PACRR
What it is
PACRR (Position-Aware Convolutional-Recurrent Relevance, Hui et al., 2017) applies convolutional filters of different sizes over the query-document term similarity matrix to capture n-gram-level interactions. Max-pooling retains the strongest signal; an LSTM aggregates query-term scores while preserving positional order. Compared to DRMM and KNRM, PACRR more explicitly models phrase-level and proximity matches.
[illustrate: Similarity matrix → CNN filters of size 1×n → max-pool → per-query-term feature → LSTM → relevance score]
How it works
-
Similarity matrix:
- Cosine similarity between query and document terms using word embeddings
-
Convolutional matching:
- CNN filters of size 1×n (n = 1, 2, 3) capture unigram, bigram, trigram document-side matches
- k-max pooling: retain top-k similarity values per filter
-
Recurrent aggregation:
- LSTM over query term features in query order
- Preserves term order within the query
-
Scoring:
- Linear combination of LSTM output → final relevance score
Variants and history
PACRR (2017) was one of several papers (alongside KNRM, DUET) developing the interaction-based neural ranking paradigm. MPACRR added multiple passages per document. Like DRMM and KNRM, PACRR was superseded by transformer models but contributed the CNN-over-similarity-matrix idea that influenced later multi-granularity matching work.
When to use it
Primarily historical. The CNN-over-interaction-matrix idea is pedagogically useful for understanding how neural models capture phrase-level relevance before transformers.