Presentation is loading. Please wait.

Presentation is loading. Please wait.

Terminology problems in literature mining and NLP

Similar presentations


Presentation on theme: "Terminology problems in literature mining and NLP"— Presentation transcript:

1 Terminology problems in literature mining and NLP
John MacMullen SILS Bioinformatics Journal Club Fall 2003

2 Assumptions of the paper
“knowledge encoded in textual documents is organized around sets of domain-specific terms, which are used as a basis for sophisticated knowledge acquisition.” [938] “Terms represent the most important concepts in a domain and characterize documents semantically.” [939] “the basic problem is to recognize domain-specific concepts and to extract instances of specific relationships among them.” [938] SILS Bioinformatics Journal Club – Fall 2003

3 Current approaches to auto term recognition
Morpho-syntactic feature identification Hybrid linguistic and statistical approaches Machine learning techniques Problems Terms are ambiguous and have variation; they are hardly ever mono-referential The lack of naming conventions (controlled vocabularies), the existence of acronyms, and the large existing heterogeneous literatures increase complexity. What are they really trying to do with these measures? Are they effective? SILS Bioinformatics Journal Club – Fall 2003

4 Context: Term variation problems in NLP
SILS Bioinformatics Journal Club – Fall 2003

5 Terminology Processing Workflow
2,082 MEDLINE abstracts related to ‘nuclear receptors’ What are the assumptions of this workflow? Nenadic, Spacsic & Ananiadou (2003), Fig 1 SILS Bioinformatics Journal Club – Fall 2003

6 SILS Bioinformatics Journal Club – Fall 2003
ATR approach C-values (“termhoods”) [940] Term frequency “Frequency of occurrence as a substring of other candidate terms” (receptor) “Number of candidate terms containing the given candidate term as a substring” “Number of words contained in the candidate term” NC-values (“termhood estimations”) [940] Includes context of candidate terms “Frequency of co-occurrence with top-ranked context words” NC-values = a linear combination of C-values and context factors for each term What are they really trying to do with these measures? Are they effective? SILS Bioinformatics Journal Club – Fall 2003

7 Clustering & Evaluation
CSL (contextual, syntactical, lexical) Clustering implies underlying perspectives or queries Evaluation Recall – the probability a relevant item will be retrieved Precision – the probability that a retrieved item will be relevant SILS Bioinformatics Journal Club – Fall 2003

8 SILS Bioinformatics Journal Club – Fall 2003
Other questions Corpus construction: “a larger corpus does not have a proportionally higher number of acronyms” [942] True? “All term variants are considered jointly for the calculation of termhood” [942] What would happen if they weren’t? In what ways is the hybrid similarity measure corpus dependent? [942] Is there bias in the sample due to the construction of the corpus? SILS Bioinformatics Journal Club – Fall 2003

9 SILS Bioinformatics Journal Club – Fall 2003
References Nenadic, G., Spasic, I., & Ananiadou, S. (2003). Terminology-driven mining of biomedical literature. Bioinformatics 19(8), SILS Bioinformatics Journal Club – Fall 2003


Download ppt "Terminology problems in literature mining and NLP"

Similar presentations


Ads by Google