Mixed Initiative Search - Prof. Dr. Maarten de Rijke

Post on 07-Jan-2017

234 views 2 download

Transcript of Mixed Initiative Search - Prof. Dr. Maarten de Rijke

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partnerspartners

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

Mixed initiative searchMaarten de Rijke

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

www.amsterdamdatascience.nlMixed initiative search

• Based on joint work with David Graus, Evangelos Kanoulas, Edgar Meij, Daan Odijk, Ridho Reinanda, Manos Tsagkias, Christophe Van Gysel, Nikos Voskarides, Wouter Weerkamp, Marcel Worring, Masrour Zoghi

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

www.amsterdamdatascience.nl

It’s all about entities

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

www.amsterdamdatascience.nlMixed initiative search

• Entities (people, locations, organizations, …) play central organizing role

• In search

• in web search, up to 70% of the queries are entity queries (Lin et al., 2012; Guo et al., 2011)

• in academic search, the proportion of queries that contain entities is over 93% (Li et al., 2016)

• increasingly, entities are retrievable items

It’s all about entities

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

www.amsterdamdatascience.nlMixed initiative search

search

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

www.amsterdamdatascience.nlMixed initiative search

search

mothers

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

www.amsterdamdatascience.nlMixed initiative search

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

www.amsterdamdatascience.nlMixed initiative search

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

www.amsterdamdatascience.nlMixed initiative search

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

www.amsterdamdatascience.nlMixed initiative search

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

www.amsterdamdatascience.nlMixed initiative search

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

www.amsterdamdatascience.nlMixed initiative search

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

www.amsterdamdatascience.nlMixed initiative search

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

www.amsterdamdatascience.nlMixed initiative search

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

www.amsterdamdatascience.nlMixed initiative search

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

www.amsterdamdatascience.nl

Where are we now?

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

www.amsterdamdatascience.nlMixed initiative search

• Diverse intents

• You know an entity by the stuff it hangs out with

• Words

• Facets

• Entities

• Organizations relations

• …

Information needs around entities

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

www.amsterdamdatascience.nlMixed initiative search

• Mine search logs to discover aspects

• Prioritizing entity display

• Providing direct action

• “Query-less” entity-oriented diversification

• Supporting complex search tasks

• Knowledge base design/construction

Discovering entity aspects

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

www.amsterdamdatascience.nlMixed initiative search

• Automatically generate human-readable explanations of related entities

• Mine large volumes of text: fragments in which entities co-occur

Entity relations

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

www.amsterdamdatascience.nlMixed initiative search

• Update description of long tail entity with information from range of sources

• Learn to adjust representations based on clicks

• Enriching the representation with additional descriptions helps improve retrieval

• Continuously updating the ranker helps

Entity updates

KBER

DCER

sim

sim

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

www.amsterdamdatascience.nlMixed initiative search

• Unsupervised model construction, efficient entity capabilities query, and semantic matching between query terms and candidate entities

• Learn mappings between words and entities, as well as distributed representations of words and entities

• Words that are strongly evidential for particular products are projected nearby those products

• Very effective, highly scalable

Unsupervised entity ranking

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

www.amsterdamdatascience.nl

What is next?

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

www.amsterdamdatascience.nlMixed initiative search

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

www.amsterdamdatascience.nlMixed initiative search

document listaction at

querystate st

userenvironment

examine document list

generate implicit feedback

reward rt

implicit feedback

evaluation measureretrieval system

agentagent

action at

environment

reward rt

state st

Image taken from K. Hofmann, S. Whiteson, and M. de Rijke. Balancing exploration and exploitation in online learning to rank. In ECIR 2011, April 2011.

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

www.amsterdamdatascience.nlMixed initiative search

• Go beyond the traditional search engine result page

• What should the search engine say?

• When should it switch?

With all of that information around an entity

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

www.amsterdamdatascience.nlMixed initiative search

• Old-fashioned search engine result page (SERP)?

• Direct answer?

• Engage in a conversation?

• Generate a news article?

• Produce a timeline?

• Multi-document summary?

• Wikipedia page?

What should a search engine say?

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

www.amsterdamdatascience.nlMixed initiative search

• Working with media professionals in an audiovisual archive

• It is their business to create narratives

• Wrapping extensive interviewing rounds with media professionals

• Next

• Annotate and mine semantic aspects of narratives media professionals create

• Learning templates for SERP generation

• Learning to re-order elements on the SERP

• Generating natural language to connect the elements

Example: narrative search

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

www.amsterdamdatascience.nlMixed initiative search

Example: Engage in a conversation

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

www.amsterdamdatascience.nlMixed initiative search

• Run an A/B test

• Explore online

• Learn from historical data (“counterfactual reasoning”)

When should should it switch?

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

www.amsterdamdatascience.nlMixed initiative search

Would people buy it?

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

www.amsterdamdatascience.nlMixed initiative search

document listaction at

querystate st

userenvironment

examine document list

generate implicit feedback

reward rt

implicit feedback

evaluation measureretrieval system

agentagent

action at

environment

reward rt

state st

Based on K. Hofmann, S. Whiteson, and M. de Rijke. Balancing exploration and exploitation in online learning to rank. In ECIR 2011, April 2011.

answer format+

answers

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

This is the presentation titlethis subline can be used for authors

partners

www.amsterdamdatascience.nlMixed initiative search

All content represents the opinion of the author(s), which is not necessarily shared or endorsed by their employer and/or sponsors.