j-hartmann commited on
Commit
56a10c2
•
1 Parent(s): 312b829

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +1 -3
README.md CHANGED
@@ -46,7 +46,7 @@ Thanks to Samuel Domdey and chrsiebert for their support in making this model av
46
 
47
  ## Appendix 📚
48
 
49
- Please find an overview of the datasets used for training below. All datasets contain English text. The table summarizes which emotions are available in each of the datasets.
50
 
51
  |Name|anger|disgust|fear|joy|neutral|sadness|surprise|
52
  |---|---|---|---|---|---|---|---|
@@ -57,6 +57,4 @@ Please find an overview of the datasets used for training below. All datasets co
57
  |MELD, Poria et al. (2019)|Yes|Yes|Yes|Yes|Yes|Yes|Yes|
58
  |SemEval-2018, EI-reg (Mohammad et al. 2018) |Yes|-|Yes|Yes|-|Yes|-|
59
 
60
- The datasets represent a diverse collection of text types. Specifically, they contain emotion labels for texts from Twitter, Reddit, student self-reports, and utterances from TV dialogues. As MELD (Multimodal EmotionLines Dataset) extends the popular EmotionLines dataset, EmotionLines itself is not included here.
61
-
62
  The model is trained on a balanced subset from the datasets listed above (2,811 observations per emotion, i.e., nearly 20k observations in total). The evaluation accuracy on a holdout test set is 66% (and significantly above the random-chance baseline of 1/7 = 14%).
 
46
 
47
  ## Appendix 📚
48
 
49
+ Please find an overview of the datasets used for training below. All datasets contain English text. The table summarizes which emotions are available in each of the datasets. The datasets represent a diverse collection of text types. Specifically, they contain emotion labels for texts from Twitter, Reddit, student self-reports, and utterances from TV dialogues. As MELD (Multimodal EmotionLines Dataset) extends the popular EmotionLines dataset, EmotionLines itself is not included here.
50
 
51
  |Name|anger|disgust|fear|joy|neutral|sadness|surprise|
52
  |---|---|---|---|---|---|---|---|
 
57
  |MELD, Poria et al. (2019)|Yes|Yes|Yes|Yes|Yes|Yes|Yes|
58
  |SemEval-2018, EI-reg (Mohammad et al. 2018) |Yes|-|Yes|Yes|-|Yes|-|
59
 
 
 
60
  The model is trained on a balanced subset from the datasets listed above (2,811 observations per emotion, i.e., nearly 20k observations in total). The evaluation accuracy on a holdout test set is 66% (and significantly above the random-chance baseline of 1/7 = 14%).