system HF staff commited on
Commit
57f849d
1 Parent(s): 573e0f0

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +109 -0
README.md ADDED
@@ -0,0 +1,109 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Multilingual Joint Fine-tuning of Transformer models for identifying Trolling, Aggression and Cyberbullying at TRAC 2020
2
+
3
+ Models and predictions for submission to TRAC - 2020 Second Workshop on Trolling, Aggression and Cyberbullying.
4
+
5
+ Our trained models as well as evaluation metrics during traing are available at: https://databank.illinois.edu/datasets/IDB-8882752#
6
+ We also make a few of our models available in HuggingFace's models repository at https://huggingface.co/socialmediaie/, these models can be further fine-tuned on your dataset of choice.
7
+
8
+ Our approach is described in our paper titled:
9
+
10
+ > Mishra, Sudhanshu, Shivangi Prasad, and Shubhanshu Mishra. 2020. "Multilingual Joint Fine-Tuning of Transformer Models for Identifying Trolling, Aggression and Cyberbullying at TRAC 2020." In Proceedings of the Second Workshop on Trolling, Aggression and Cyberbullying (TRAC-2020).
11
+
12
+ The source code for training this model and more details can be found on our code repository: https://github.com/socialmediaie/TRAC2020
13
+
14
+ NOTE: These models are retrained for uploading here after our submission so the evaluation measures may be slightly different from the ones reported in the paper.
15
+
16
+ If you plan to use the dataset please cite the following resources:
17
+
18
+ * Mishra, Sudhanshu, Shivangi Prasad, and Shubhanshu Mishra. 2020. "Multilingual Joint Fine-Tuning of Transformer Models for Identifying Trolling, Aggression and Cyberbullying at TRAC 2020." In Proceedings of the Second Workshop on Trolling, Aggression and Cyberbullying (TRAC-2020).
19
+ * Mishra, Shubhanshu, Shivangi Prasad, and Shubhanshu Mishra. 2020. “Trained Models for Multilingual Joint Fine-Tuning of Transformer Models for Identifying Trolling, Aggression and Cyberbullying at TRAC 2020.” University of Illinois at Urbana-Champaign. https://doi.org/10.13012/B2IDB-8882752_V1.
20
+
21
+
22
+ ```
23
+ @inproceedings{Mishra2020TRAC,
24
+ author = {Mishra, Sudhanshu and Prasad, Shivangi and Mishra, Shubhanshu},
25
+ booktitle = {Proceedings of the Second Workshop on Trolling, Aggression and Cyberbullying (TRAC-2020)},
26
+ title = {{Multilingual Joint Fine-tuning of Transformer models for identifying Trolling, Aggression and Cyberbullying at TRAC 2020}},
27
+ year = {2020}
28
+ }
29
+
30
+ @data{illinoisdatabankIDB-8882752,
31
+ author = {Mishra, Shubhanshu and Prasad, Shivangi and Mishra, Shubhanshu},
32
+ doi = {10.13012/B2IDB-8882752_V1},
33
+ publisher = {University of Illinois at Urbana-Champaign},
34
+ title = {{Trained models for Multilingual Joint Fine-tuning of Transformer models for identifying Trolling, Aggression and Cyberbullying at TRAC 2020}},
35
+ url = {https://doi.org/10.13012/B2IDB-8882752{\_}V1},
36
+ year = {2020}
37
+ }
38
+ ```
39
+
40
+
41
+ ## Usage
42
+
43
+ The models can be used via the following code:
44
+
45
+ ```python
46
+ from transformers import AutoModel, AutoTokenizer, AutoModelForSequenceClassification
47
+ import torch
48
+ from pathlib import Path
49
+ from scipy.special import softmax
50
+ import numpy as np
51
+ import pandas as pd
52
+
53
+ TASK_LABEL_IDS = {
54
+ "Sub-task A": ["OAG", "NAG", "CAG"],
55
+ "Sub-task B": ["GEN", "NGEN"],
56
+ "Sub-task C": ["OAG-GEN", "OAG-NGEN", "NAG-GEN", "NAG-NGEN", "CAG-GEN", "CAG-NGEN"]
57
+ }
58
+
59
+ model_version="databank" # other option is hugging face library
60
+ if model_version == "databank":
61
+ # Make sure you have downloaded the required model file from https://databank.illinois.edu/datasets/IDB-8882752
62
+ # Unzip the file at some model_path (we are using: "databank_model")
63
+ model_path = next(Path("databank_model").glob("./*/output/*/model"))
64
+ # Assuming you get the following type of structure inside "databank_model"
65
+ # 'databank_model/ALL/Sub-task C/output/bert-base-multilingual-uncased/model'
66
+ lang, task, _, base_model, _ = model_path.parts
67
+ tokenizer = AutoTokenizer.from_pretrained(base_model)
68
+ model = AutoModelForSequenceClassification.from_pretrained(model_path)
69
+ else:
70
+ lang, task, base_model = "ALL", "Sub-task C", "bert-base-multilingual-uncased"
71
+ base_model = f"socialmediaie/TRAC2020_{lang}_{lang.split()[-1]}_{base_model}"
72
+ tokenizer = AutoTokenizer.from_pretrained(base_model)
73
+ model = AutoModelForSequenceClassification.from_pretrained(base_model)
74
+
75
+ # For doing inference set model in eval mode
76
+ model.eval()
77
+ # If you want to further fine-tune the model you can reset the model to model.train()
78
+
79
+ task_labels = TASK_LABEL_IDS[task]
80
+
81
+ sentence = "This is a good cat and this is a bad dog."
82
+ processed_sentence = f"{tokenizer.cls_token} {sentence}"
83
+ tokens = tokenizer.tokenize(sentence)
84
+ indexed_tokens = tokenizer.convert_tokens_to_ids(tokens)
85
+ tokens_tensor = torch.tensor([indexed_tokens])
86
+
87
+ with torch.no_grad():
88
+ logits, = model(tokens_tensor, labels=None)
89
+ logits
90
+
91
+
92
+ preds = logits.detach().cpu().numpy()
93
+ preds_probs = softmax(preds, axis=1)
94
+ preds = np.argmax(preds_probs, axis=1)
95
+ preds_labels = np.array(task_labels)[preds]
96
+ print(dict(zip(task_labels, preds_probs[0])), preds_labels)
97
+ """You should get an output as follows:
98
+
99
+ ({'CAG-GEN': 0.06762535,
100
+ 'CAG-NGEN': 0.03244293,
101
+ 'NAG-GEN': 0.6897794,
102
+ 'NAG-NGEN': 0.15498641,
103
+ 'OAG-GEN': 0.034373745,
104
+ 'OAG-NGEN': 0.020792078},
105
+ array(['NAG-GEN'], dtype='<U8'))
106
+
107
+ """
108
+
109
+ ```