Upgrade to Pro — share decks privately, control downloads, hide ads and more …

A Semantic Search Approach to Task-Completion Engines

A Semantic Search Approach to Task-Completion Engines

Date: July 8, 2018
Venue: Ann Arbor, MI, USA. Doctoral Consortium at the 41st International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR '18)

Please cite, link to or credit this presentation when using it or part of it in your work.

#InformationRetrieval #IR #TaskBasedSearch #TaskCompletionEngines #SemanticSearch

Darío Garigliotti

July 08, 2018
Tweet

More Decks by Darío Garigliotti

Other Decks in Research

Transcript

  1. About me • I'm in the third year of my

    PhD at IAI, UiS, Norway • My advisor is Prof. Krisztian Balog • My work aims to understand: • which challenges in semantic search are favorable for supporting task-completion engines, • which methods prove effective to model these challenges, • and how to integrate them into task-based search.
  2. Semantic Search and beyond • More users, greater expectations: understanding

    the search query • Search engines are becoming answer engines • Multiple techniques for query semantics
  3. Semantic Search and beyond • More users, greater expectations: understanding

    the search query • Search engines are becoming answer engines • Multiple techniques for query semantics • "With great power comes great responsibility"
  4. Task completion engines • Underlying search goal is often a

    complex and knowledge-intensive task • For example, to plan a travel • How to get there? • Where to stay? • What to do? • Task completion would provide a set of useful properties • To extend strategies of semantic search to help users complete their tasks
  5. Challenges in Semantic Search for Task-Completion Engines Entity type information

    for entity retrieval Entity-oriented search intents
  6. Challenges in Semantic Search for Task-Completion Engines Entity type information

    for entity retrieval Entity-oriented search intents Query suggestions to support task-based search
  7. Entity types • A characteristic property of entities is that

    they are typed • Types are organized in hierarchies • (or taxonomies) … Scientist … … … Person Agent … Enrico Fermi
  8. Query target types • Target types: types of entities sought

    by the query … Scientist Artist Writer … … … Person Agent … italian nobel prize winners
  9. Type-aware Entity Retrieval query entity Olympic games target types Rio

    de Janeiro term-based similarity type-based similarity … … entity types • Type information is known to improve entity retrieval • Unlike what it seems, it is a multifaceted problem
  10. Identifying and utilizing entity type information • How to utilize

    entity type information, with respect to dimensions as • the type taxonomy, • the type representation, • and the retrieval model? Entity type information for entity retrieval
  11. • Type taxonomy • Type representation • Retrieval model •

    We assume oracle-given type information Type-aware Entity Retrieval • We conduct an evaluation of dimensions in utilizing entity type information [1]
  12. • Type taxonomy • Wikipedia categories • Type representation •

    Retrieval model • We assume oracle-given type information Type-aware Entity Retrieval • We conduct an evaluation of dimensions in utilizing entity type information [1]
  13. • Type taxonomy • Wikipedia categories • Type representation •

    Most specific types • Retrieval model • We assume oracle-given type information Type-aware Entity Retrieval • We conduct an evaluation of dimensions in utilizing entity type information [1] t3 t3 t2 t2 t5 t5 t4 t4 t9 t9 t8 t8 e t6 t6 t12 t12 t7 t7 … t10 t10 t11 t11 t0 t0 t1 t1 …
  14. • Type taxonomy • Wikipedia categories • Type representation •

    Most specific types • Retrieval model • Interpolation • We assume oracle-given type information Type-aware Entity Retrieval • We conduct an evaluation of dimensions in utilizing entity type information [1] t3 t3 t2 t2 t5 t5 t4 t4 t9 t9 t8 t8 e t6 t6 t12 t12 t7 t7 … t10 t10 t11 t11 t0 t0 t1 t1 …
  15. Target Type Identification • How to automatically identify the target

    types for a query, from a given type taxonomy? • "We assume oracle-given type information"
  16. Target Type Identification • How to automatically identify the target

    types for a query, from a given type taxonomy? • We build a test collection for this task • We develop a Learning-to-Rank approach [2] • Our supervised learning method outperforms existing baselines by a large margin, and does consistently so across all query categories • "We assume oracle-given type information"
  17. Identifying and utilizing entity type information • We evaluated multiple

