🧠 Structured Sentiment Analysis

IIIT Delhi · Natural Language Processing Project

A structured approach to opinion mining — extracting who expresses what sentiment toward whom, with a polarity label.

---

📋 Table of Contents

• Problem Statement
• Architecture
• Methodology
  • 1. Initial Processing & Tokenization
  • 2. Tag Classification Layer (BIO Tagging)
  • 3. Conditional Random Field (CRF)
  • 4. Polarity Classification
  • 5. Structured Opinion Generation
• Baseline System
• Features
• Datasets
• Results
• Installation
• Usage
• Project Structure
• Further Work

---

🎯 Problem Statement

Structured Sentiment Analysis (SSA) extends traditional sentiment analysis by extracting complete sentiment graphs from raw text. Instead of simply classifying a text as positive or negative, SSA identifies all the structured opinion tuples present in a sentence.

Formal Definition

Given a text $T = O_1, O_2, \ldots, O_n$, the goal is to predict all opinion tuples of the form:

$$\mathcal{G} = {(h, b, t, p)}$$

Where each element is defined as:

Symbol	Role	Description
$h$	Holder	The entity who expresses the opinion
$b$	Expression	The polar expression (words conveying sentiment)
$t$	Target	The entity toward which the opinion is directed
$p$	Polarity	The sentiment label: positive, negative, or neutral

Illustrative Example

Sentence: "Even though the price is decent for Paris, I would not recommend this hotel."

Opinion Tuple 1:
  Holder     → "I"
  Expression → "would not recommend"
  Target     → "this hotel"
  Polarity   → NEGATIVE

Opinion Tuple 2:
  Holder     → (implicit)
  Expression → "decent"
  Target     → "the price"
  Polarity   → POSITIVE

Unlike standard sentiment analysis, SSA captures complex, overlapping, and multi-target opinions within a single sentence — making it far more expressive and practically useful.

---

🏗️ Architecture

Our model follows a unified end-to-end architecture built on top of a pretrained BERT-based encoder, enhanced with a CRF decoding layer and a separate polarity head.

                    ┌──────────────────────────────────────────┐
                    │              Input Text                   │
                    └───────────────────┬──────────────────────┘
                                        │
                    ┌───────────────────▼──────────────────────┐
                    │   [CLS]   w₁    w₂   ...   wₙ   [SEP]   │
                    │         Token Embedding Layer             │
                    └───────────────────┬──────────────────────┘
                                        │
                    ┌───────────────────▼──────────────────────┐
                    │        Norwegian BERT Encoder             │
                    │      (NbAiLab / nb-bert-base)             │
                    │   Contextual Hidden State Representations │
                    └──────────┬────────────────────┬──────────┘
                               │                    │
               ┌───────────────▼──────┐   ┌─────────▼────────────┐
               │   CRF for BIO Tagging│   │  Polarity Head        │
               │                      │   │  ([CLS] token)        │
               │  B-Source  I-Source  │   │                       │
               │  B-Target  I-Target  │   │  ┌──────────────┐    │
               │  B-Polar   I-Polar   │   │  │   Positive   │    │
               │  O                   │   │  │   Negative   │    │
               └──────────┬───────────┘   │  │   Neutral    │    │
                          │               │  └──────────────┘    │
                          │               └─────────┬────────────┘
                          │                         │
               ┌──────────▼─────────────────────────▼──────────┐
               │         Structured Opinion Generation           │
               │                                                 │
               │   Merge CRF spans + polarity predictions       │
               │   Apply offset mapping & span filtering        │
               │                                                 │
               │  Output: (Source, Target, Expression, Polarity)│
               └─────────────────────────────────────────────────┘

Model Components at a Glance

Component	Description	Technology
Encoder	Contextual representation of input tokens	RoBERTa / BERT (multilingual)
BIO Tagger	Sequence labeling for span identification	Linear Projection
CRF Decoder	Structured decoding with label constraints	Conditional Random Field
Polarity Head	Classifies sentiment of identified spans	Multi-layer Classifier on [CLS]
Opinion Builder	Assembles final structured output tuples	Post-processing pipeline

