CEMTM: Contextual Embedding-based Topic Model

A novel multimodal topic modeling framework that leverages Vision-Language Models (VLMs) to discover coherent topics from documents containing both text and images.

Abstract

We introduce CEMTM, a context-enhanced multimodal topic model designed to infer coherent and interpretable topic structures from both short and long documents containing text and images. CEMTM builds on fine-tuned large vision language models (LVLMs) to obtain contextualized embeddings, and employs a distributional attention mechanism to weight token-level contributions to topic inference. A reconstruction objective aligns topic-based representations with the document embedding, encouraging semantic consistency across modalities. Unlike existing approaches, CEMTM can process multiple images per document without repeated encoding and maintains interpretability through explicit word-topic and document-topic distributions. Extensive experiments on six multimodal benchmarks show that CEMTM consistently outperforms unimodal and multimodal baselines, achieving a remarkable average LLM score of 2.61. Further analysis shows its effectiveness in downstream few-shot retrieval and its ability to capture visually grounded semantics in complex domains such as scientific articles.

Overview

CEMTM (Contextual Embedding-based Topic Model) addresses the limitations of traditional topic models by incorporating multimodal information through contextualized embeddings. This approach:

• Leverages VLMs: Uses pretrained Vision-Language Models to extract rich contextual embeddings from text-image pairs
• Token-level Analysis: Learns importance weights for individual tokens to identify the most relevant content for topic discovery
• Multimodal Integration: Seamlessly combines textual and visual information in a unified embedding space
• Contextual Understanding: Captures semantic relationships that traditional bag-of-words approaches miss
• Flexible Architecture: Supports various VLM backbones and can be adapted to different multimodal datasets

Key Contributions

1. Novel Architecture: Introduction of an importance network that learns to weight tokens based on their relevance to topic discovery
2. Multimodal Topic Discovery: First approach to use contextualized embeddings from VLMs for joint text-image topic modeling
3. Comprehensive Evaluation: Extensive evaluation on multiple datasets with both automatic metrics and human evaluation
4. Superior Performance: Demonstrates improved topic coherence and diversity compared to traditional and neural topic models

Architecture

CEMTM consists of four key components working together to extract topics from multimodal documents:

1. VLM2Vec Module

• Purpose: Extracts contextualized embeddings from text-image pairs
• Implementation: Wrapper around Vision-Language Models
• Output:
  • Contextual token embeddings: H ∈ ℝ^(N×D) where N is sequence length, D is embedding dimension
  • Document-level embedding: e_d ∈ ℝ^D

2. Topic Encoder

• Purpose: Maps contextual embeddings to topic space
• Function: t_i = softmax(W_t h_i) where t_i ∈ ℝ^K (K topics)
• Output: Token-level topic distributions

3. Importance Network

• Purpose: Learns which tokens are most relevant for topic discovery
• Architecture: Multi-layer Transformer encoder
• Function: Outputs μ and σ² parameters for importance distribution
• Sampling: α_i ~ N(μ_i, σ_i²) followed by β_i = softmax(α) for importance weights

4. Reconstruction Module

• Purpose: Ensures learned representations preserve document semantics
• Function: e_d' = f_dec(∑_i β_i t_i) where f_dec is a linear decoder
• Loss: Minimizes reconstruction error between e_d and e_d'

Mathematical Formulation

The model optimizes the following objective:

L = L_rec + λ_ent L_ent + λ_kl L_kl

Where:

• L_rec: Reconstruction loss ||e_d - e_d'||²
• L_ent: Entropy regularization ∑_i β_i log β_i
• L_kl: KL divergence between learned and prior importance distributions

Requirements

System Requirements

• Python 3.8+
• CUDA-compatible GPU (recommended)
• 8GB+ RAM

Dependencies

Install the required packages:

pip install -r requirements.txt

Additional Setup

1. NLTK Data: Required for text preprocessing

   python -c "import nltk; nltk.download('stopwords')"
   

2. Word Embeddings (for evaluation metrics): Download GloVe embeddings for the WE (Word Embedding) metric:

   mkdir -p data/embeddings
   wget http://nlp.stanford.edu/data/glove.6B.zip
   unzip glove.6B.zip -d data/embeddings/
   

3. OpenAI API Key (optional, for LLM evaluation):

   export OPENAI_API_KEY="your-api-key-here"
   

Supported Datasets

CEMTM has been evaluated on multiple multimodal datasets:

Text-Image Datasets

• MS-COCO: Image captioning dataset with rich visual-textual associations
• WikiWeb2M: Large-scale Wikipedia articles with associated images
• SPIQA: Scientific paper abstracts with figures and diagrams
• TQA: Textbook question-answering with educational diagrams
• VIST: Visual storytelling dataset with image sequences and narratives

Meme and Social Media Datasets

• FHM (Facebook Hateful Memes): Multimodal content analysis
• T4SA: Twitter sentiment analysis with images

Each dataset loader handles the specific format and preprocessing requirements automatically.

