In the garden and in the jungle:
comparing genres in the BNC and in the Internet

Serge Sharoff

The jungle metaphor is quite common in genre studies. The subtitle of David Lee's seminal paper on genre classification is `navigating a path through the BNC jungle'. According to Adam Kilgarriff, the BNC is a jungle only when compared to smaller Brown-type corpora, while it looks more like an English garden when compared to the Web . A corpus from the web can easily surpass the BNC in size, see 160 million words of I EN or 2 billion words of ukWac (http://wacky.sslmit.unibo.it/).

However, we know little about the domains and genres of texts in corpora collected in this way. This webpage lists resources that make I-EN, I-RU (a web corpus of Russian) and ukWac a bit similar to the BNC.

Below I report two ways of approaching the question of genre classification. One involves a traditional typology of typical genre labels, as applicable to the Web, for example, texts aimed at instructing, reporting or entertaining the reader. Another approach involves designing a topology to assess how similar individual texts are to a prototypical webpage, for example, a typical news item is aimed at reporting, but some of them also aim at entertaining, so that such texts are positioned between reporting and entertaining texts.

Typology for text classification

The typology and the procedure for automatic annotation of Web texts is described in:

Serge Sharoff, In the garden and in the jungle: comparing genres in the BNC and Internet. In Genres on the Web, Mehler, A., Sharoff., S., Santini, M., (editors) Springer 2010. PDF

According to this approach, the texts in I-EN, I-RU and ukWac have been automatically classified using the following classes:

  1. discussion - all texts expressing positions and discussing a state of affairs
  2. information - catalogues, lists (mostly containing incomplete sentences)
  3. instruction - how-tos, FAQs, tutorials
  4. propaganda - adverts, political pamphlets
  5. recreation - fiction and popular lore
  6. regulations - laws, small print, rules
  7. reporting - newswires and informative broadcasts, police reports

Source data

Validated samples

Classified files

The accuracy of this classification is about 73-84% (see the paper above for argumentation), so you have one chance in four that a link is not of the correct type. Let me know if you have ideas on how to improve the accuracy.

Legacy classes

I-EN and ukWac files have been also classified using David Lee's BNC classification (70 genres in total) and the four main genres from the Brown corpus (press, fiction, nonfiction and misc):

The accuracy of this classification has not been validated. Presumably it is quite low (especially for the 70-genres classification from the BNC). I made a quick check for the genre distribution for 8310 pages from The Guardian website, which is a newspaper, so it should be classified as 'press' according to the Brown Corpus, but the genre of feature articles, biographies, reviews can be different from what is assumed by `press' in the Brown Corpus (it corresponds to 'reporting' in the classification used above):

10.01% fiction
29.07% misc
16.68% nonfiction
44.24% press

The following is the distribution of genres assigned to the same set of 8310 pages according to the BNC-trained classifier (only the 10 most frequent labels are listed):

3.14% W_newsp_other_social
3.21% W_newsp_brdsht_nat_editorial
3.29% S_speech_unscripted
3.35% W_newsp_brdsht_nat_commerce
3.61% W_newsp_brdsht_nat_sports
4.16% W_fict_prose
5.57% W_pop_lore
5.93% W_newsp_brdsht_nat_arts
6.45% W_biography
8.19% W_newsp_brdsht_nat_misc
11.01% W_misc

Not all items are treated as coming from newspapers, but many of them are (in the BNC genre scheme, brdsht_nat means `national broadsheets', newsp_other means either regional or tabloid). Webpages automatically classified as all forms of W_newsp account for 41% of The Guardian subcorpus in ukWac.

Topology for text classification

The topological approach is described in the currently unpublished draft of my paper:

Serge Sharoff, Functional Text Dimensions for annotation of Web corpora. PDF

The resources consist in multi-annotated webpages for Russian and English (along with translations of some pages into Chinese, French and German) as described in the following table:

Corpus Language #Docs #Words Annotations
5g, part1 en,de,fr,ru,zh 113 306302 5g-p1.tgz
5g, part 2 en,fr,ru,zh 133 505468 5g-p2.tgz
ukWac, random en 257 211549 ukwac-sample.tgz
GICR, random ru 618 919972 gicr-random.tgz
GICR, blogs ru 285 83829 gicr-blogs.tgz
Total   1406 2027120

 

About

The resources listed on this page have been developed by Serge Sharoff (Centre for Translation Studies, University of Leeds). Get in touch with me if you have any comments or suggestions.

Note: for files from the `Genres on the Web' colloquium (2007), see the original colloquium page

Note: for the description of a Google Research Award project, see the project webpage



Serge Sharoff 2015-12-20