Towards Efficient and Effective Semantic Table Interpretation

Towards Efficient and Effective Semantic Table Interpretation
Ziqi Zhang
Department of Computer Science, University of Sheffield

Outline
•Define semantic table interpretation
•State-of-the-art and motivation
•The method – TableMiner
•Evaluation
Z. Zhang / Towards Efficient and Effective Semantic Table Interpretation

Semantic Table Interpretation
•Input
• Ontology
• Relational table
•Goals/Tasks
• Label columns by concepts
• Link cells to named entities
• Connect columns by relations
Thing
Work
Artist
Location
… …
Ent:USA
Ent:UK
…
Film
Actor/ Actress
Country
Name
Film
Country
1
Tom Hanks
Philadelphia
USA
2
Jamie Foxx
Ray
USA
3
Kate Winslet
The Reader
UK
99
Charlize Theron
Monster
South Africa
Table of Best Actor/Actress
< … … >
… …
Rel:performIn
Rel:performIn

Semantic Table Interpretation
•Input
• Ontology
• Relational table
•Goals/Tasks
• Label columns by concepts
• Link cells to named entities
• Connect columns by relations
Column classification/ header disambiguation
Relation interpretation
Cell disambiguation

Motivation and State-of-the-art
•154 mil. relational tables on the Web and growing [Cafarella2008]
•Classic Information Extraction methods do not work [Limaye2010, Lu2013]
• They cannot model the complex interdependence among table components

•SoA semantic table interpretation methods, e.g. [Limaye2010, Venetis2011, Mulwad2013]
Limitation 1
Inference is ‘exhaustive’, but unnecessary
Name
Film
Country
1
Tom Hanks
Philadelphia
USA
2
Jamie Foxx
Ray
USA
3
Kate Winslet
The Reader
UK
99
Charlize Theron
Monster
South Africa
< … … >
Goal: Assign a concept to this column
Hint: Content in the column gives useful clues
How much do we need for inference (99 rows in this example)?
- Human: SOME (learn by example)
- SoA: ALL

•SoA semantic table interpretation methods, e.g. [Limaye2010, Venetis2011, Mulwad2013]
Limitation 2
Contextual features for inference
SoA: features only from within the table
Context outside the table also makes hint for interpretation. E.g., the words in the paragraph are often found in descriptions of actors

TableMiner
•Two tasks:
• Column classification
• Cell disambiguation
•Non-exhaustive inference in a bootstrapping pattern
• phase 1 – inference with partial content
• phase 2 – propagation and update
•Contextual features both inside and outside tables

TableMiner – Phase 1 I-Inf
•Incremental inference with stopping (I-Inf) Tj – a column; Cj – candidate concepts for the column; Ei,j candidate entities for a cell

Itr.1
….
(until stop)
Ei,j=
{<e1,s1>, <e2,s2>, …}

Itr.1
…. (until stop)
Ei,j= {<e1,s1>, <e2,s2>, …}
concepts = {<c1,s1>, <c2,s2>, …}
Cj=
{<c1,s1’>, <c2,s2‘>}

Itr.1
….
(until stop)
Ei,j=
{<e1,s1>, <e2,s2>, …}
concepts = {<c1,s1>, <c2,s2>, …}
Cj= {<c1,s1’>, <c2,s2‘>}
|H(Cj) – H(prevCj)|<t?
Yes – stop No – next itr.

…. (until stop)
concepts = {<c1,s1>, <c3,s3>, …}
Cj= {<c1,s1’>, <c2,s2‘>, <c3,s3‘>}
Ei,j= {<e1,s1>, <e2,s2>, …}
Itr.2

….
(until stop)
Itr.3
Ei,j= {<e1,s1>, <e2,s2>, …}
concepts = {<c11,s11>}
Cj= {<c1,s1’>, <c2,s2‘>, <c3,s3‘>, …. <c11,s11‘>}

•To compute scores of candidate named entities (e.g. <e1,s1>) and concepts (e.g., <c1,s1’>)
•Candidate NE
•Build a feature vector of a candidate using the ontology
•Build a feature vector of the cell/column header using its context
•Compute vector similarity
•Candidate concept: same principle, but also depends on score of contributing NEs

