BpHigh@TamilNLP-ACL2022: Effects of Data Augmentation on Indic-Transformer based classifier for Abusive Comments Detection in Tamil
Kernels used for model training using Kaggle Notebooks and other experimental kernels as well as Kernels used for data augmentation purposes.
-
The shared task on Abusive comment detection in Tamil-ACL 2022 is a comment classification problem that can be further described as a multi-class text classification problem in Tamil native script and Tamil-English code-mixed.
-
Given a YouTube comment, the systems submitted by the participants should classify its abusive categories.
-
The participants were provided with development, training and test dataset in Tamil and Tamil-English.
-
The dataset is tagged using various classes namely, Homophobia, Misandry, Counter-speech, Misogyny, Xenophobia, Transphobic and hope speech.
-
The dataset consists of rows that contain the comment text and the label assigned to that comment
- Bhavish Pahwa 🏄♂️ (GitHub)
-
We build a classifier using the MURIL Transformer as our embedding layer(all layers frozen) and attach a classifier head by adding subsequent convolution and dense layers. The final output dense layer has softmax activation, which gives us the final predictions.
-
We use two data Augmentation approaches to improve our model performance.
-
We define an equation to generate a balanced form of the original shared task dataset through our augmentation approaches.
-
We take the help of the NlpAug library , which provides the methods to perform word-level augmentation using contextual models as well as non-contextual word embeddings like Word2vec, fastText, and Glove.
-
The above equation shows us the multiplier value M, used while generating the augmented sentences. M refers to the value by which the number of occurrences of a label should change, and N is the number of occurrences of a label, also called the value count of a label. L refers to the set of class labels.
-
In terms of words, the above equation conveys that the multiplier value M(i) for label i is equal to the floor division of the value count for the label having maximum count and the value count for label i.
-
We use the MURIL Transformer again as a "Contextual Word Embedding Augmenter" to generate word-level augmented sentences. Then we train our classifier using this new balanced version of the train dataset.
-
We use the IndicNLP tokenizer for Indian languages for pre-processing the input sentences and the Tamil fastText model from the IndicNLP suite as a ’Word Embeddings Augmenter’ to generate word-level augmented sentences. Then we train our classifier using this new balanced version of the train dataset.
