Skip to content

Latest commit

 

History

History
198 lines (148 loc) · 6.15 KB

README.md

File metadata and controls

198 lines (148 loc) · 6.15 KB

GoEmotions Pytorch

Pytorch Implementation of GoEmotions with Huggingface Transformers

What is GoEmotions

Dataset labeled 58000 Reddit comments with 28 emotions

  • 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

Training Details

  • Use bert-base-cased (Same as the paper's code)

  • In paper, 3 Taxonomies were used. I've also made the data with new taxonomy labels for hierarchical grouping and ekman.

    1. Original GoEmotions (27 emotions + neutral)
    2. Hierarchical Grouping (positive, negative, ambiguous + neutral)
    3. Ekman (anger, disgust, fear, joy, sadness, surprise + neutral)

Vocabulary

  • I've replace [unused1], [unused2] to [NAME], [RELIGION] in the vocab, respectively.
[PAD]
[NAME]
[RELIGION]
[unused3]
[unused4]
...
  • I've also set special_tokens_map.json as below, so the tokenizer won't split the [NAME] or [RELIGION] into its word pieces.
{
  "unk_token": "[UNK]",
  "sep_token": "[SEP]",
  "pad_token": "[PAD]",
  "cls_token": "[CLS]",
  "mask_token": "[MASK]",
  "additional_special_tokens": ["[NAME]", "[RELIGION]"]
}

Requirements

  • torch==1.4.0
  • transformers==2.11.0
  • attrdict==2.0.1

Hyperparameters

You can change the parameters from the json files in config directory.

Parameter
Learning rate 5e-5
Warmup proportion 0.1
Epochs 10
Max Seq Length 50
Batch size 16

How to Run

For taxonomy, choose original, group or ekman

$ python3 run_goemotions.py --taxonomy {$TAXONOMY}

$ python3 run_goemotions.py --taxonomy original
$ python3 run_goemotions.py --taxonomy group
$ python3 run_goemotions.py --taxonomy ekman

Results

Best Result of Macro F1

Macro F1 (%) Dev Test
original 50.16 50.30
group 69.41 70.06
ekman 62.59 62.38

Pipeline

  • Inference for multi-label classification was made possible by creating a new MultiLabelPipeline class.
  • Already uploaded finetuned model on Huggingface S3.
    • Original GoEmotions Taxonomy: monologg/bert-base-cased-goemotions-original
    • Hierarchical Group Taxonomy: monologg/bert-base-cased-goemotions-group
    • Ekman Taxonomy: monologg/bert-base-cased-goemotions-ekman

1. Original GoEmotions Taxonomy

from transformers import BertTokenizer
from model import BertForMultiLabelClassification
from multilabel_pipeline import MultiLabelPipeline
from pprint import pprint

tokenizer = BertTokenizer.from_pretrained("monologg/bert-base-cased-goemotions-original")
model = BertForMultiLabelClassification.from_pretrained("monologg/bert-base-cased-goemotions-original")

goemotions = MultiLabelPipeline(
    model=model,
    tokenizer=tokenizer,
    threshold=0.3
)

texts = [
    "Hey that's a thought! Maybe we need [NAME] to be the celebrity vaccine endorsement!",
    "it’s happened before?! love my hometown of beautiful new ken 😂😂",
    "I love you, brother.",
    "Troll, bro. They know they're saying stupid shit. The motherfucker does nothing but stink up libertarian subs talking shit",
]

pprint(goemotions(texts))

# Output
 [{'labels': ['neutral'], 'scores': [0.9750906]},
 {'labels': ['curiosity', 'love'], 'scores': [0.9694574, 0.9227462]},
 {'labels': ['love'], 'scores': [0.993483]},
 {'labels': ['anger'], 'scores': [0.99225825]}]

2. Group Taxonomy

from transformers import BertTokenizer
from model import BertForMultiLabelClassification
from multilabel_pipeline import MultiLabelPipeline
from pprint import pprint

tokenizer = BertTokenizer.from_pretrained("monologg/bert-base-cased-goemotions-group")
model = BertForMultiLabelClassification.from_pretrained("monologg/bert-base-cased-goemotions-group")

goemotions = MultiLabelPipeline(
    model=model,
    tokenizer=tokenizer,
    threshold=0.3
)

texts = [
    "Hey that's a thought! Maybe we need [NAME] to be the celebrity vaccine endorsement!",
    "it’s happened before?! love my hometown of beautiful new ken 😂😂",
    "I love you, brother.",
    "Troll, bro. They know they're saying stupid shit. The motherfucker does nothing but stink up libertarian subs talking shit",
]

pprint(goemotions(texts))

# Output
[{'labels': ['positive'], 'scores': [0.9989434]},
 {'labels': ['ambiguous', 'positive'], 'scores': [0.99801123, 0.99845874]},
 {'labels': ['positive'], 'scores': [0.99930394]},
 {'labels': ['negative'], 'scores': [0.9984231]}]

3. Ekman Taxonomy

from transformers import BertTokenizer
from model import BertForMultiLabelClassification
from multilabel_pipeline import MultiLabelPipeline
from pprint import pprint

tokenizer = BertTokenizer.from_pretrained("monologg/bert-base-cased-goemotions-ekman")
model = BertForMultiLabelClassification.from_pretrained("monologg/bert-base-cased-goemotions-ekman")

goemotions = MultiLabelPipeline(
    model=model,
    tokenizer=tokenizer,
    threshold=0.3
)

texts = [
    "Hey that's a thought! Maybe we need [NAME] to be the celebrity vaccine endorsement!",
    "it’s happened before?! love my hometown of beautiful new ken 😂😂",
    "I love you, brother.",
    "Troll, bro. They know they're saying stupid shit. The motherfucker does nothing but stink up libertarian subs talking shit",
]

pprint(goemotions(texts))

# Output
 [{'labels': ['joy', 'neutral'], 'scores': [0.30459446, 0.9217335]},
 {'labels': ['joy', 'surprise'], 'scores': [0.9981395, 0.99863845]},
 {'labels': ['joy'], 'scores': [0.99910116]},
 {'labels': ['anger'], 'scores': [0.9984291]}]

Reference