Simple BM25 extension to multiple weighted fields

By Stephen E. Robertson, Hugo Zaragoza, Michael J. Taylor. International Conference on Information and Knowledge Management, 2004

812 citations recorded by the source corpus.

Subfields: Artificial Intelligence, Information Systems

View Paper: https://www.semanticscholar.org/paper/10b2752c9940e9c761a161a0fef7ae08cae66301

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

1. Tie-Yan Liu
2. Jimmy J. Lin
3. Tao Qin
4. Nick Craswell
5. Bhaskar Mitra
6. Hang Li
7. Olivier Chapelle
8. Jamie Callan

In Information Systems

1. Olivier Chapelle
2. Hang Li
3. Nick Craswell
4. Tie-Yan Liu
5. Yi Chang
6. Bhaskar Mitra
7. Jun Xu
8. Tao Qin

Citations in the corpus

• Towards more accurate retrieval of duplicate bug reports (2011, background): “Unfortunately, the accuracy of proposed techniques for duplicate bug report detection through ranking [3]–[6] is still low.”
• Understanding user's query intent with wikipedia (2009, methodology): “It has been proven to be quite effective especially in text document retrieval [2].”
• Voting for candidates: adapting data fusion techniques for an expert search task (2006, methodology): “[29] showed improved retrieval effectiveness in Web tasks when the contribution of each field to the document ranking was controlled by the use of weights.”
• Query dependent pseudo-relevance feedback based on wikipedia (2009, methodology): “Robertson et al. [26] gives a detailed analysis of the disadvantages of linear combination of the scores obtained from scoring every field, and recommends a simple weighted field extension.”
• Overview of Touché 2022: Argument Retrieval (2022, neutral): “Their argument search engine args.me uses BM25F [71] to rank the indexed arguments, giving conclusions more weight than premises.”
• BISON: BM25-weighted Self-Attention Framework for Multi-Fields Document Search (2020, background): “Lots of studies (36; 27) reveal that different fields contain complementary information.”
• Improving bug localization using structured information retrieval (2013, background): “While we believe modeling source code structure is novel for bug localization, we also note that the concept of modeling document structure in IR is quite old (e.g., Google in 1998 [4] and more recent BM25F [31]).”
• An Introduction to Neural Information Retrieval (2018, neutral): “Incorporating these differernt sources of document description is useful to determine the relevance of the document to a search query [Robertson et al., 2004].”
• Improving web search ranking by incorporating user behavior information (2006, positive): “The scoring function and field-specific tuning is described in detail in [23].”
• Galago: A Modular Distributed Processing and Retrieval System (2012, methodology): “We have additionally used combinations of operators and traversals to implement the Relevance Model [8], the field-based PRMS model [7], BM25 scoring [11] and it’s field-based variant [10], to name a few of the implemented models.”
• Scalable clustering of news search results (2011, methodology): “Given a query q and a document d, BM25 score can be computed as K x¯d,t (1) BM25F (d)= wt K1 +¯xd,t t.q.d xd,t x¯d,t = |d| 1+ B( - 1) |avgdl| (1) where wt is the usual relevance weight of the term t in the document, which can often be the inverse document fre­quency, xd,t is the frequency of the t”
• An anatomy for neural search engines (2019, background): “Furthermore, oftentimes a document might be comprised of several components, like title, main body, meta tags and anchor texts, and to perform feature engineering from these sources, plus the decision of how to weight them in the final f matching function, is not trivial [Robertson et al., 2004].”
• A Study of Information Retrieval Weighting Schemes for Sentiment Analysis (2010, background): “…approach doesn’t take into consideration a number of tested notions from IR, such as the non-linearity of term frequency to document relevancy (e.g. Robertson et al. (2004)) according to which, the probability of a document being relevant to a query term is typically sublinear in relation to the…”
• Understanding the Semantic Structure of Noun Phrase Queries (2010, methodology): “Knowing the semantic role of each query constituent, we can reformulate the query into a structured form or reweight different query constituents for structured data retrieval (Robertson et al., 2004; Kim et al., 2009; Paparizos et al., 2009).”
• Spark: top-k keyword query in relational databases (2007, background): “We note that a similar observation and remedy about the need of non-linear term frequency attenuation was also made by IR researchers [26].”
• LambdaMerge: merging the results of query reformulations (2011, background): “Traditional IR rankers such as BM25 [22] have few parameters, so that it is possible to tune them by hand.”
• Zero-shot Neural Retrieval via Domain-targeted Synthetic Query Generation (2020, background): “For instance, BM25 does exceptionally well on BIOASQ, but very poorly on most other sets.”
• Yahoo! Learning to Rank Challenge Overview (2010, methodology): “It is a combination of BM25F (Robertson et al., 2004) and Metzler’s sequential dependence (SD) model (Metzler and Croft, 2005), which provides an effective framework for term proximity matching.”