NL Reference Corpus

Revision as of 11:05, 22 February 2013

The NL Reference Corpus (NC) is the corpus used to prepare and to assess grammars for sentence-based UNLization. It is divided in 6 different levels according to the Framework of Reference for UNL (FoR-UNL):

• NC-A1: NL Reference Corpus A1
• NC-A2: NL Reference Corpus A2
• NC-B1: NL Reference Corpus B1
• NC-B2: NL Reference Corpus B2
• NC-C1: NL Reference Corpus C1
• NC-C2: NL Reference Corpus C2

Methodology

As a natural language corpus, the NC varies for each language. It is derived from a base corpus to be compiled and processed according to the following criteria:

1. The Base Corpus must have at least 5,000,000 tokens (strings isolated by blank space and other word boundary markers). It must be representative of the contemporary standard use of the written language, and should include documents from as many different genres and domains as possible.
2. The Base Corpus must be segmented (in sentences) and tagged for POS.
3. The segmented corpus is used to calculate the average sentence length (ASL), which is the median of the length (in words) of all sentences.
4. The tagged corpus is used to extract the syntactic surface structures (SSS), which are sequences of POS, and to calculate their frequency of occurrence.
5. The average sentence length (ASL) and the syntactic surface structures (SSS) are used to generate the NC templates, as follows:
   • NC-A1 = 500 most frequent SSS's where length < (ASL/2) (500 most frequent shortest syntactic structures)
   • NC-A2 = 1,000 most frequent SSS's where length < (ASL/2) (1,000 most frequent shortest syntactic structures)
   • NC-B1 = 2,000 most frequent SSS's where length < ASL (2,000 most frequent short syntactic structures)
   • NC-B2 = 3,000 most frequent SSS's where length < ASL (3,000 most frequent short syntactic structures)
   • NC-C1 = 4,000 most frequent SSS's
   • NC-C2 = 5,000 most frequent SSS's
6. The NC templates are used to compile the NC corpora: the training corpora and the testing corpora. The training corpora consists of 1 exemplar of each SSS, and will be used to prepare the grammar. The testing corpora consists of 4 exemplars of each SSS randomly selected in the Base Corpus. The whole NC corpora (i.e., 5 exemplars for each SSS) is used to calculate the F-measure, which is the parameter for assessing the precision and the recall of the grammars.

Files

• Arabic
  • Source: Wikipedia
  • Total number of distinct sentences: 801,258
  • ASL = 16
  • Raw files (with POS tags)
    • Corpus NC_ara_A (sentences < ASL/2): 121,747 (9 MB)
    • Corpus NC_ara_B (ASL/2 <= sentences ASL): 256,352 (45 MB)
    • Corpus NC_ara_C (sentences >= ASL): 423,159