    dimensions of type information • We proposed an effective approach for type detection • There are benefits in the type-level representations Entity type information for entity retrieval
  18. Search intents and refiners • Intent: the underlying user need

    in a search query • For example, the intent of booking a hotel room • Refiner: a way to express an intent in an entity- oriented query • For example, for booking a hotel room: "booking", "book", "reservation", "rooms"
  19. Understanding and modeling search intents • A large proportion of

    entity- oriented search queries • What do those queries ask for, and how can they be better fulfilled? • How can we model search intents in a structured way? Entity-oriented search intents
  20. Towards an understanding of search intents • We define a

    scheme of intent categories [3] • Website, Service, Property, Other messi instagram => Website lebron james net worth => Property michigan league taxi => Service
  21. Towards an understanding of search intents • We define a

    scheme of intent categories [3] • Website, Service, Property, Other messi instagram => Website lebron james net worth => Property michigan league taxi => Service Property: 28.6% Service: 54.06% Website: 5.34% Other: 12.08%
  22. A Knowledge Base of entity- oriented search intents 1. Intents

    searched for a type of entities paris map, sydney map => [city] map 2. Categories assigned to refiners messi instagram => Website lebron james net worth => Property michigan league taxi => Service 3. Multiple refiners expressing an intent "booking", "book", "make a reservation", "rooms"
  23. 1. Intents searched for a type of entities paris map,

    sydney map => [city] map • (intent ID, searchedForType, entity type, confidence) 2. Categories assigned to refiners messi instagram => Website lebron james net worth => Property michigan league taxi => Service 3. Multiple refiners expressing an intent "booking", "book", "make a reservation", "rooms" A Knowledge Base of entity- oriented search intents
  24. 1. Intents searched for a type of entities paris map,

    sydney map => [city] map • (intent ID, searchedForType, entity type, confidence) 2. Categories assigned to refiners messi instagram => Website lebron james net worth => Property michigan league taxi => Service • (intent ID, ofCategory, intent category, confidence) 3. Multiple refiners expressing an intent "booking", "book", "make a reservation", "rooms" A Knowledge Base of entity- oriented search intents
  25. 1. Intents searched for a type of entities paris map,

    sydney map => [city] map • (intent ID, searchedForType, entity type, confidence) 2. Categories assigned to refiners messi instagram => Website lebron james net worth => Property michigan league taxi => Service • (intent ID, ofCategory, intent category, confidence) 3. Multiple refiners expressing an intent "booking", "book", "make a reservation", "rooms" • (intent ID, expressedBy, refiner, confidence) A Knowledge Base of entity- oriented search intents
  26. Our pipeline approach Refiners acquisition Refiners categorization Intents discovery [hotel]

    airport [hotel] spa [hotel] booking ... [hotel] airport: Service [hotel] address: Property [hotel] expedia: Website ... taxi arrive Hotel_Arriving booking make a reservation Hotel_Booking address Hotel_Address KB construction Intent ID Predicate Object Confidence Hotel_Booking searchedForType [hotel] c1 Hotel_Booking ofCategory Service c2 Hotel_Booking expressedBy "booking" c3 Hotel_Booking expressedBy "make a reservation" c4 Hotel_Booking expressedBy "rooms" c5
  27. Our pipeline approach Refiners acquisition Refiners categorization Intents discovery [hotel]

    airport [hotel] spa [hotel] booking ... [hotel] airport: Service [hotel] address: Property [hotel] expedia: Website ... taxi arrive Hotel_Arriving booking make a reservation Hotel_Booking address Hotel_Address Intent profile { KB construction Intent ID Predicate Object Confidence Hotel_Booking searchedForType [hotel] c1 Hotel_Booking ofCategory Service c2 Hotel_Booking expressedBy "booking" c3 Hotel_Booking expressedBy "make a reservation" c4 Hotel_Booking expressedBy "rooms" c5
  28. Knowledge base construction - Application of the pipeline to extract

