💻 computer science

Semantically Enriching Investor Micro-blogs for Opinion-Aware Emotion Analysis: A Practical Approach

This paper proposes a practical approach to enhance investor micro-blog analysis by augmenting the StockEmotions dataset with semantically structured opinion graphs derived from StockTwits, demonstrating that integrating these opinion semantics via Graph Neural Networks significantly improves emotion classification performance.

Original authors: Gaurav Negi, Paul Buitelaar

Published 2026-05-06

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Gaurav Negi, Paul Buitelaar

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 trying to understand why a crowd of investors is shouting on a social media platform called StockTwits.

The Old Way: Just Listening to the Volume
Traditionally, computers analyzing these posts only looked at the "volume" of the voice. They asked a simple question: Is this person happy (positive) or sad (negative)?

The Problem: This is like hearing someone scream "I hate this!" and only knowing they are angry. You don't know what they hate, why they hate it, or if they are actually talking about a specific stock or just the weather. It's too blunt an instrument for the complex world of finance.

The New Approach: Adding a "Why" and a "Who"
This paper introduces a smarter way to listen. The authors took a dataset of 10,000 investor comments and used a powerful AI (a Large Language Model) to act like a super-detective. This detective didn't just label the emotion; it broke every sentence down into a detailed map of opinions.

Think of it like upgrading from a simple "Happy/Sad" sticker to a full diagram that includes:

The Target: What are they talking about? (e.g., "Tesla")
The Aspect: What specific part? (e.g., "The battery life")
The Holder: Who is saying it? (e.g., "I")
The Qualifier: Under what conditions? (e.g., "When doing heavy computing")
The Sentiment: How strong is the feeling? (e.g., "Strongly negative")

The Magic Tool: The "Opinion Graph"
Once the AI created these detailed diagrams for every sentence, the researchers built a special machine to read them. They used something called a Graph Neural Network (GNN).

The Analogy: Imagine the text of a tweet is a straight line of words. The new "Opinion Graph" turns that line into a spiderweb.
- The words are the nodes (dots).
- The relationships (like "I" holding an opinion about "Tesla") are the threads connecting them.
- The "Sentiment" is the color of the thread.

By feeding this spiderweb into their computer model, the machine doesn't just read the words; it understands the structure of the thought. It sees how the "Holder" connects to the "Target" and how the "Qualifier" changes the meaning.

What Happened When They Tried It?
The researchers tested this new "Spiderweb" method against standard models and even against giant, famous AI models (like GPT-5 and Qwen).

The Small, Specialized Models Won: The models enhanced with the "Opinion Graph" (specifically the BERT-GNN and RoBERTa-GNN) got much better at guessing the specific emotions (like "Anxiety," "Excitement," or "Confusion") than the giant AI models did.
The Giants Struggled: The huge, general-purpose AI models were surprisingly bad at spotting tricky emotions like "Ambiguity" or "Disgust." They often got them wrong, scoring very low.
The "Why" Matters: By understanding the structure of the opinion (the graph), the model could distinguish between a confused investor and an angry one much more accurately. For example, it improved the detection of "Anxiety" and "Excitement" significantly.

The Bottom Line
The paper claims that if you want to understand investor emotions, you can't just look at the words; you have to map out the relationships between the words. By turning text into a structured "opinion map" and feeding it into a model designed to read maps, they made the computer much better at understanding the nuanced, messy reality of human emotion in finance.

In short: They taught the computer to stop just hearing the noise and start reading the blueprint of the thought.

Technical Summary: Semantically Enriching Investor Micro-blogs for Opinion-Aware Emotion Analysis

Problem Statement

While sentiment analysis is a staple of financial Natural Language Processing (NLP), capturing the nuances of why a sentiment exists remains a significant challenge. Existing approaches often analyze investor emotions alongside sentiment but lack the granularity to identify the specific targets of these emotions. Current literature has transitioned from binary sentiment to richer emotional categories (e.g., Ekman's or Plutchik's classifications), yet a gap persists in integrating these emotional categories with fine-grained opinions that include specific targets and deep semantic context. This paper addresses the need to move beyond simple emotional categorization to a representation that integrates opinion semantics, specifically targeting the interaction between investor emotional states and the structured semantics of opinions.

Methodology

The proposed approach involves a three-stage pipeline: dataset augmentation, graph-based semantic encoding, and feature fusion for classification.

1. Dataset Augmentation via LLM

The authors augment the StockEmotions dataset (collected from StockTwits) with fine-grained opinion representations based on the Unified Opinion Concepts (UOC) ontology.

