A retrieval approach that represents each document as a probability distribution over terms (a language model) and ranks documents by the likelihood that the model would generate the query, typically P(query | document-model), with smoothing to handle unseen terms.

Principle

Rank documents by how probable the observed query is under each document's generative language model; smoothing trades off document evidence with collection-level statistics to avoid zero likelihoods.

Demonstration

Illustrative scenario → Short query of two terms. Recognition → Build a smoothed term distribution for each document. Action → Compute P(query | document-model) as product or log-sum of term probabilities and rank documents by this score. Consequence → Documents that better 'generate' the query rise to the top, provided smoothing prevents zeroing out for unseen words.

Misapplication

Assuming query-likelihood scores equal P(relevant | document). The plausible mistake is because both are probabilistic statements; the error is failing to distinguish a generative likelihood (how well a document explains the query) from a posterior relevance probability.

Consequence

Offers a coherent generative interpretation, enables principled smoothing and collection-level prior incorporation, and often yields competitive baseline performance; retrieval quality depends on smoothing choice and the extent to which term-generation models capture relevance signals beyond lexical overlap.

Reversal

When relevance depends on document-level patterns not captured by term-generation (phrase structure, semantics, discourse) or when queries are long and user intent complex, query-likelihood ordering can misrank documents; discriminative or neural methods that model relevance directly may improve results.

Boundary

Clearly within: document-as-generator scoring by P(query | document-model) with explicit smoothing. Boundary case: using KL-divergence between query and document models as ranking equivalently related to query-likelihood under certain assumptions. Clearly outside: purely heuristic TF–IDF scoring without generative interpretation.

Semantic Tension

Tension between generative (model P(query|document)) and discriminative perspectives (directly model P(relevant|features)); each offers different trade-offs in interpretability and empirical optimization.

Synthesis

Treating documents as generators clarifies smoothing and collection priors; however, the approach's success depends on whether term-generation adequately captures the notion of relevance for the user's information need.