WDAqua introduction presentation
0 likes485 views
(1) Handling the dynamic and temporal aspects of web data is challenging as facts may only be valid during certain time periods and new facts can contradict or modify existing facts. (2) Embedding models can help evaluate new facts by modeling relationships between entities as translations in vector space and assessing if new facts have support. (3) Reasoning over multiple supporting facts is needed to answer questions involving temporal and counting logic.
WDAqua introduction presentation
- 1. Handling Dynamicity and Temporality of Web Data Hady Elsahar hadyelsahar@gmail.com Jean Monnet University Saint-Étienne, France
- 2. First try with Question Answering Weet it : Natural language interface for Linked Data (ElSahar et al. ‘11 )
- 3. ● Most of the current knowledge bases focus on static facts and ignore the temporal dimension of facts. ● Aspects of temporality and Dynamicity of Datasets : ○ Aspect 1 : Many facts are valid only during a particular time period. ○ Aspect 2 : New extracted facts can contradict with, verify or modify new ones ○ Aspect 3 : Some Facts are collectively induced from a series of Events Handling Dynamicity of Data
- 4. Challenges and Motivations (1) : Stephen Hawking Many facts are valid only during a particular time period. Use Case : Questions about Temporal facts ● Who is first Wife of Stephen Hawiking ? ● Who is the 10th President of France ? ● Who is the past CEO of google ?
- 5. Extraction and Represenation of Temporal data Extraction and representation of Temporal Facts and Events ❏ Representation : ❏ Keeping the last updated fact is not enough (DBpedia) ❏ Higher order fact (Erdal and Weikum ‘11) ❏ f1:Bill_Clinton isPresidentOf USA. ❏ f2:f1 startedOnDate 20-01-1993 ❏ Wikidata Qualifiers (Vrandečić ‘12) ❏ Temporal fact and event extraction: ❏ Free Text and structured data from wikipedia (patterns and pattern induction) (Erdal and Weikum ‘11)
- 6. Annotation of temporal facts in documents for Question answering SemEval-2015 Task 5: QA TempEval
- 7. SemEval-2015 Task 5: QA TempEval Question Examples in the Evaluation Dataset : Yes / No: ● “Did the the Indonesian stock market rise again after it’s last fall ? List: ● “What happened after the crash?” ● “What happened between the crash and yesterday?” When (Factoid): ● “When did the Oscar ceremony end yesterday ?” Applications ?
- 8. Challenges and Motivations (2) : Stephen Hawking In Highly dynamic datasets, new extracted facts can contradict with, verify or modify new ones. Existing facts New Extracted Fact Matt Smith is dbo:starring of ■ dbr:Womb_(film) ■ dbr:Lost_River_(film) ■ dbr:Bert_and_Dickie ■ dbr:The_Science_of_Doctor_Who “Matt Smith is the doctor” (Matt Smith, occupation, Medicine) confidence : 0.1
- 9. (Frank Sinatra, profession, Singer) confidence : 0.9 (Jared leto, influenced_by, Frank Sinatra) confidence : 0.8 ● People influenced by Writers are probably writers as well ● people are probably born at the same place of their siblings Challenges and Motivations (2) : Stephen Hawking In Highly dynamic datasets, new extracted facts can contradict with, verify or modify new ones.
- 10. Evaluation of new facts using Link prediction Link Prediction ● Add new facts without extra knowledge ● Assess the validity of an unknown fact
- 11. Embedding Models for knowledge bases TransE : Modeling Relations as Translations (Bordes et al. ’13): ● Modeling Facts as translations between vectors of entities VSubject + VRelation ≅ VObject ● distance is used to Quantify confidence in facts ● Training objective: Find the representations that Minimizes distances across all true facts and maximize across “corrupted” facts ( s’ , o’ ):
- 12. Other Embedding Models: ● Structured Embeddings (SE) (Bordes et al ‘11 ) ● Collective Matrix Factorization (RESCAL) (Nickel et al., ’11) ● Neural Tensor Networks (socher et al. ‘13) ● TATEC (Garcia-Duran et al., ’14) Embedding Models for Text + Knowledge bases: ● Joint Learning of Words and Meaning Representations (Bordes et al. ‘12) ● Knowledge Graph and Text Jointly Embedding (Wang et al ‘14) Link prediction using Embedding Models
- 13. Applications ? ● Verification of new Extracted Facts ● Completeness of new added datasets ● Modeing literals dataypes (length, date ..etc ) not only relations and entities. Embedding Models other benefits ? (collaboration potential) ● Entity Disambiguation for Fact Extraction and QA (Bordes et al. ‘12) ● Paraphrase Detection for Questions, (PARALEX) (Fader et al. ‘13)
- 14. Challenges and Motivations (3) : Reasoning with more than one supporting facts ● Reasoning about positions (ex: Geo Data) ● Reasoning about Counts ● Reasoning about sizesFact 1 : 55 passengers crammed into the smuggler’s boat. Fact 2 : The boat made it to the Greek island. Question : Where are the passengers ? Stephen Hawking Facts induced from a series of Events ● Towards AI-Complete QA: A Set of Prerequisite Toy Tasks (Wetson et al ‘15) ● Memory Networks (Wetson et al ‘14)