Process: A declarative Large Language Model (LLM) pipeline, specifically utilizing Qwen3.5-35B and the Multi-prompt Instruction Proposal Optimiser (MIPRO), generates opinion annotations for 10,000 comments. This approach avoids naive prompting by automatically optimizing instructions and in-context examples.
Annotation Schema: The augmentation extracts five key components for each opinion:
- Sentiment: Polarity (positive/negative/neutral), intensity, and expression.
- Target: The entity being discussed (e.g., "Battery").
- Aspect: The specific attribute of the target (e.g., "battery life").
- Holder: The individual or organization expressing the opinion.
- Qualifier: The group or context to which the opinion pertains.

2. Graph Neural Network (GNN) Architecture

The core innovation lies in integrating these opinion semantics into a classification model using Graph Neural Networks.

Graph Construction: For each text sequence, the system constructs opinion sub-graphs where nodes represent the Holder, Target, Aspect, Qualifier, and Sentiment. Edges encode the relationships and sentiment polarity between these nodes.
Message Passing: The model employs a Graph Attention Network (GATv2) to aggregate information from neighborhood nodes. This allows the model to learn parametric semantic representations of the opinion structure.
Feature Fusion: The graph-level representations are mapped back to the sequence level and fused with the baseline transformer embeddings (from BERT or RoBERTa). The authors experiment with three fusion mechanisms:
1. Concatenation: Simple vector concatenation.
2. Gated Fusion: A mechanism allowing the model to dynamically weight the importance of text vs. opinion semantics.
3. Attention Fusion: A dot-product attention mechanism modeling interactivity between feature spaces.
Residual Connection: A residual stage preserves the original textual context by adding the baseline sequence features back to the fused representation.

Key Contributions

Dataset Augmentation: The creation of an augmented StockEmotions dataset enriched with detailed, fine-grained opinion annotations derived from the UOC ontology using an LLM-driven pipeline.
Architectural Integration: The proposal of a novel method to integrate UOC opinion semantics into pre-trained language model classifiers using Graph Neural Networks.
Empirical Evaluation: A comprehensive evaluation of state-of-the-art LLMs (GPT-5, Qwen-3.5) against fine-tuned encoder models (BERT, RoBERTa) enhanced with GNNs for emotion classification in financial texts.

Results

The study evaluates performance across 12 distinct emotion categories and aggregated valence groups (Positive, Negative, Ambiguous).

Performance Gains: The integration of GNNs significantly improved the performance of the BERT baseline. The BERT-GNN variant achieved an F1-macro score of 37.88, an improvement of over 4.5 points compared to the standard BERT baseline (33.25), with statistical significance ( $p \approx 0.001$ ).
Category Specifics:
- BERT-GNN showed notable improvements in "Anxiety" (50.19), "Excitement" (48.25), and "Ambiguity" (23.12).
- RoBERTa-GNN achieved the highest scores for "Depression" (41.86) and "Disgust" (45.36), though the overall macro-F1 increase was marginal (37.57 to 38.01) and not statistically significant via McNemar's test, though label assignment changes were significant.
- Baseline Encoders vs. Generative LLMs: Fine-tuned encoder models (BERT/RoBERTa) significantly outperformed large generative models (GPT-5, GPT-5-MINI, Qwen-3.5-35B) in a zero-shot setting. The generative models struggled particularly with "Ambiguity," "Belief," and "Disgust," often dropping into single-digit F1 scores.
Valence Analysis: The GNN-enhanced models demonstrated superior robustness in classifying "Ambiguous" and "Negative" emotions compared to generative LLMs, which performed well on "Positive" emotions but failed on the others.

Significance and Claims

The paper claims that incorporating semantically rich opinion representations significantly enhances the accuracy of predicting investor emotional states, particularly for complex or ambiguous emotions. The authors assert that:

Opinion Semantics Matter: The "why" behind sentiment (captured via UOC) provides necessary granularity that improves classification performance across different emotional spectrums.
Task-Specific Encoders Outperform General LLMs: For specific financial emotion classification tasks, fine-tuned encoder models augmented with graph-based semantic reasoning outperform larger, general-purpose generative LLMs, especially in detecting negative and ambiguous sentiments.
Practical Utility: The proposed declarative LLM pipeline offers a practical approach to enriching existing datasets without requiring parametric learning for the annotation process itself, bridging the gap between raw text and structured opinion analysis.

The work concludes that while large language models show promise in zero-shot scenarios for positive sentiment, they currently lack the precision required for nuanced financial emotion analysis, where opinion-aware, graph-enhanced architectures provide a more effective solution.