---

🔬 Methodology

1. Initial Processing & Tokenization

• SSA extends basic sentiment analysis by extracting four key components: Holder, Target, Expression, Polarity
• We use a pre-trained RoBERTa model (PolitAi Lab/nb-bert-base, [Liu et al., 2019]) as our encoder
• The encoder transforms raw input text into rich contextual embeddings for each token
• We benchmarked RoBERTa with bert-base and bert-medium — RoBERTa consistently outperforms both

# Encoder produces hidden states for every token position
hidden_states = roberta_encoder(input_ids, attention_mask)
# Shape: [batch_size, seq_len, hidden_dim]

---

2. Tag Classification Layer (BIO Tagging)

• A linear projection layer maps contextual embeddings to BIO tag logits
• We use the BIO (Beginning-Inside-Outside) tagging scheme to represent the boundaries of each opinion component

BIO Tag Space

O           → Outside any opinion component
B-Source    → Beginning of a Holder span
I-Source    → Inside a Holder span
B-Target    → Beginning of a Target span
I-Target    → Inside a Target span
B-Polar     → Beginning of a Polar Expression span
I-Polar     → Inside a Polar Expression span

Projection Formula

$$Z = W \cdot z + b$$

where $W \in \mathbb{R}^{t \times H}$ is the weight matrix, $z^{t,i} \in \mathbb{R}^H$ represents the hidden state at position $i$, and $t$ is the number of possible BIO tags.

$$\hat{Y} = \mathbb{R}^{N \times t}$$

---

3. Conditional Random Field (CRF)

After obtaining token-level hidden states from BERT/RoBERTa, a CRF layer is applied to decode BIO tag sequences in a globally-consistent manner.

• The CRF learns transition weights between adjacent tags
• This enforces structural constraints (e.g., I-Source cannot follow B-Target)
• Enables span detection while explicitly modeling tag dependencies

CRF Training Objective

$$\mathcal{L}_{CRF} = -\log\frac{\exp(\text{score}(\text{gold tags}))}{\sum_{\hat{y}} \exp(\text{score}(\hat{y}, \text{ valid tags}))}$$

Key Insight: CRF outperforms a plain softmax decoder because it captures inter-tag dependencies critical for multi-span structured extraction.

---

4. Polarity Classification

• A separate multi-layer classification head takes the [CLS] token representation as input
• Predicts one of three sentiment labels:

Sentiment Labels:
  ✅ Positive   →  Favorable / approving sentiment
  ❌ Negative   →  Critical / disapproving sentiment
  ➖ Neutral    →  Factual / non-opinionated expression

• The model struggles with ambiguous sentences like:

  "The food was great, but the service was terrible."

  where polarity is mixed at the span level — a key challenge that motivates future work on span-level polarity.

---

5. Structured Opinion Generation

In the final stage, we combine CRF-predicted spans with polarity predictions to produce the complete structured opinion quadruple (holder, expression, target, polarity).

Post-processing pipeline:

1. Convert span boundaries to character offsets using BERT's offset mapping with adjacent token merging
2. Filter short spans (< 3 characters) to remove noise
3. Assemble final output as JSON-formatted opinion tuples

⚠️ This filtering step leads to a small loss (~a few percent) of true predictions.

---

🔁 Baseline System

As a comparison point, we implement a Sequence Labeling + Relation Classification pipeline using BiLSTM models.

Pipeline Overview

Step 1: Span Extraction
        ├── BiLSTM → Extract Holders
        ├── BiLSTM → Extract Targets
        └── BiLSTM → Extract Polar Expressions

Step 2: Relation Prediction
        └── BiLSTM + Max Pooling
              ├── Full text representation
              ├── Holder / Target representation
              └── Expression representation
                    └── Concatenate → Linear → Sigmoid
                          → Binary: has_relation? (threshold = 0.5)

Step 3: Assemble tuples → (holder, target, expression, polarity)

The baseline uses GloVe / FastText embeddings and trains three separate BiLSTMs, one per annotation type, followed by a relation prediction model.