    all quadruples from 581 unseen types - 155K quadruples, 31K intent profiles Excerpt of the KB, for intent ID <aviation.airline-65-customer_service>
  29. Understanding and modeling search intents • We proposed a scheme

    of intent categories • We built a high-quality knowledge base • There is a large proportion of service-oriented intents Entity-oriented search intents
  30. Task-based search Cheap wedding cake Make your own invitations Buy

    a used wedding gown Excerpt from TREC Tasks test dataset low wedding budget 1 low budget wedding dresses 0 low wedding budget cars 1 find a gown ... 0 wedding flowers 1 cup cake wedding 1 wedding cakes ... 2 wedding invitation 1 find wedding invitation templates 0 designer dresses wedding ...
  31. Task-based search Cheap wedding cake Make your own invitations Buy

    a used wedding gown Excerpt from TREC Tasks test dataset } low wedding budget 1 low budget wedding dresses 0 low wedding budget cars 1 find a gown ... 0 wedding flowers 1 cup cake wedding 1 wedding cakes ... 2 wedding invitation 1 find wedding invitation templates 0 designer dresses wedding ...
  32. Task-based search Cheap wedding cake Make your own invitations Buy

    a used wedding gown Excerpt from TREC Tasks test dataset } } low wedding budget 1 low budget wedding dresses 0 low wedding budget cars 1 find a gown ... 0 wedding flowers 1 cup cake wedding 1 wedding cakes ... 2 wedding invitation 1 find wedding invitation templates 0 designer dresses wedding ...
  33. Task-based search Cheap wedding cake Make your own invitations Buy

    a used wedding gown Excerpt from TREC Tasks test dataset } } } low wedding budget 1 low budget wedding dresses 0 low wedding budget cars 1 find a gown ... 0 wedding flowers 1 cup cake wedding 1 wedding cakes ... 2 wedding invitation 1 find wedding invitation templates 0 designer dresses wedding ...
  34. Supporting task-based search with query suggestions • How to generate

    query suggestions to support task-based search? • Can existing methods generate high-quality query suggestions? Query suggestions to support task-based search
  35. Query suggestions for task-based search choose bathroom cabinets lightning choose

    bathroom decoration style bathroom get ideas renew floor bathroom changing furniture bathroom choose bathroom choose bathroom • Given an initial query, to get a ranked list of query suggestions that cover all the possible subtasks related to the task the user is trying to achieve.
  36. • Components: • Source importance • We propose a generative

    probabilistic model [4] • We conduct a principled estimation of each of its components Query suggestions for task-based search API SUGGS. WEB SNIPPETS WEB DOCS. WH q0
  37. • Components: • Source importance • Document importance • We

    propose a generative probabilistic model [4] • We conduct a principled estimation of each of its components Query suggestions for task-based search API SUGGS. WEB SNIPPETS WEB DOCS. WH q0 API SUGGS. WEB SNIPPETS WEB DOCS. WH q0
  38. • Components: • Source importance • Document importance • Keyphrase

    relevance • We propose a generative probabilistic model [4] • We conduct a principled estimation of each of its components Query suggestions for task-based search API SUGGS. WEB SNIPPETS WEB DOCS. WH q0 API SUGGS. WEB SNIPPETS WEB DOCS. WH q0 Keyphrases
  39. • Components: • Source importance • Document importance • Keyphrase

    relevance • Query suggestion • We propose a generative probabilistic model [4] • We conduct a principled estimation of each of its components Query suggestions for task-based search API SUGGS. WEB SNIPPETS WEB DOCS. WH q0 API SUGGS. WEB SNIPPETS WEB DOCS. WH Query suggestions q0 Keyphrases
  40. Query suggestions for task-based search Search living in india cost

    of living in india american expats in india indian classical music india tourism India Live TV Search choose bathroom choose bathroom brass choose bathroom cabinets choose bathroom colors choose bathroom warmers choose bathroom lighting (a) query completion (b) query refinement • How to jointly generate query suggestions in query completion and refinement modes? • Which are the most useful information sources?
  41. Supporting task-based search with query suggestions Query suggestions to support

