📝 Stance Detection Project

This project focuses on stance detection by extracting keywords/targets, performing sentiment analysis, and mapping relationships between targets within articles. It combines statistical, semantic, and embedding-based NLP techniques to build a full pipeline from raw articles → sentiment-aware keyword graphs.

---

🚀 Workflow

1. Keyword Classification

• Input: Article file

• Output: keywords/targets.csv

• Extracts keywords using RAKE, YAKE, and KeyBERT.

• Preprocessing: cleaned sentences → concatenated into a single paragraph for better context.

• Threshold-based keyword selection:

  • RAKE & KeyBERT → keep high-score keywords.
  • YAKE → keep low-score keywords (higher relevance).

• Results compiled into a CSV file.

Feature / Method	RAKE	YAKE	KeyBERT
Type	Statistical (frequency & co-occurrence)	Statistical (multi-feature scoring)	Semantic (embedding-based)
Strengths	Fast, simple	Language-independent, adaptable	Captures meaning & nuance
Limitations	Ignores semantics	Sensitive to weights	Slower, needs embeddings

---

2. Sentence Extraction

• Input: Article file
• Output: Sentences list
• Uses NLTK Punkt tokenizer.
• Handles punctuation and avoids splitting inside abbreviations.
• Cleans leading/trailing whitespace.

---

3. Sentiment Analysis

• Input: Sentences list

• Output: sentences + sentiment.csv

• Based on ELECTRA architecture:

  1. Pre-trained fine-tuned model (jbeno/electra-base-classifier-sentiment)
  2. Fine-tuned Google ELECTRA model (google/electra-base-discriminator)

• Current experiments use the jbeno model applied to the demonetization dataset.

---

4. Finding Relations Between Targets

• Input: Sentences+sentiment file + targets list file
• Output: keywords/targets + sentiments.csv
• Matches keywords in sentences via regex.
• If multiple keywords co-occur in a sentence → relation is formed.
• Relation inherits sentence sentiment.

---

5. Target Generalization

• Input: keywords/targets.csv
• Output: clustered keywords.csv
• Uses SentenceTransformer (all-MiniLM-L6-v2) to generate embeddings.
• Applies hierarchical clustering with cosine distance.
• Threshold parameter controls granularity (strict vs. loose clustering).

---

6. Keyword Generalization Mapping

• Input: Clustered keywords map + targets+sentiment file
• Output: Unified file with keywords → generalized form + sentiment
• Ensures consistent grouping of keywords (e.g., AI → Artificial Intelligence).

---

7. Keyword Sentiment Graph

• Input: Final CSV (generalized keywords + sentiments)

• Output: Sentiment graph

• Graph shows keyword connections with sentiment-aware edges:

  • ✅ Green → Positive
  • ❌ Red → Negative
  • ⚪ Gray → Neutral

• Nodes = keywords, edges = co-occurrence + sentiment context.

---

📊 Example Datasets

• Demonetization dataset (with/without sampling)
• Ball tampering dataset
• Catalan
• Harvey

---

🛠️ Tech Stack

• Python (NLP pipeline)
• NLTK (sentence tokenizer)
• RAKE, YAKE, KeyBERT (keyword extraction)
• Hugging Face Transformers (ELECTRA models)
• SentenceTransformers (embeddings & clustering)
• Matplotlib / NetworkX (graphs & visualization)

---

📌 Future Improvements

• Fine-tune ELECTRA on stance datasets
• Experiment with contextual target detection (beyond regex)
• Improve clustering thresholds dynamically
• Interactive graph visualization

---

👩‍💻 Contributors

• Shiny Chakma
• Priya Keshri
• Toko Saniya