---

✨ Features

Feature	Description
🌍 Multilingual Support	Works across Norwegian, English, Spanish, Catalan, and Basque
🏷️ BIO Sequence Labeling	Precise span-level identification using structured tagging
🔗 CRF Decoding	Globally-consistent tag sequence prediction with structural constraints
🎭 Polarity Classification	3-class (Positive / Negative / Neutral) sentiment head
🧩 Quadruple Extraction	Complete (holder, target, expression, polarity) output
📊 Weighted F1 Evaluation	Partial-overlap scoring using token-level Jaccard intersection
🔄 Cross-lingual Transfer	Train on English, evaluate on low-resource target languages
📦 Modular Architecture	Encoder, CRF, and classifier heads are independently configurable

---

📦 Datasets

Subtask 1 — Monolingual

Dataset	Language	Domain
norec	Norwegian	Professional reviews (multi-domain)
opener_en	English	Hotel reviews
opener_es	Spanish	Hotel reviews
multibooked_ca	Catalan	Hotel reviews
multibooked_eu	Basque	Hotel reviews
darmstadt_unis	English	University reviews (online)
mpqa	English	News (opinion annotations)

Subtask 2 — Cross-Lingual

Trained on high-resource English data, evaluated on:

Test Dataset	Language
opener_es	Spanish
multibooked_ca	Catalan
multibooked_eu	Basque

Data Format

Each dataset is in JSON format with the following schema:

{
  "sent_id": "opener/en/hotel/english00164-6",
  "text": "Even though the price is decent for Paris, I would not recommend this hotel.",
  "opinions": [
    {
      "Source":           [["I"],                    ["44:45"]],
      "Target":           [["this hotel"],           ["66:76"]],
      "Polar_expression": [["would not recommend"],  ["46:65"]],
      "Polarity":         "negative",
      "Intensity":        "average"
    }
  ]
}

---

📊 Results

Monolingual Performance

Average across all 7 datasets

Metric	Score
SF1 (Sentiment F1)	0.46
SP (Sentiment Precision)	0.62
SR (Sentiment Recall)	0.65

Per-dataset Breakdown:

Dataset	SF1	SP	SR
Opener_en	0.41	0.37	0.47
Opener_es	0.35	0.33	0.38
NoReC	0.23	0.30	0.18
Multibooked_ca	0.57	0.53	0.63
Multibooked_eu	0.53	0.40	0.71
darmstadt_unis	0.55	0.58	0.00
MPQA	0.52	0.55	0.00

---

Cross-Lingual Performance

Average across 3 target language datasets

Metric	Score
SF1 (Sentiment F1)	0.35
SP (Sentiment Precision)	0.85
SR (Sentiment Recall)	0.63

Per-dataset Breakdown:

Dataset	SF1	Precision	Recall
Opener_es	0.000	0.000	0.000
Multibooked_ca	0.481	0.461	0.503
Multibooked_eu	0.671	0.618	0.733

---

🔍 Key Observations

• 🟢 BERT significantly improves span extraction results over CRF baseline alone — especially for English and language-similar corpora
• 🟡 Multibooked_eu performs best in cross-lingual settings — likely due to the smaller dataset size and consistent hotel-review characteristics
• 🔴 Complex/ambiguous expressions (different polarity in different contexts) present a challenge across all datasets
• 📌 Character-level BERT representations outperform word-level representations as a comparison baseline

---

🚀 Installation

Prerequisites

• Python ≥ 3.8
• PyTorch ≥ 1.9
• CUDA (optional, for GPU acceleration)

1. Clone the Repository

git clone https://github.com/your-username/Structured-Sentiment-Analysis-IIITD-NLP-PROJECT.git
cd Structured-Sentiment-Analysis-IIITD-NLP-PROJECT

2. Install Core Dependencies

pip install torch torchvision transformers
pip install nltk scikit-learn tqdm gensim

3. Install Baseline Dependencies

pip install -r baseline/sequence_labeling/requirements.txt

4. Install Data Processing Dependencies

pip install -r data/requirements.txt

5. Prepare External Datasets

MPQA 2.0 — Download from the MPQA website and run:

cd data/mpqa && bash process_mpqa.sh

Darmstadt Service Review Corpus — Download from TU Darmstadt and run:

cd data/darmstadt_unis && bash process_darmstadt.sh

---

🛠️ Usage

Evaluation

Run the official evaluation script on model predictions:

python evaluate.py <input_dir> <output_dir>

Where:

• <input_dir>/res/ contains your predictions.json per dataset
• <input_dir>/ref/data/ contains the gold test files

