SPLADE

Definition

SPLADE (Sparse Lexical and Expansion / SParse Lexical AnD Expansion) is a neural sparse retrieval model that creates learned sparse embeddings via BERT’s Masked Language Model (MLM) head. It combines semantic understanding from transformers with the efficiency of inverted-index retrieval.

Developed by Thibault Formal, Stéphane Clinchant, and Benjamin Piwowarski at NAVER LABS Europe.

Core Mechanism

Input text → BERT → MLM head → 30,522-dim vocabulary distribution per token
→ Log-ReLU activation
→ Max-pool across tokens
→ Sparse vector (≈100-200 non-zero entries out of 30,522)

For each token in the input, the MLM head predicts probability over the entire vocabulary — enabling term expansion (predicting relevant terms not in the original text) and compression (suppressing uninformative terms).

Example: For document about binary numbers, SPLADE:

  • Expands: adds “computing”, “digit” (semantically related)
  • Compresses: removes conjunctions and articles
  • Result: 23-term sparse vector from a 60-term passage

Key Technical Components

Log saturation: Prevents single terms from dominating scores.

FLOPS regularizer: Penalizes computation cost to encourage sparsity, acts as implicit stop-word removal.

Versions

VersionInnovation
SPLADE v1Original: both query & document expansion
SPLADE v2Max pooling over the vocabulary axis; document-only expansion (faster queries)
SPLADE-V3 (2024)Updated models; Hugging Face release

vs. BM25 and Dense Retrieval

BM25SPLADEDense (Bi-Encoder)
VocabularyFixed (term frequency)Expanded (learned)None (continuous)
SemanticsNoneGood (via BERT)Excellent
SpeedVery fastFast (inverted index)Fast (ANN)
InterpretabilityHighHigh (vocabulary terms)Low
Domain adaptationManualLearnedLearned

Domain Fine-Tuning

Public SPLADE checkpoints are trained on MS MARCOweb queries against Wikipedia passages. In e-commerce that mismatch shows up directly in the learned term weights and synonym expansions, which reflect encyclopedic text rather than a product catalog.

Fine-tuning on catalog data measurably closes the gap. In Fine-Tuning Sparse Embeddings for E-Commerce Search, training from distilbert-base-uncased on the Amazon ESCI Dataset gave nDCG@10 0.389 vs 0.305 for BM25 (+27.5%), where off-the-shelf SPLADE reached only 0.326 (+7.2%). What changed: brand names gained weight, generic words like “good” lost it, and domain vocabulary such as “refurbished” became meaningful.

The cost is generality. The ESCI-tuned model lost to off-the-shelf SPLADE on Home Depot data (0.384 vs 0.391) and fell far below BM25 on MS MARCO (0.751 vs 0.915). Multi-domain training recovers cross-catalog consistency at the price of peak in-domain accuracy.

Full vs inference-free SPLADE

The inference-free variant skips the encoder pass on the query side to save latency. Evgeniya Sukhodolskaya argues against it for e-commerce: the domain is intent-heavy, and “Apple juice” vs “Apple iPhone” are two distinct intents around the same token — precisely the distinction the query encoder provides.

Vocabulary limits

SPLADE’s output dimensions are the base model’s vocabulary, so out-of-vocabulary terms cannot be represented and there is no fallback path — you must pick a base model that already knows your tokens. miniCOIL addresses this directly by reverting to a BM25 weight for untrained words.

Tooling: Sentence Transformers v5 (MLMTransformer + SpladePooling, SpladeLoss) and qdrant-sparse-finetune.

Advantages Over Dense Retrieval

  • No Vector Search infrastructure needed — works with standard inverted indexes
  • Interpretable representations (vocabulary-dimension vectors)
  • Strong zero-shot performance
  • Easier integration into existing Hybrid Search pipelines

Articles

Videos

People