    task-based search • We proposed a query generation approach • We studied best combinations of sources and modes • The different methods generate unique candidates
  42. Future work • Using target types automatically detected, and dealing

    with missing type information • Providing actionable responses, to fulfill the variety of categories of entity-oriented search intents
  43. Future work • Using target types automatically detected, and dealing

    with missing type information • Providing actionable responses, to fulfill the variety of categories of entity-oriented search intents • Exploiting search intents to generate query suggestions for supporting task-based search
  44. Thank you! Darío Garigliotti [email protected] @DGarigliotti References: [1] Garigliotti, Darío

    and Balog, Krisztian. On Type-Aware Entity Retrieval. 2017. In: Procs. of ICTIR. [2] Garigliotti, Darío and Hasibi, Faegheh and Balog, Krisztian. Target Type Identification for Entity- Bearing Queries. 2017. In: Procs. of SIGIR. [3] Garigliotti, Darío and Balog, Krisztian. Towards an Understanding of Entity-Oriented Search Intents. 2018. In: Procs. of ECIR. [4] Garigliotti, Darío and Balog, Krisztian. Generating Query Suggestions to Support Task-Based Search. 2017. In: Procs. of SIGIR.
  45. Our pipeline approach Refiners acquisition Refiners categorization [hotel] airport [hotel]

    spa [hotel] booking ... [hotel] airport: Service [hotel] address: Property [hotel] expedia: Website ...
  46. Our pipeline approach Refiners acquisition Refiners categorization Intents discovery [hotel]

    airport [hotel] spa [hotel] booking ... [hotel] airport: Service [hotel] address: Property [hotel] expedia: Website ...
  47. Our pipeline approach Refiners acquisition Refiners categorization Intents discovery [hotel]

    airport [hotel] spa [hotel] booking ... [hotel] airport: Service [hotel] address: Property [hotel] expedia: Website ... booking make a reservation Hotel_Booking
  48. Our pipeline approach Refiners acquisition Refiners categorization Intents discovery [hotel]

    airport [hotel] spa [hotel] booking ... [hotel] airport: Service [hotel] address: Property [hotel] expedia: Website ... taxi arrive Hotel_Arriving booking make a reservation Hotel_Booking
  49. Our pipeline approach Refiners acquisition Refiners categorization Intents discovery [hotel]

    airport [hotel] spa [hotel] booking ... [hotel] airport: Service [hotel] address: Property [hotel] expedia: Website ... taxi arrive Hotel_Arriving booking make a reservation Hotel_Booking address Hotel_Address
  50. Our pipeline approach Refiners acquisition Refiners categorization Intents discovery [hotel]

    airport [hotel] spa [hotel] booking ... [hotel] airport: Service [hotel] address: Property [hotel] expedia: Website ... taxi arrive Hotel_Arriving booking make a reservation Hotel_Booking address Hotel_Address KB construction
  51. Our pipeline approach Refiners acquisition Refiners categorization Intents discovery [hotel]

    airport [hotel] spa [hotel] booking ... [hotel] airport: Service [hotel] address: Property [hotel] expedia: Website ... taxi arrive Hotel_Arriving booking make a reservation Hotel_Booking address Hotel_Address KB construction Intent ID Predicate Object Confidence Hotel_Booking searchedForType [hotel] c1 Hotel_Booking ofCategory Service c2 Hotel_Booking expressedBy "booking" c3 Hotel_Booking expressedBy "make a reservation" c4 Hotel_Booking expressedBy "rooms" c5
  52. Our pipeline approach Refiners acquisition Refiners categorization Intents discovery [hotel]

    airport [hotel] spa [hotel] booking ... [hotel] airport: Service [hotel] address: Property [hotel] expedia: Website ... taxi arrive Hotel_Arriving booking make a reservation Hotel_Booking address Hotel_Address Intent profile { KB construction Intent ID Predicate Object Confidence Hotel_Booking searchedForType [hotel] c1 Hotel_Booking ofCategory Service c2 Hotel_Booking expressedBy "booking" c3 Hotel_Booking expressedBy "make a reservation" c4 Hotel_Booking expressedBy "rooms" c5