Update README.md

README.md

@@ -13,7 +13,7 @@ size_categories:
 
 # EmoBench-UA: Emotions Detection Dataset in Ukrainian Texts
 
-<img alt="EmoBench-UA" src="intro_logo.png">
+<img alt="EmoBench-UA" style="width: 50%; height: 50%" src="intro_logo.png">
 
 **EmoBench-UA**: the first of its kind emotions detection dataset in Ukrainian texts. This dataset covers the detection of basic emotions: Joy, Anger, Fear, Disgust, Surprise, Sadness, or None.
 Any text can contain any amount of emotion -- only one, several, or none at all. The texts with *None* emotions are the ones where the labels per emotions classes are 0.
@@ -21,14 +21,33 @@ Any text can contain any amount of emotion -- only one, several, or none at all.
 *Binary*: specifically this dataset contains binary labels indicating simply presence of any emotion in the text.
 
 ## Data Collection
-The data collection was done via [Toloka.ai](https://toloka.ai}{https://toloka.ai) crowdsourcing platform. As per original Ukrainian texts, we used the opensourced [corpus of Ukrainian tweets](https://github.com/kateryna-bobrovnyk/ukr-twi-corpus).
+The data collection was done via [Toloka.ai](https://toloka.ai}{https://toloka.ai) crowdsourcing platform. 
+For original Ukrainian texts, we used the opensourced [corpus of Ukrainian tweets](https://github.com/kateryna-bobrovnyk/ukr-twi-corpus).
+
+Firstly, we did data pre-filtering:
+
+**Length** We applied a length-based filter, discarding texts that were too short (N words < 5), as such samples often consist of hashtags or other non-informative tokens. Similarly, overly long texts (N words >= 50) were excluded, as longer sequences tend to obscure the central meaning and make it challenging to accurately identify the expressed emotions.
+
+**Toxicity** While toxic texts can carry quite strong emotions, to ensure annotators well-being and general appropriateness of our corpus, we filtered out too toxic instances using our opensourced [toxicity classifier](https://huggingface.co/ukr-detect/ukr-toxicity-classifier).
+
+**Emotional Texts Pre-selection** To avoid an excessive imbalance toward emotionless texts, we performed a pre-selection step aimed at identifying texts likely to express emotions. Specifically, we applied the English emotion classifier [DistillRoBERTa-Emo-EN](https://huggingface.co/michellejieli/emotion_text_classifier) on translated Ukrainian texts with [NLLB](https://huggingface.co/facebook/nllb-200-distilled-600M) model.
+
+Then, we utilized several control strategies to ensure the quality of the data:
+* annotators were native Ukrainian speakers;
+* annotators were obliged to finish training and examination before being allowed to perform annotation;
+* annotators were permanently banned if they submitted the last three task pages in under 15 seconds each, indicating low engagement;
+* a one-day ban was triggered if three consecutive pages were skipped;
+* annotators were asked to take a 30-minute break after completing 25 consecutive pages;
+* control tasks were randomly injected to check.
+
+Finally, each sample were annotated by 5 annotators. We took for datasets only instances with 90% confidence score.
 
 ## Splits Statistics
 
 Overall, the datasets contains of 4949 labelled instances. Krippendorff's alpha annotation agreement score is 0.85. 
 Then, we partitioned the dataset into fixed train/development/test subsets following a 50/5/45% split ratio.
 
-<img alt="Splits Statistic" src="data_stats.png">
+<img alt="Splits Statistic" style="width: 50%; height: 50%" src="data_stats.png">
 
 ## Citation