure it doesn't accidentally put a "happy" flower in the "sad" section.

The result is a Super-Dictionary (eNRC) that is much bigger and understands modern language much better.

3. The Experiment: Testing the New Dictionary

The authors took this Super-Dictionary and taught a computer model to use it. They tested it on five different "town squares" (datasets) ranging from short social media comments to long student essays.

They compared three teams:

The No-Dictionary Team: The computer tried to guess the stance without any emotion help.
The Old-Dictionary Team: The computer used the original, outdated dictionary.
The Super-Dictionary Team (eNRC): The computer used the new, expanded dictionary.

The Results:

The Super-Dictionary Team won almost every time.
On the hardest test (student essays), they improved their score by 6.2% compared to the team with no help.
Even a very smart, massive AI (a Large Language Model like Qwen) struggled to guess the stance correctly on its own, but the authors' method, which combined a standard AI with their emotion dictionary, performed better.

4. Why This Matters

Think of it like a detective solving a crime.

Logic alone is like looking at the suspect's alibi. It's important, but people lie.
Emotion is like noticing the suspect's shaking hands or sweating. It often reveals the truth even when the words are tricky.

By giving the computer a better "emotion radar," the authors made it much better at understanding controversial debates. This is crucial for:

Stopping Misinformation: Detecting when bots are trying to manipulate public opinion by using fake emotional outrage.
Understanding Public Opinion: Seeing what people really feel about climate change or politics, not just what they say they think.

The Takeaway

The authors didn't just build a better computer; they built a better translator for human feelings. They showed that if you want a computer to understand a heated argument, you can't just teach it logic; you have to teach it to understand the emotional color of the words being used. And by using AI to expand their emotional dictionary, they made that translator much sharper.

1. Problem Statement

Argumentative Stance Classification involves identifying whether an author's text supports or opposes a specific controversial topic (e.g., "waste separation" or "death penalty"). While argumentation mining has advanced, existing approaches face three critical limitations:

Neglect of Emotion: Arguments on controversial topics often rely heavily on emotional appeals (Pathos). However, most prior work focuses on propositional content (Logos) and fails to systematically incorporate fine-grained emotion analysis.
Lexicon Limitations: Standard emotion lexicons (like NRC) are manually curated and often miss modern, context-specific, or semantically related emotional terms, leading to poor coverage in diverse domains.
Generalizability Issues: Previous studies often focus on single domains, non-argumentative texts, or specific datasets, limiting the applicability of models to broader, real-world argumentative contexts.

The paper aims to address these gaps by systematically integrating an expanded, context-aware emotion lexicon into a neural stance classification framework to improve performance across diverse controversial topics.

2. Methodology

The proposed approach consists of two main components: the development of an Expanded NRC Lexicon (eNRC) and a Neural Argumentative Stance Classification (NASCA) model.

A. Development of eNRC (Lexicon Expansion)

The authors propose a method to expand the existing Bias-Corrected NRC Emotion Lexicon using contextualized embeddings to capture semantically similar emotional terms.

Embedding Generation: Words from the original NRC lexicon are embedded into a 768-dimensional space using DistilBERT.
Clustering & Dimensionality Reduction:
- Principal Component Analysis (PCA) reduces embeddings to 3 dimensions to capture dominant variance.
- A Gaussian Mixture Model (GMM) with three components clusters the words. This reveals a trimodal distribution, identifying distinct semantic regions within the lexicon.
Cluster-Based Normalization: To handle varying densities across clusters, similarity scores between candidate words and lexicon words are normalized within their respective clusters. This prevents high-density clusters from dominating the expansion process.
- Formula: $s_{norm} = \text{clip}(\frac{s - \mu_i}{\sigma_i} + \frac{\sigma_c}{2\sigma_c}, 0, 1)$ , where $\mu_i$ and $\sigma_i$ are cluster statistics.
Threshold-Based Expansion:
- Candidate words are assigned emotion labels if their normalized similarity to a lexicon word exceeds a threshold $\theta$ .
- The authors identified $\theta = 0.4$ as the optimal balance between coverage and semantic precision, avoiding "semantic drift" (assigning emotions to unrelated words) while maximizing vocabulary.

B. NASCA Model Architecture

The classification model is an end-to-end neural framework designed for argumentative text.

Input Construction: The input is formatted as [CLS] Topic: t [SEP] Argument: x [SEP], allowing the model to jointly encode the target topic and the argumentative segment.
Encoder: A pre-trained BERT encoder generates contextualized representations. The [CLS] token representation ( $h_{CLS}$ ) serves as the aggregate sentence embedding.
Feature Fusion:
- The model extracts emotion features ( $f_{emo}$ ) from the text using the eNRC (or baseline NRC/noNRC).
- These features are concatenated with the BERT [CLS] vector: $h_{combined} = [h_{CLS}; f_{emo}]$ .
Classifier: The combined vector passes through two hidden layers (ReLU activation, 10% dropout) and a final linear layer for binary classification (For/Against).

3. Key Contributions

eNRC Lexicon: The creation of an expanded, context-aware emotion lexicon using DistilBERT embeddings and cluster-based normalization. This systematically identifies emotionally charged terms missed by manual curation.
Novel Framework: A neural stance classification framework that explicitly integrates emotion lexicon features as an adapter to BERT, specifically tailored for argumentative microtexts and essays.
Comprehensive Evaluation: The approach is evaluated on five diverse corpora (AMT1, AMT2, Persuasive Essays, IBM Debater, UKP) covering various domains (web discourse, student essays, parliamentary debates) and topics, addressing the generalizability gap in prior work.
Resource Release: Full reproducibility is ensured by releasing the eNRC, adapted corpora, and model code.

4. Results

The model was evaluated using Macro F1 scores across the five datasets, comparing three variants: noNRC (baseline), NRC (original lexicon), and eNRC (expanded lexicon).

Performance Gains:
- eNRC outperformed the noNRC baseline across all datasets, with the most significant improvement on the Persuasive Essays (PE) dataset (+6.2% F1, reaching 62.9%).
- eNRC outperformed the original NRC on four of the five datasets, with gains up to +3.0% (e.g., +2.6% on PE).
- On the UKP dataset, eNRC achieved 68.4% F1, the highest absolute performance.
Comparison with SOTA:
- The eNRC model surpassed previous State-of-the-Art (SOTA) results on the IBM (65.9% vs. 64.5%) and UKP (68.4% vs. 63.2%) datasets.
- It significantly outperformed a large language model (Qwen2.5-7B) on most corpora, highlighting that specialized feature engineering (emotion lexicons) remains crucial for argumentative tasks even in the era of LLMs.
Ablation Studies:
- Threshold Sensitivity: Performance peaked at $\theta = 0.4$ . Lower thresholds introduced noise; higher thresholds limited coverage.
- BERT Fine-tuning: The eNRC model showed robust performance even with frozen BERT weights, suggesting the emotion features provide sufficient discriminative power to compensate for lack of encoder fine-tuning. This reduces computational costs for deployment.

5. Significance

Bridging Emotion and Logic: The study empirically validates that emotional framing is a critical, often overlooked signal in argumentative stance detection, particularly for controversial topics where logic and emotion are intertwined.
Scalability and Efficiency: By demonstrating that high-quality emotion features can stabilize performance without full model fine-tuning, the paper offers a computationally efficient solution for stance classification in resource-constrained environments.
Generalizability: The success across five distinct datasets proves that the method is not domain-specific, offering a robust template for future argumentation mining research.
Practical Application: The improved ability to detect stances in polarized media ecosystems aids in identifying misinformation campaigns, understanding public discourse, and distinguishing genuine sentiment from amplified extremes.