Update README.md
Browse files
README.md
CHANGED
@@ -9,21 +9,10 @@ datasets:
|
|
9 |
|
10 |
|
11 |
# Model Description
|
|
|
|
|
12 |
|
13 |
-
|
14 |
-
|
15 |
-
The fine-tuned distilbert-base-uncased can be found [here](https://huggingface.co/bhadresh-savani/bert-base-go-emotion).
|
16 |
-
|
17 |
-
### Limitations and bias
|
18 |
-
|
19 |
-
GoEmotions:
|
20 |
-
1) Demographics of Reddit Users
|
21 |
-
2) Imbalanced class distribution
|
22 |
-
3) ...
|
23 |
-
|
24 |
-
EmpatheticDialogues:
|
25 |
-
1) Unable to ascertain the degree of cultural specificity for the context that a respondent described when given an emotion label (i.e., p(description | emotion, *culture*))
|
26 |
-
2) ...
|
27 |
|
28 |
|
29 |
## Training data
|
@@ -34,3 +23,9 @@ EmpatheticDialogues:
|
|
34 |
|
35 |
## Evaluation results
|
36 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
9 |
|
10 |
|
11 |
# Model Description
|
12 |
+
Yet another Transformer model fine-tuned for approximating another non-linear mapping between X and Y? That's right!
|
13 |
+
This is your good ol' emotion classifier - given an input text, the model outputs a probability distribution over a set of pre-selected emotion words. In this case, it is 32, which is the number of emotion classes in the [Empathetic Dialogues](https://huggingface.co/datasets/bdotloh/empathetic-dialogues-contexts) dataset.
|
14 |
|
15 |
+
This model is built "on top of" a [distilbert-base-uncased model fine-tuned on the go-emotions dataset](https://huggingface.co/bhadresh-savani/bert-base-go-emotion). Y'all should really check out that model, it even contains a jupyter notebook file that illustrates how the model was trained (bhadresh-savani if you see this, thank you!).
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
16 |
|
17 |
|
18 |
## Training data
|
|
|
23 |
|
24 |
## Evaluation results
|
25 |
|
26 |
+
### Limitations and bias
|
27 |
+
Well where should we begin...
|
28 |
+
|
29 |
+
EmpatheticDialogues:
|
30 |
+
1) Unable to ascertain the degree of cultural specificity for the context that a respondent described when given an emotion label (i.e., p(description | emotion, *culture*))
|
31 |
+
2) ...
|