SamLowe/bge-small-en-go_emotions-classifier-onnx

Overview

This is a multi-label, multi-class linear classifer for emotions that works with BGE-small-en embeddings, having been trained on the go_emotions dataset.

Labels

The 28 labels from the go_emotions dataset are:

['admiration', 'amusement', 'anger', 'annoyance', 'approval', 'caring', 'confusion', 'curiosity', 'desire', 'disappointment', 'disapproval', 'disgust', 'embarrassment', 'excitement', 'fear', 'gratitude', 'grief', 'joy', 'love', 'nervousness', 'optimism', 'pride', 'realization', 'relief', 'remorse', 'sadness', 'surprise', 'neutral']

Metrics (exact match of labels per item)

This is a multi-label, multi-class dataset, so each label is effectively a separate binary classification. Evaluating across all labels per item in the go_emotions test split the metrics are shown below.

Optimising the threshold per label to optimise the F1 metric, the metrics (evaluated on the go_emotions test split) are:

• Precision: 0.429
• Recall: 0.483
• F1: 0.439

Weighted by the relative support of each label in the dataset, this is:

• Precision: 0.457
• Recall: 0.585
• F1: 0.502

Using a fixed threshold of 0.5 to convert the scores to binary predictions for each label, the metrics (evaluated on the go_emotions test split, and unweighted by support) are:

• Precision: 0.650
• Recall: 0.189
• F1: 0.249

Metrics (per-label)

This is a multi-label, multi-class dataset, so each label is effectively a separate binary classification and metrics are better measured per label.

Optimising the threshold per label to optimise the F1 metric, the metrics (evaluated on the go_emotions test split) are:

	f1	precision	recall	support	threshold
admiration	0.561	0.517	0.613	504	0.25
amusement	0.647	0.663	0.633	264	0.20
anger	0.324	0.238	0.510	198	0.10
annoyance	0.292	0.200	0.541	320	0.10
approval	0.335	0.297	0.385	351	0.15
caring	0.306	0.221	0.496	135	0.10
confusion	0.360	0.400	0.327	153	0.20
curiosity	0.461	0.392	0.560	284	0.15
desire	0.411	0.476	0.361	83	0.25
disappointment	0.204	0.150	0.318	151	0.10
disapproval	0.357	0.291	0.461	267	0.15
disgust	0.403	0.417	0.390	123	0.20
embarrassment	0.424	0.483	0.378	37	0.30
excitement	0.298	0.255	0.359	103	0.15
fear	0.609	0.590	0.628	78	0.25
gratitude	0.801	0.819	0.784	352	0.30
grief	0.500	0.500	0.500	6	0.75
joy	0.437	0.453	0.422	161	0.20
love	0.641	0.693	0.597	238	0.30
nervousness	0.356	0.364	0.348	23	0.45
optimism	0.416	0.538	0.339	186	0.25
pride	0.500	0.750	0.375	16	0.65
realization	0.247	0.228	0.269	145	0.10
relief	0.364	0.273	0.545	11	0.30
remorse	0.581	0.529	0.643	56	0.25
sadness	0.525	0.519	0.532	156	0.20
surprise	0.301	0.235	0.418	141	0.10
neutral	0.626	0.519	0.786	1787	0.30

The thesholds are stored in thresholds.json.

Use with ONNXRuntime

The input to the model is called logits, and there is one output per label. Each output produces a 2d array, with 1 row per input row, and each row having 2 columns - the first being a proba output for the negative case, and the second being a proba output for the positive case.

# Assuming you have embeddings from BAAI/bge-small-en for the input sentences
# E.g. produced from sentence-transformers E.g. huggingface.co/BAAI/bge-small-en
#      or from an ONNX version E.g. huggingface.co/Xenova/bge-small-en

print(embeddings.shape)  # E.g. a batch of 1 sentence
> (1, 384)

import onnxruntime as ort

sess = ort.InferenceSession("path_to_model_dot_onnx", providers=['CPUExecutionProvider'])

outputs = [o.name for o in sess.get_outputs()]  # list of labels, in the order of the outputs
preds_onnx = sess.run(_outputs, {'logits': embeddings})
# preds_onnx is a list with 28 entries, one per label,
# each with a numpy array of shape (1, 2) given the input was a batch of 1

print(outputs[0])
> surprise
print(preds_onnx[0])
> array([[0.97136074, 0.02863926]], dtype=float32)

# load thresholds.json and use that (per label) to convert the positive case score to a binary prediction

Commentary on the dataset

Some labels (E.g. gratitude) when considered independently perform very strongly, whilst others (E.g. relief) perform very poorly.

This is a challenging dataset. Labels such as relief do have much fewer examples in the training data (less than 100 out of the 40k+, and only 11 in the test split).

But there is also some ambiguity and/or labelling errors visible in the training data of go_emotions that is suspected to constrain the performance. Data cleaning on the dataset to reduce some of the mistakes, ambiguity, conflicts and duplication in the labelling would produce a higher performing model.