Dataset Setup

Choose one of the supported datasets and set up the data directory:

Option 1: MS-COCO (Recommended for testing)

mkdir -p data/corpus/annotations data/corpus/train2014
# Download MS-COCO 2014 train images and annotations
# Place captions_train2014.json in data/corpus/annotations/
# Place images in data/corpus/train2014/

Option 2: Other Supported Datasets

• WikiWeb2M: Place wikiweb2m.jsonl and image folders in data/corpus/
• SPIQA: Place metadata.json and images/ folder in data/corpus/
• TQA: Place tqa.json and images/ folder in data/corpus/
• FHM: Place hateful_memes.jsonl and img/ folder in data/corpus/
• T4SA: Place t4sa.csv and images/ folder in data/corpus/
• VIST: Place stories.json and images/ folder in data/corpus/

Configuration

Key Hyperparameters

Understanding the main configuration parameters:

Model Architecture

model:
  d_model: 4096              # VLM embedding dimension (depends on chosen VLM)
  n_topics: 50               # Number of topics to discover
  transformer_layers: 2      # Layers in importance network
  transformer_heads: 8       # Multi-head attention heads
  dropout: 0.1              # Regularization

Training Configuration

training:
  batch_size: 4             # Adjust based on GPU memory
  num_epochs: 20            # Training epochs
  gradient_clip: 1.0        # Gradient clipping for stability
  device: cuda              # Use GPU for faster training

Loss Weights

loss:
  lambda_entropy: 0.01      # Entropy regularization weight
  lambda_kl: 0.1           # KL divergence weight

Optimization

optimizer:
  name: adamw
  lr: 5e-5                 # Learning rate
  weight_decay: 0.01       # L2 regularization

scheduler:
  name: linear             # Learning rate schedule
  warmup_steps: 500        # Warmup period

Usage

Training

Train the CEMTM model:

python scripts/train_cemtm.py --config config/config.yaml

Training outputs:

• Model checkpoints saved to outputs/
• Training logs in logs/

Evaluation

Evaluate trained model on topic quality metrics:

python scripts/evaluate_topics.py --config config/config.yaml

Vocabulary Quality Check

After training, you can verify the quality of the generated vocabulary using the vocabulary checker tool:

python scripts/check_vocabulary.py --vocab_path outputs/vocabulary.json

This tool provides:

• Vocabulary Statistics: Total size, word length distribution
• Quality Checks: Detects duplicates, suspicious short words, numeric tokens
• Sample Words: Shows most and least frequent words
• Recommendations: Suggests improvements based on vocabulary characteristics

When to Use:

• After training to verify vocabulary was built correctly
• If you're getting low NPMI scores (may indicate vocabulary issues)
• When tuning vocab_size or min_word_freq parameters
• To ensure text preprocessing is working properly

Configuration Parameters: You can adjust vocabulary generation in config/config.yaml:

data:
  vocab_size: 2000         # Maximum vocabulary size
  min_word_freq: 5         # Minimum frequency threshold (filters rare words)

Tips:

• Low NPMI scores? Increase min_word_freq to filter more rare words (e.g., 10-20)
• Too small vocabulary? Decrease min_word_freq or increase vocab_size
• Too many noisy words? Improve text preprocessing or increase min_word_freq

Evaluation Metrics

CEMTM provides comprehensive evaluation through multiple metrics:

Automatic Metrics

• Topic Diversity (TD): Measures the proportion of unique words across all topics

  • Higher values indicate more diverse topics
  • Range: [0, 1], where 1 means all topic words are unique

• Inverse Rank-Biased Overlap (I-RBO): Measures topic distinctiveness

  • Computes 1 - average RBO across all topic pairs
  • Higher values indicate more distinct topics

• Normalized Pointwise Mutual Information (NPMI): Measures topic coherence

  • Uses co-occurrence statistics from the corpus
  • Range: [-1, 1], where higher values indicate more coherent topics

• Word Embedding Score (WE): Semantic coherence based on word embeddings

  • Computes average cosine similarity between topic words
  • Requires pretrained word embeddings (GloVe/Word2Vec)

LLM-based Evaluation

• LLM Score: GPT-based topic quality assessment
  • Uses OpenAI API to rate topic coherence on a 1-3 scale
  • Provides human-like evaluation of topic interpretability

Clustering Metrics (when ground truth available)

