SamLowe/bge-small-en-v1.5-go_emotions-classifier-onnx

Overview

This is a multi-label, multi-class linear classifer for emotions that works with BGE-small-en-v1.5 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: 0.445
• Recall: 0.476
• F1: 0.449

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

• Precision: 0.472
• Recall: 0.582
• F1: 0.514

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: 0.602
• Recall: 0.250
• F1: 0.303

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	0.583	0.574	0.593	504	0.30
amusement	0.668	0.722	0.621	264	0.25
anger	0.350	0.309	0.404	198	0.15
annoyance	0.299	0.318	0.281	320	0.20
approval	0.338	0.281	0.425	351	0.15
caring	0.321	0.323	0.319	135	0.20
confusion	0.384	0.313	0.497	153	0.15
curiosity	0.467	0.432	0.507	284	0.20
desire	0.426	0.381	0.482	83	0.20
disappointment	0.210	0.147	0.364	151	0.10
disapproval	0.366	0.288	0.502	267	0.15
disgust	0.416	0.409	0.423	123	0.20
embarrassment	0.370	0.341	0.405	37	0.30
excitement	0.313	0.368	0.272	103	0.25
fear	0.615	0.677	0.564	78	0.40
gratitude	0.828	0.810	0.847	352	0.25
grief	0.545	0.600	0.500	6	0.85
joy	0.455	0.429	0.484	161	0.20
love	0.642	0.673	0.613	238	0.30
nervousness	0.350	0.412	0.304	23	0.60
optimism	0.439	0.417	0.462	186	0.20
pride	0.480	0.667	0.375	16	0.70
realization	0.232	0.191	0.297	145	0.10
relief	0.353	0.500	0.273	11	0.50
remorse	0.643	0.529	0.821	56	0.20
sadness	0.526	0.497	0.558	156	0.20
surprise	0.329	0.318	0.340	141	0.15
neutral	0.634	0.528	0.794	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-v1.5 for the input sentences
# E.g. produced from sentence-transformers E.g. huggingface.co/BAAI/bge-small-en-v1.5
#      or from an ONNX version E.g. huggingface.co/Xenova/bge-small-en-v1.5

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.