Channel: #retrieval-and-search

The Retrieval and Search Subprogram drives foundational and applied research in the design, evaluation, and deployment of next-generation retrieval systems. With growing demand for multi-modal, multi-lingual, and context-aware search, our work spans theory, algorithms, and systems across classical IR and neural architectures.

Leads: 

Focus Areas

Neural Information Retrieval (Neural IR)

• Dense retrieval using dual encoders, late interaction (ColBERT), and generative retrievers.

• Specialized architectures for domain-adapted retrieval (scientific, biomedical, multilingual corpora).

• In-context retrieval strategies for grounding large language models (LLMs).

• Retrieval-Augmented Generation (RAG) optimization techniques.
  
  

2. Representation Learning for Search

• Pretraining and fine-tuning of text, image, and multimodal encoders for semantic similarity.

• Contrastive learning and mutual information maximization.

• Evaluation of learned embeddings for clustering, nearest-neighbor search, and knowledge transfer.

• Cross-modal representation learning (image-text, code-docstring, audio-text).
  
  

3. Advances in Ranking and Reranking Models

• Learning-to-Rank (LTR) frameworks.

• Differentiable ranking approximations and unbiased learning-to-rank from click data.

• Sparse vs. dense retrieval trade-offs and hybrid reranking pipelines.

• Fine-grained ranking in recommendation and personalized search.
  
  

4. Multimodal and Multilingual Search

• Retrieval models that generalize across modalities (e.g., image→text, table→text).

• Query reformulation and disambiguation in multilingual retrieval scenarios.

• Multilingual dense retrieval benchmarks and transfer learning across languages.

• Language-agnostic embedding spaces via knowledge distillation and adversarial learning.
  
  

5. Recommendation Systems and Personalization

• Sequential, graph-based, and session-aware recommenders.

• Large-scale implicit feedback modeling with user/item embeddings.

• Causal modeling in recommendations (counterfactual reasoning, uplift modeling).

• Representation learning for user behavior and item metadata integration.

• Late and Early fusion methods.
  
  

6. Efficient and Scalable Retrieval

• Vector search algorithms (e.g., ScaNN, HNSW) for billion-scale corpora.

• Memory-efficient approximate nearest neighbor (ANN) and Product Quantization techniques.

• Knowledge distillation for retrieval model compression.

• Index-aware training and optimization and latency-recall trade-offs in production settings.
  
  

7. Evaluation, Robustness, and Fairness

• Robustness of retrieval and ranking models under distributional shift.

• Evaluation metrics for semantic retrieval beyond precision/recall (e.g., nDCG, MRR, etc.).

• Fairness, diversity, and explainability in ranking and recommendation.

• Behavioral testing and model auditing for retrievers and recommenders.

And much more.....