2006 CombiningLinguisticExtractRelations

• [Suchanek et al, 2006] => F. M. Suchanek, G. Ifrim and G. Weikum. (2006). Combining Linguistic and Statistical Analysis to Extract Relations from Web Documents. In Proc. of KDD-2006.

---

Keywords: Relation Recognition from Text Algorithm, Dependency Grammar-based Relation Recognition Algorithm, LEILA

---

Summary

• Presents a Supervised Relation Recognition Algorithm for binary Semantic Relations.
• The patterns are undirected link structures generated from deep syntactic analysis (Link Grammar) where the two ends of the links are the two entities sought.
• The algorithm uses a machine learning approach to discriminate which patterns to retain.
  • Technical Report: http://domino.mpi-sb.mpg.de/internet/reports.nsf/NumberView/2006-5-004

Contributions

1. Applies [[Syntactic_Parsing?]] information to Supervised Relation Recognition with high F-score.
2. First known application of [[kNN?]] and [[SVM?]] to Pattern selection. Past approaches use Heuristics or [[.
3. Proposes a [[Text_Graph?]] representation that is expressive and robust.
4. The patterns reported in the empirical studies are interesting and reasonable. (see Comments)
5. Achieves good empirical results on several datasets.
6. Their system is available to test.

---

Input/Output

Input

• positive examples
• (relationship direction? one-to-one, one-to-many, many-to-one?)
• annotated data (POS, Link, NER(e.g. person,date,...))

Output

• relation instances (& duplicates)
• and their associated patterns

---

Patterns

• Uses [[Syntactic_Patterns?]] based on the Dependency Grammar-like Link Grammar.
• Nouns and Adjectives are only matched for Part-of-Speech Role confirm.
• Bridge: the (unique) shortest path from one Entity

   placeholder to the other

• A "pattern matches a linkage if the bridge of the pattern appears in the linkage"

Questions, Patterns

• How are the boundaries of multi-word entities detected?

---

Algorithm

1. Discovery Phase (Train Phase)

• patterns are constructed from all sentences with both entities.

2. Assessment Phase (Train Phase)

• negative examples (counterexamples) are gathered
• negative patterns are constructed
• discriminative model induced

3. Harvesting Phase (Test Phase)

Questions, Algorithm

• How does the counterexample selection know which entity is required to be fixed. E.g. in "Chopin"/"1957", would the algorithm know that the relationship is one-to-many and that the 'many' is associated with the year?

---

General Questions

• What is the relationship of their approach with [Bunescu and Mooney, 2005] and with [Chiang and Yu, 2005]?
  • E.g. what is the difference in their definition of "shortest-path"?
• It would be interesting to see all competitors evaluated.
• The patterns reported in the empirical studies do not demonstrate the benefit of shortest path or of replacement nouns and adjectives. For example, the pattern used in figure 3 "<X> was <ADJECTIVE> among <Y>", is not one of the patterns reported in 4.2.1.
• Why the discrepancy between the Snowball F-score performance of ~30% they report and the 80%+ figures reported in [Agichtein and Gravano, 2000]. This especially disconcerting given that the main pattern reported is the same one that Snowball (and DIPRE) have been shown to discover: "Y-based X".
• The statements around the contribution of anaphora resolution are unconvincing. Clearly Snowball would also benefit directly from accurate tagging of anaphoras as entities.
• There is no support for text that occurs before or after the entities, as in Snowball. This is a minor issue given that the value of prefiller and postfiller patterns.
• The size of the test sets are very small; in the hundreds.
• The size of the train sets is not reported

---

BibTeX

@InProceedings{conf/kdd/SuchanekIW06,
  title =	"Combining linguistic and statistical analysis to
		 extract relations from web documents",
  author =	"Fabian M. Suchanek and Georgiana Ifrim and Gerhard
		 Weikum",
  bibdate =	"2006-10-05",
  bibsource =	"DBLP,
		 http://dblp.uni-trier.de/db/conf/kdd/kdd2006.html#SuchanekIW06",
  booktitle =	"KDD",
  booktitle =	"Proceedings of the Twelfth {ACM} {SIGKDD}
		 International Conference on Knowledge Discovery and
		 Data Mining, Philadelphia, {PA}, {USA}, August 20-23,
		 2006",
  publisher =	"ACM",
  year = 	"2006",
  editor =	"Tina Eliassi-Rad and Lyle H. Ungar and Mark Craven and
		 Dimitrios Gunopulos",
  ISBN = 	"1-59593-339-5",
  pages =	"712--717",
  URL =  	"http://doi.acm.org/10.1145/1150402.1150492",}