Baseline — Training

# Train all BiLSTM baseline models across datasets
cd baseline/sequence_labeling
bash get_baselines.sh

Baseline — Inference

# Run inference on a specific dataset and split
python baseline/sequence_labeling/inference.py \
    --DATADIR opener_en \
    --FILE dev.json

Output will be saved to:

baseline/sequence_labeling/saved_models/relation_prediction/<DATADIR>/prediction.json

Predictions Format

Prediction files must match the gold data format. Each entry should look like:

{
  "sent_id": "unique-sentence-id",
  "text": "Raw input sentence here.",
  "opinions": [
    {
      "Source":           [["holder text"],     ["start:end"]],
      "Target":           [["target text"],     ["start:end"]],
      "Polar_expression": [["expression text"], ["start:end"]],
      "Polarity":         "positive"
    }
  ]
}

---

📁 Project Structure

Structured-Sentiment-Analysis-IIITD-NLP-PROJECT/
│
├── 📄 evaluate.py                  ← Official evaluation script (SF1 / SP / SR)
│
├── 📁 baseline/
│   └── sequence_labeling/
│       ├── extraction_module.py    ← BiLSTM span extractor (Holder / Target / Expr)
│       ├── relation_prediction_module.py  ← BiLSTM relation classifier
│       ├── inference.py            ← End-to-end inference pipeline
│       ├── convert_to_bio.py       ← Convert JSON data to BIO format
│       ├── convert_to_rels.py      ← Convert predictions to relation pairs
│       ├── utils.py                ← Data loading & vocabulary utilities
│       ├── WordVecs.py             ← Pretrained word embedding loader
│       ├── get_baselines.sh        ← Script to train all baseline models
│       └── requirements.txt
│
├── 📁 data/
│   ├── norec/                      ← Norwegian multi-domain reviews
│   │   ├── train.json
│   │   ├── dev.json
│   │   └── test.json
│   ├── opener_en/                  ← English hotel reviews
│   ├── opener_es/                  ← Spanish hotel reviews
│   ├── multibooked_ca/             ← Catalan hotel reviews
│   ├── multibooked_eu/             ← Basque hotel reviews
│   ├── mpqa/                       ← MPQA news corpus
│   ├── darmstadt_unis/             ← English university reviews
│   └── README.md                   ← Data format documentation
│
└── 📁 predictions/
    └── norec/
        └── predictions.json        ← Sample model predictions

---

🔭 Further Work

We identify several promising directions to build upon this work:

🕸️ Dependency Graph Parsing

In the future, we would likely move to a dependency graph parsing approach for structured sentiment — augmenting the token-level representation with their heads in a dependency tree (Kurtz et al., 2020). This allows richer relational reasoning between opinion components.

🌐 Multi-task & Cross-lingual Learning

• Exploring multi-task learning across languages to better leverage shared structure in multilingual sentiment graphs
• Joint training on all monolingual datasets with language-specific adapters

🔗 Advanced Graph Parsers

• Point Network (Samuel & Straka, 2020) — a strong graph parser for SSA
• PERIN — a permutation-invariant structured prediction framework

📐 Span-level Polarity

• Currently, polarity is predicted globally per sentence via the [CLS] token
• Moving to span-level polarity prediction would handle cases like "The food was great, but the service was terrible"

🤖 Serialized Large Language Models

• Explore whether large pre-trained models (e.g., GPT-4, LLaMA) can directly predict structured opinion tuples via in-context learning or fine-tuning, without an explicit CRF layer

---

📚 References

Barnes, J. et al. (2021). SemEval-2022 Task 10: Structured Sentiment Analysis.
  Proceedings of the 16th Workshop on Semantic Evaluation.

Liu, Y. et al. (2019). RoBERTa: A Robustly Optimized BERT Pretraining Approach.
  arXiv:1907.11692.

Kurtz, et al. (2020). Improving Low-Resource NMT through Relevance Based Linguistic Features.
  ACL 2020.

Samuel, D. & Straka, M. (2020). ÚFAL at MRP 2020: Permutation-Invariant Semantic Parsing.
  CoNLL 2020 Shared Task.

---

Made with ❤️ at IIIT Delhi | NLP Course Project

For questions or contributions, please open an issue or pull request.