pratikratadiya/toxicity_detection

Detection of toxicity in a given comment using various sequence models

Toxicity_detection

Detection of toxic nature of comment using the toxic comment dataset(https://github.com/vzhou842/profanity-check/blob/master/profanity_check/data/clean_data.csv)

The data includes both tweets as well as wikipedia edits as obtained from the Toxic comment classification challenge on Kaggle.

The following implementations are done:

1. LSTM
2. LSTM with GLoVE 100D word embeddings
3. LSTM with GLoVE 300D word embeddings
4. Attention mechanism with GLoVe 300D word embeddings

The results obtained were as follows:

Model	Training accuracy(%)	Testing accuracy(%)
LSTM	82	80
LSTM + GLoVe(100D)	95	96.12
LSTM + GLoVe(300D)	95.34	95.94
Multi layer LSTM + GLoVe(300D)	97.06	96.14
LSTM + Attention + GLoVe(300D)	98.16	95.87

Future scope includes improvement in the attention layer to increase testing accuracy.

Also, transformer architecture can be used to improve the performance further

toxicity_detection's Issues

TypeError: add_weight() got multiple values for argument 'name'

`---------------------------------------------------------------------------
TypeError Traceback (most recent call last)
in ()
6 x1 = Bidirectional(LSTM(60, return_sequences=True))(x1)
7 x1 = Dropout(0.3)(x1)
----> 8 x1 = Attention(max_len)(x1)
9
10 x2 = Embedding(vocab_size + 1, embedding_dim, weights=[embedding_matrix], input_length=max_len, trainable=True)(comment_input_post)

2 frames
in build(self, input_shape)
42 name='{}_W'.format(self.name),
43 regularizer=self.W_regularizer,
---> 44 constraint=self.W_constraint)
45 self.features_dim = input_shape[-1]
46

TypeError: add_weight() got multiple values for argument 'name'`