Added additional information about the dataset. (#1)

- Added additional information about the dataset. (9c31147fd9e88db89d42ff7fda7b7675e44df6c4)


Co-authored-by: Taja Kuzman <[email protected]>

README.md

@@ -67,8 +67,30 @@ size_categories:
 
 ### Dataset Summary
 
-The Slovene Web genre identification corpus GINCO 1.0 contains 10002 web texts, manually annotated with genre. The corpus allows for automated genre identification and genre analyses as well as other web corpora research.
-The dataset is split into 602 training, 200 validation, and 200 test texts by the original authors.  
+The Slovene Web genre identification corpus GINCO 1.0 contains 1,002 web texts (478,969 words), manually annotated with genres.
+The corpus allows for automated genre identification and genre analyses as well as other web corpora research.
+
+This dataset was extracted from the manually-annotated subcorpus (GINCO-1.0-suitable.json.zip) from the original GINCO dataset,
+published on the [CLARIN.SI repository](http://hdl.handle.net/11356/1467).
+
+The dataset is split into 602 training, 200 validation, and 200 test texts by the original authors.
+
+The texts in the suitable subset are annotated with up to three genre categories,
+where the primary label is the most prevalent, and secondary and tertiary labels denote presence of additional genre(s).
+The secondary and tertiary labels are available for multilabel classification,
+while for most use cases, we suggest that only the primary label is used.
+
+The labels are provided in three levels of detail (three category sets), allowing experiments with the full set (24 labels),
+set of 21 labels (labels with less than 5 instances are merged with label Other) and set of 12 labels (similar labels are merged).
+For most use cases, we suggest that the smallest set -- set of 12 labels is used (`primary_level_3`),
+and that the category "Other" is regarded as a "throw-away" category to detect texts
+for which the classifier could not predict any of the concrete labels.
+
+Additionally, the corpus contains some metadata about the text (e.g. url, domain, year)
+and its paragraphs (e.g. near-duplicates and their usefulness for the genre identification).
+
+More details on dataset construction, manual annotation and results of machine learning experiments
+are provided in the paper ["The GINCO Training Dataset for Web Genre Identification of Documents Out in the Wild"](https://aclanthology.org/2022.lrec-1.170/) (Kuzman et al., 2022).
 
 ### Languages
 
@@ -121,15 +143,15 @@ A sample instance from the dataset:
   - 'text': text of the paragraph;
   - 'duplicate': true if the text is a near-duplicate;
   - 'keep': true, if the text is useful for the genre identification, false, if not;
-- 'primary_level_1': first genre category, level of detail 1;
-- 'primary_level_2': first genre category, level of detail 2;
-- 'primary_level_3': first genre category, level of detail 3;
-- 'secondary_level_1': second genre category, level of detail 1;
-- 'secondary_level_2': second genre category, level of detail 2;
-- 'secondary_level_3': second genre category, level of detail 3;
-- 'tertiary_level_1': third genre category, level of detail 1;
-- 'tertiary_level_2': third genre category, level of detail 2;
-- 'tertiary_level_3': third genre category, level of detail 3;
+- 'primary_level_1': first genre category, most detailed category set;
+- 'primary_level_2': first genre category, category set where too infrequent categories are merged to Other;
+- 'primary_level_3': first genre category, compact and most useful category set;
+- 'secondary_level_1': second genre category, most detailed category set;
+- 'secondary_level_2': second genre category, category set where too infrequent categories are merged to Other;
+- 'secondary_level_3': second genre category, compact and most useful category set;
+- 'tertiary_level_1': third genre category, most detailed category set;
+- 'tertiary_level_2': third genre category, category set where too infrequent categories are merged to Other;
+- 'tertiary_level_3': third genre category, compact and most useful category set;
 - 'split': example can belong to the 'train', 'dev', or 'test' split;
 - 'domain': domain address of the website where the text originates from.
 
@@ -160,6 +182,8 @@ Texts are annotated using 24 genre categories:
 - List of Summaries/Excerpts,
 - Other.
 
+See the [Appendix in the paper](https://aclanthology.org/2022.lrec-1.170/) for descriptions of the labels.
+
 ## Additional Information
 
 ### Dataset Curators
@@ -172,6 +196,8 @@ CC BY-SA 4.0
 
 ### Citation Information
 
+To cite the dataset:
+
 ```
 @misc{11356/1467,
  title = {Slovene Web genre identification corpus {GINCO} 1.0},
@@ -184,6 +210,17 @@ CC BY-SA 4.0
 }
 ```
 
+To cite the paper on the dataset construction and manual annotation:
+```
+@inproceedings{kuzman2022ginco,
+  title={The GINCO Training Dataset for Web Genre Identification of Documents Out in the Wild},
+  author={Kuzman, Taja and Rupnik, Peter and Ljube{\v{s}}i{\'c}, Nikola},
+  booktitle={Proceedings of the Thirteenth Language Resources and Evaluation Conference},
+  pages={1584--1594},
+  year={2022}
+}
+```
+
 ### Contributions
 
-Thanks to Hana Skitek for adding this dataset.
+Thanks to Hana Skitek for adding this dataset, and Taja Kuzman for extending the readme with additional information on the dataset.