TableMiner – Phase 2 Propagate, Update
•When I-Inf stops
• Select the highest scoring candidate concept c+ to label the column
• Propagate: use c+ as constrain to disambiguate remaining cells – candidate NEs not belonging to c+ are discarded
• Update:
•Re-compute c+ after all cells are disambiguated
•If the new c+ is different, revise disambiguation across the entire column with it as new constraint
•Repeat until no change
Cj= {<c1,s1’>, <c2,s2‘>, <c3,s3‘>, …. <c11,s11‘>} c+
Rank and select
Use as constraint to disambiguate cells

Evaluation

TableMiner – Evaluation
•Data
• Freebase as reference ontology/background knowledge
• Limaye112 – 112 Web tables from Limaye2010 originally annotated with Wikipedia
•Cells are automatically mapped to Freebase – some are unmapped
•Columns are manually annotated
• IMDB – 7,354 “cast” tables of films mapped to Freebase

TableMiner – Evaluation
•Baselines (both uses exhaustive inference)
• Bfirst - cell disambiguation: choose the top ranked NE candidate in the Freebase search result - column classification: each disambiguated cell casts a vote to the set of concepts the NEs belong to, and the majority wins
• Bsim - cell disambiguation: string similarity + feature vector similarity (in-table context only) - column classification: the majority vote method as above + string similarity

TableMiner – Evaluation Results
•Cell disambiguation Manual validation of 932 cell annotations in Limaye112 not covered by the above results (i.e., unmapped cells)
If only consider those cells where at least one system predicts correctly

•Column classification best only – a column is labelled correctly only if the concept is suitable for the data in the column and is specific enough best or ok – a column is labelled correctly if the concept is suitable for the data in the column, though not very specific (E.g., ‘Film Actors’ may be the best, while ‘Artist’ or ‘Person’ is OK, but ‘Engineer’ is incorrect)

•Efficiency – TableMiner is efficient because
•Column classification: processes partial content from a column (avg. 57% Limaye112, 43% IMDB)
•Cell disambiguation: constrained by column classification, resulting in smaller NE candidate space (avg. 32% reduction Limaye32, 24% IMDB)
• Fewer candidates => less time spent on retrieval and feature space creation (typically >90% of CPU in the pipeline, Limaye2010)

TableMiner – Conclusion
•TableMiner take-home messages
•How can it be more effective?
• Use both context within and outside tables as features for inference
•How can it be more efficient?
• Perform inference with partial data and follow the bootstrapping pattern of learning
Message 1
Message 2

References
•[Cafarella2008] Cafarella, M.J., Halevy, A., Wang, D.Z., Wu, E., Zhang, Y. 2008: Webtables: exploring the power of tables on the web. Proceedings of VLDB Endowment 1(1), 538–549
•[Limaye2010] Limaye, G., Sarawagi, S., Chakrabarti, S. 2010: Annotating and searching web tables using entities, types and relationships. Proceedings of the VLDB Endowment 3(1-2), 1338–134
•[Lu2013] Lu, C., Bing, L., Lam, W., Chan, K., Gu, Y. 2013: Web entity detection for semi-structured text data records with unlabeled data. International Journal of Computational Linguistics and Applications
•[Mulwad2013] Mulwad, V., Finin, T., Joshi, A. 2013: Semantic message passing for generating linked data from tables. In: International Semantic Web Conference (1). pp. 363–378. Lecture Notes in Computer Science, Springer
•[Venetis2011] Venetis, P., Halevy, A., Madhavan, J., Pas¸ca, M., Shen,W.,Wu, F., Miao, G.,Wu, C. 2011: Recovering semantics of tables on the web. Proceedings of VLDB Endowment 4(9), 528–538

Thank you

Towards Efficient and Effective Semantic Table Interpretation

More Related Content

Similar to Towards Efficient and Effective Semantic Table Interpretation (20)

Recently uploaded (20)

Towards Efficient and Effective Semantic Table Interpretation