ETD

• aggregation
• BM25F
• CW22-B
• encoder
• IR
• multi-field
• retrieval
• web

Web documents are not simple blocks of text. A page has a title, a URL, a body and a set of incoming anchor texts, each describing the page in a different way and with a different degree of reliability. Classical lexical retrieval has exploited this structure for two decades through BM25F, which merges per-field term frequencies before the saturation step and therefore treats the document as a single weighted pseudo-field. Modern neural retrieval, in contrast, has mostly ignored document structure: standard dense retrievers concatenate the available text, truncate it to the encoder’s maximum sequence length and produce a single embedding per document. This gap between the two approaches motivates this thesis.
We introduce NeuralBMF, a family of multi-field neural retrieval models that keeps the fields separate throughout the retrieval pipeline. Each field is independently encoded by a frozen bi-encoder, and its relevance score is computed as the dot product between the query and field embeddings. A lightweight aggregation function then combines the field-level scores into a final document score. We isolate this aggregation step as the main object of study, comparing parameterized functions, such as linear combinations and query-conditioned attention, with parameter-free alternatives, such as maximum pooling and a convex combination of the maximum with a properly tuned BM25F score. A cross-encoder stage, NeuralBMF-CE, re-ranks the first- stage shortlist by concatenating the fields into a single sequence for MonoT5, with field order treated as an additional design variable. Experiments are conducted on ClueWeb22-B using MS MARCO Web Search queries and are replicated across two bi-encoders, a second query set (Researchy Questions), and three BEIR collections.
The results show that multi-field scoring consistently outperforms single-field dense retrieval. Surprisingly, simple aggregation strategies based on maximum pooling outperform more complex learned heads, suggesting that additional aggregation capacity provides little benefit. The title and URL provide most of the signal, while the anchor field mainly benefits the cross-encoder. Overall, the findings generalize within the web domain and partially transfer to body-dominated collections.