Search and Ranking Systems
Staff Engineer @ BloomReach — Oct 2013 to Mar 2018
Location: Bengaluru, India
Overview
Led the design and implementation of an automated synonym generation pipeline for BloomReach Discovery, processing 100 million+ product descriptions and 30 million+ queries weekly to extract high-quality synonym pairs for e-commerce search.
As deep learning started emerging, designed deep NNs to match user queries with document text & images.
Key Features
- Data Mining & Representation:
- Extract candidate phrase pairs from product catalog text, Web crawl, and search logs.
- Build contextual embeddings via co-occurrence contexts, click-through distributions, document-vector averages, and session proximity.
- Similarity Metrics & Filtering:
- Combine symmetric (JS divergence) and asymmetric (hyponym/hypernym) measures, augmented by edit-distance heuristics.
- Apply phrase-specific thresholds (e.g. ensuring D(P, Q) < D(P, P′) where P′ is a known variant) to filter high-confidence pairs.
- Train a supervised classifier on multi-score feature vectors to rank and grade synonyms.
- Continuous Learning: Weekly batch updates to capture evolving language, new trends, and seasonal shifts in user queries.
Patents
Resources
- Blog: Synonym Generation