240 likes | 346 Views
On Enhancing the User Experience in Web Search Engines. Franco Maria Nardini. About Me. I joined the HPC Lab in 2006 Master Thesis Ph.D. in 2011, University of Pisa Thesis: “Query Log Mining to Enhance User Experience in Search Engines” m ail: francomaria.nardini@isti.cnr.it
E N D
On Enhancing the User Experiencein Web SearchEngines Franco Maria Nardini
About Me • I joined the HPC Lab in 2006 • Master Thesis • Ph.D. in 2011, University of Pisa • Thesis: “Query Log Mining to Enhance User Experience in Search Engines” • mail: francomaria.nardini@isti.cnr.it • web: http://hpc.isti.cnr.it/~nardini • skype: francomaria.nardini
Query Suggestion with Daniele Broccolo, Lorenzo Marcon RaffaelePerego, FabrizioSilvestri
Our Contribution: Search Shortcuts • SearchShortcuts: • Ituses the “happy ending” stories in the query log to help new users; • Efficient: • All the “stuff” isstored on a invertedindex: retrievalproblem; • Effective: (head, torso, tail) • New evaluationmethodologyconfirmingthisevidencies: TREC DiversityTrack. Daniele Broccolo, Lorenzo Marcon, Franco Maria Nardini, Fabrizio Silvestri, Raffaele Perego, GeneratingSuggestionsforQueries in the Long TailwithanInvertedIndex, IP&M, 2011.
What’sNext?! • WhynottouseMachineLearning? • Machinelearningishelping a lot in the IR community; • Better and “fine-graned” ranking asitcould take into account importantsignalsthatare notfully-exploitednowadays; • Itmayhelps in filteringredundantsuggestions and choosing the “best” expressiveones (for eachintent). under exploration with • MarcinSydow (PJIIT), RaffaelePerego, FabrizioSilvestri
Signals • Whichsignalswewouldliketocapture? • Relevance to the givenquery; • Diversity with respect to a subtopiclist; • Serendipityofsuggestions; • Novelty with respect to news/trends on Twitter; • How do we catch them? • How do we combine them? • The “training” set is a problem.
QuerySuggestion: Ranking • A two-step architecture • First step to produce a list of candidates; • Second stepas a MLarchitecturecomposed of twodifferent (cascade) stages of ranking: • First round to ranksuggestionsw.r.t. the query; • Second round to understand“diversity”.
Diversification ofWeb Search Engine Results with Gabriele Capannini, RaffaelePerego, FabrizioSilvestri
Our Contribution • We design a method for efficiently diversify results from Web search engines. • Same effectiveness of other state-of-the-art approaches; • Extremely fast in doing the “hard” work; • Intents behind “ambiguous” queries are mined from query logs; Capannini G., Nardini F.M., Silvestri F., Perego R., A Search Architecture Enabling Efficient Diversification of Search Results, Proc. DDR Workshop 2011. Capannini G., Nardini F.M., Silvestri F., Perego R., Efficient Diversification of Web Search Results. Proceedings of VLDB 2011 (PVLDB), Volume 4, Issue 7.
What’s Next? • A modern ranking architecture: • Effective: • Users should be happy of the results they receive; • Efficient: • Low response times (< 0.1 s); • Easy to adapt: • Continuous crawling from the Web; • Continuous users’ feedback; with BerkantBarlaCambazoglu (Yahoo! Barcelona), Gabriele Capannini, RaffaelePerego, FabrizioSilvestri
Let’s Plug All Together • A way for efficiently diversifying “ambiguous” queries; • SS teaches how to “diversify” the current user query; • Scorerdiv computes the diversity “signal” of each document and rerank the final results list; First Phase Second Phase Index BM25 Scorer1 Scorerdiv Scorern … Possible intents behind the query SS Query Results
Retrieval over Query Sessions with M-DyaaAlBakour (University of Glasgow)
MainGoals • Question1) • Can Web searchenginesimprovetheir performancebyusingprevioususerinteractions? (includingpreviousqueries, clicks on rankedresults, dwelltimes, etc.) • Question2) • How do weevaluatesystem performance over an entirequery session instead of a single query?
TREC Session Track • Twoeditions of the challenge: 2010, 2011 • query, previousqueries; • urls + docs, urls + docs + dwell time; • Twodifferentevaluations: last subtop., allsubtop. • “Query expansion” with SearchShortcuts: • weighted by means of userinteraction data; • “history-based” recommendation; • Follow-up withtuningof the parameters. Ibrahim Adeyanju, Franco Maria Nardini, M-DyaaAlbakour, Dawei Song, UdoKruschwitz, RGU-ISTI-Essex at TREC 2011 Session Track, TREC Conference, 2011. Franco Maria Nardini, M-DyaaAlbakour, Ibrahim Adeyanju, UdoKruschwitz, Studying Search Shortcuts in a Query Log to Improve Retrieval Over Query Sessions, SIR 2012 in conjunction with ECIR 2012.
Some Results • Entity-based representation of the user session. • to reduce the “sparsity” of the space. • What’s Next?
Challenges • How those systems really affect (and modify) the behavior of the user? • Is it possible to quantify it? (metrics?) • What do we need to observe? • Toward the “perfect result page”: • accurate models for blending different sources of results.
Little Announcement • Models and TechniquesforTouristFacilities • Evaluation and Test Collections • UserInteraction and Interfaces http://tf.isti.cnr.it Paper Deadline 06/25/2012