Combined Score Normalization (CombSUM)

The core mechanism of CombSUM is the unweighted summation of normalized scores. For a given document d, the final score is calculated as Score(d) = Σ norm_score_i(d) across all i retrieval systems. This assumes that higher combined scores indicate stronger overall relevance, and that all constituent systems contribute equally valuable evidence. The technique is mathematically simple, making it highly interpretable and easy to debug compared to learned fusion models.

CombSUM is critically dependent on a preceding normalization step. Raw relevance scores from different systems (e.g., BM25 and cosine similarity) exist in vastly different, incomparable ranges. Without normalization, a system producing scores in the thousands would dominate one producing scores between 0 and 1. Common normalization techniques used before CombSUM include:

• Min-Max Normalization: Rescales scores to a fixed [0, 1] range.
• Z-Score Normalization: Standardizes scores based on the mean and standard deviation of the result set.

CombSUM operates on the principle that a document is more likely to be relevant if it is consistently retrieved by multiple independent systems. This is known as the 'chorus effect'. If a document appears in the top results of both a sparse (lexical) and a dense (semantic) retriever, its summed score will naturally be high. This implicitly penalizes documents that are a strong match for only one system but completely missed by others, acting as a simple form of noise reduction.

CombSUM is often compared to its variant, CombMNZ (Multiply by Non-Zero). While CombSUM simply sums the scores, CombMNZ multiplies the sum by the number of systems that returned the document: Score(d) = Σ norm_score_i(d) * count_i(d). This gives an explicit bonus to documents found by more systems. CombSUM is a more conservative approach, as it doesn't artificially inflate the score based on retrieval count, making it preferable when the overlap between systems is already high or when a single system's score is a very strong signal.

A practical consideration for CombSUM is the treatment of documents not retrieved by a specific system. A missing document is typically assigned a normalized score of zero for that system's contribution. This ensures the document can still be ranked based on the scores from systems that did retrieve it. This behavior is fundamental to the 'chorus effect', as a document found by only one system will have a lower combined score than one found by all, but it will not be discarded entirely.

CombSUM is a classic late fusion technique. It is applied after each independent retrieval pipeline (e.g., a sparse index and a dense vector index) has returned its own ranked list. The fused, re-ranked list produced by CombSUM can then be passed to a final, more computationally expensive re-ranker, such as a Cross-Encoder. This architecture allows the system to combine diverse signals cheaply before applying a precision-focused model to the top candidates.