gsarti commited on
Commit
619d371
•
1 Parent(s): b4a2497

Updated with style transfer task

Browse files
Files changed (1) hide show
  1. README.md +19 -2
README.md CHANGED
@@ -31,6 +31,7 @@ paperswithcode_id: glue
31
  - [Dataset Description](#dataset-description)
32
  - [Dataset Summary](#dataset-summary)
33
  - [Supported Tasks and Leaderboards](#supported-tasks-and-leaderboards)
 
34
  - [Languages](#languages)
35
  - [Dataset Structure](#dataset-structure)
36
  - [Data Instances](#data-instances)
@@ -58,15 +59,31 @@ paperswithcode_id: glue
58
  The CHANGE-IT dataset contains approximately 152,000 article-headline pairs, collected from two Italian newspapers situated at opposite ends of the political spectrum, namely la Repubblica (left) and Il Giornale (right), with the two newspapers equally represented. The dataset has been used in the context
59
  of the [CHANGE-IT task](https://sites.google.com/view/change-it) during the [Evalita 2020 evaluation campaign](http://www.evalita.it/2020). CHANGE-IT is a generation task for Italian – more specifically, a style transfer task for headlines of Italian newspapers. Given a (collection of) headlines from one newspaper, namely Il Giornale (G) or La Repubblica (R), it challenges automatic systems to change all G-headlines to headlines in style R, and all R-headlines to headlines in style G. Although the task only concerns headline change, the dataset comprehends both the headlines as well as their respective full articles.
60
 
61
- **Disclaimer**: *The CHANGE-IT dataset is hosted by the [European Language Grid](https://live.european-language-grid.eu/) and licensed under the [Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License](https://creativecommons.org/licenses/by-nc-sa/4.0/). To use the dataset using* 🤗 *Datasets, download and unzip the folder and pass it to the* `load_dataset` *method as:* `datasets.load_dataset('gsarti/change_it', data_dir='path/to/unzipped/folder')`
62
 
63
  ### Supported Tasks and Leaderboards
64
 
 
65
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
66
 
67
  ### Languages
68
 
69
- The language data in GLUE is in Italian (BCP-47 `it`)
70
 
71
  ## Dataset Structure
72
 
 
31
  - [Dataset Description](#dataset-description)
32
  - [Dataset Summary](#dataset-summary)
33
  - [Supported Tasks and Leaderboards](#supported-tasks-and-leaderboards)
34
+ - [Style Transfer](#style-transfer)
35
  - [Languages](#languages)
36
  - [Dataset Structure](#dataset-structure)
37
  - [Data Instances](#data-instances)
 
59
  The CHANGE-IT dataset contains approximately 152,000 article-headline pairs, collected from two Italian newspapers situated at opposite ends of the political spectrum, namely la Repubblica (left) and Il Giornale (right), with the two newspapers equally represented. The dataset has been used in the context
60
  of the [CHANGE-IT task](https://sites.google.com/view/change-it) during the [Evalita 2020 evaluation campaign](http://www.evalita.it/2020). CHANGE-IT is a generation task for Italian – more specifically, a style transfer task for headlines of Italian newspapers. Given a (collection of) headlines from one newspaper, namely Il Giornale (G) or La Repubblica (R), it challenges automatic systems to change all G-headlines to headlines in style R, and all R-headlines to headlines in style G. Although the task only concerns headline change, the dataset comprehends both the headlines as well as their respective full articles.
61
 
62
+ **Disclaimer**: *The CHANGE-IT dataset is hosted by the [European Language Grid](https://live.european-language-grid.eu/) and licensed under the [Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License](https://creativecommons.org/licenses/by-nc-sa/4.0/). To use the dataset using* 🤗 *Datasets, download and unzip the folder from its [ELG page](https://live.european-language-grid.eu/catalogue/corpus/7373) and pass it to the* `load_dataset` *method as:* `datasets.load_dataset('gsarti/change_it', data_dir='path/to/unzipped/folder')`
63
 
64
  ### Supported Tasks and Leaderboards
65
 
66
+ #### Style Transfer
67
 
68
+ The following table is taken from Table 4 of the original paper, where a *pointer-network* architecture is used as a baseline to perform style transfer in two settings. In the **rep2gio** variant the system is trained to summarize Repubblica headlines from full texts (vice versa for **gio2rep**), and the style transfer is performed by summarizing full texts of the other newspaper in the source newspaper's headline style. **avg** is the average of the two settings.
69
+
70
+ | | HH| AH|Main|Compliancy|
71
+ |--------:|---:|---:|---:|---------:|
72
+ |`rep2gio`|.649|.876|.799| .449|
73
+ |`gio2rep`|.639|.871|.435| .240|
74
+ | `avg`|.644|.874|.616| .345|
75
+
76
+ Here **Main**, **HH** and **AH** are all BERT-base models trained to evaluate the quality of style transfer as follows:
77
+
78
+ - **Main**: the model is trained to classify a generated headline either as `ilgiornale` or `repubblica`, achieving ~80% F1 score on gold data. Tests whether the transfer has been successful.
79
+ - **Headline-Headline (HH)**: the model is trained to check the compatibility between original and generated headlines. Tests whether the generation is coherent with the reference.
80
+ - **Article-Headline (AH)**: the model is trained to check the compatibility between original fulltext article and generated headlines. Tests whether the generation is coherent with the source article.
81
+
82
+ The final metric, **Overall compliancy**, is a binary metric that is positive if the other three metrics match (**Main** decision is reversed, **HH** and **AH** predict match), and negative otherwise. Refer to Section 3 of the original paper for more details.
83
 
84
  ### Languages
85
 
86
+ The language data in CHANGE-IT is in Italian (BCP-47 `it`)
87
 
88
  ## Dataset Structure
89