Bilkent University
Department of Computer Engineering
M.S.THESIS PRESENTATION
COSMIC: Cosine-Supervised Modeling of Informative Coefficients for Sparse-Term Retrieval
Ekrem Polat
Master Student
(Supervisor:Prof.Dr.Özgür Ulusoy)
Computer Engineering Department
Bilkent University
Abstract: Learned sparse retrieval improves first-stage ranking by replacing raw term-frequency statistics with learned importance scores, while remaining compatible with inverted indexes and BM25-style lexical matching. How these scores should be supervised, however, remains an open question. Existing methods rely on query-term recall heuristics, document-level ranking objectives, or vocabulary-wide expansion. While highly effective, these approaches either learn term importance indirectly from aggregate relevance signals or produce sparse representations that extend beyond the terms observed in the original text. This thesis proposes CoSMIC, a learned sparse retrieval framework that supervises term importance directly at the token level. Its starting point is a simple observation: the token-level MaxSim cosine similarities computed by a late-interaction model already indicate which passage terms are responsible for relevance. CoSMIC distils this signal from a frozen ColBERTv2 teacher, converting these similarities into sharpened, sparse term-importance targets. A BERT-based term-weighting model is then trained to reproduce them. The predicted coefficients are quantized into integer pseudo-term frequencies and stored in an ordinary inverted index, so that retrieval requires no query encoder and no query-time neural computation. Such term-level supervision, however, constrains term weights only within a passage. The thesis therefore introduces CoSMIC+, which adds pairwise margin calibration from a cross-encoder teacher. The two signals are complementary: the late-interaction teacher shapes term weights within a passage, while the cross-encoder teacher calibrates the resulting scores across competing passages. CoSMIC is evaluated on MS MARCO Passage Ranking, the TREC Deep Learning 2019 and 2020 passage-ranking query sets, and MS MARCO Document Ranking. It consistently improves over BM25 and over comparable contextual term-weighting baselines, and CoSMIC+ yields further gains. Although the base framework is deliberately expansion-free, combining CoSMIC+ with DocT5Query-expanded passages improves effectiveness further. This shows that learned term weighting and document expansion are complementary techniques
DATE: July 22, Wednesday @ 14:30
Place: EA 516