• Purity: Measures cluster homogeneity
• Adjusted Rand Index (ARI): Similarity between predicted and true clusters
• Normalized Mutual Information (NMI): Information-theoretic clustering metric

Directory Structure

CEMTM/
├── config/
│   └── config.yaml           # Configuration file
├── data/
│   ├── dataset.py           # Dataset loaders
│   └── preprocessing.py     # Text/image preprocessing
├── evaluation/
│   ├── llm_api.py          # OpenAI API interface
│   ├── metrics.py          # Evaluation metrics
│   └── topic_extraction.py # Topic analysis utilities
├── losses/
│   └── losses.py           # Loss functions
├── model/
│   ├── cemtm.py           # Main CEMTM model
│   ├── encoder.py         # Topic encoder
│   ├── importance_net.py  # Importance network
│   └── vlm2vec.py        # Vision-Language Model wrapper
├── scripts/
│   ├── evaluate_topics.py # Evaluation script
│   └── train_cemtm.py    # Training script
├── trainer/
│   ├── trainer.py        # Training logic
│   └── utils.py         # Training utilities
└── requirements.txt     # Dependencies

Troubleshooting

Common Issues

1. CUDA Out of Memory:

   • Reduce batch_size in config.yaml
   • Use device: cpu for CPU-only training

2. Import Errors:

   • Ensure you're running from the CEMTM root directory
   • Install all requirements: pip install -r requirements.txt

3. Dataset Loading Errors:

   • Check dataset path in config.yaml
   • Ensure dataset files are in correct format and location

4. Evaluation Errors:

   • Download word embeddings for WE metric
   • Set OPENAI_API_KEY for LLM evaluation

Computational Requirements

• GPU Memory: 8-12GB VRAM recommended for batch_size=4
• Training Time: 2-4 hours on RTX 3080 for 20 epochs on MS-COCO subset
• Inference: Real-time topic extraction for new documents

Advantages over Traditional Methods

• Contextual Understanding: Captures semantic relationships beyond word co-occurrence
• Multimodal Integration: Leverages both text and visual information
• Better Coherence: Produces more interpretable and coherent topics
• Scalability: Efficient processing of large multimodal datasets

Related Work

CEMTM builds upon and extends several lines of research:

• Topic Modeling: LDA, Neural Topic Models, BERTopic
• Vision-Language Models: CLIP, LLaVA, BLIP
• Multimodal Learning: Cross-modal attention, multimodal transformers
• Contextualized Embeddings: BERT, RoBERTa, contextual topic models

License

This project is licensed under the MIT License - see the LICENSE file for details.

Contributing

1. Fork the repository
2. Create a feature branch
3. Make your changes
4. Add tests if applicable
5. Submit a pull request

Contact

For questions or issues, please:

• Open a GitHub issue for bug reports or feature requests
• Contact the authors: [amirhossein.abaskohi@gmail.com]
• Check the paper for theoretical details and experimental results

Acknowledgements

This research was supported by the Natural Sciences and Engineering Research Council of Canada (NSERC).
Ce projet a été financé par le Conseil de recherches en sciences naturelles et en génie du Canada (CRSNG).

Citation

@inproceedings{abaskohi-etal-2025-cemtm,
    title = "{CEMTM}: Contextual Embedding-based Multimodal Topic Modeling",
    author = "Abaskohi, Amirhossein  and
      Li, Raymond  and
      Li, Chuyuan  and
      Joty, Shafiq  and
      Carenini, Giuseppe",
    editor = "Christodoulopoulos, Christos  and
      Chakraborty, Tanmoy  and
      Rose, Carolyn  and
      Peng, Violet",
    booktitle = "Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing",
    month = nov,
    year = "2025",
    address = "Suzhou, China",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.emnlp-main.590/",
    doi = "10.18653/v1/2025.emnlp-main.590",
    pages = "11675--11692",
    ISBN = "979-8-89176-332-6",
    abstract = "We introduce CEMTM, a context-enhanced multimodal topic model designed to infer coherent and interpretable topic structures from both short and long documents containing text and images. CEMTM builds on fine-tuned large vision language models (LVLMs) to obtain contextualized embeddings, and employs a distributional attention mechanism to weight token-level contributions to topic inference. A reconstruction objective aligns topic-based representations with the document embedding, encouraging semantic consistency across modalities. Unlike existing approaches, CEMTM can process multiple images per document without repeated encoding and maintains interpretability through explicit word-topic and document-topic distributions. Extensive experiments on six multimodal benchmarks show that CEMTM consistently outperforms unimodal and multimodal baselines, achieving a remarkable average LLM score of 2.61. Further analysis shows its effectiveness in downstream few-shot retrieval and its ability to capture visually grounded semantics in complex domains such as scientific articles."
}