rushi-the-neural-arch / openai-nlp-hackathon-symbiosis-university

This repository is made in lieu of submission towards the solution of problem statement 2 of the OPEN AI NLP hackathon. The objective here is to classify the voice recordings of a call center proceeding by treating them as consumer complaints into the said categories of the automotive industry

OpenAI-NLP-Hackathon-Symbiosis-University

This repository is made in lieu of submission towards the solution of problem statement 2 of the OPEN AI NLP hackathon. The objective here is to classify the voice recordings of a call center proceeding by treating them as consumer complaints into the said categories of the automotive industry

Consumer-Complaint-Classification-OPEN-AI

This repository is made in lieu of submission towards the solution of problem statement 2 of the OPEN AI NLP hackathon. The objective here is to classify the voice recordings of a call center proceeding by treating them as consumer complaints into the said categories of the automotive industry.

Dataset

https://www.kaggle.com/dushyantv/consumer_complaints

Methodology

Pre-Training

• Batching complaints from the common pool as a pair with SQUAD like format.
• Evaluation model for pre-training used is: SQUAD 2.0 dev

2D Feature Space Generation

• For each pair from the pre-trained pool, a 2d vector is plotted on the feature space.
• The resultant magnitude of that vector is taken as a sentence embeddinng.
• This embedding is validated using the gensim wordmodel, to create the sentence level representations.
• These representations are further clustered by biased c-means to find the average number of classes.
• Further, the number of classes is mapped against the existing classes of consumer complaints.

Model Training

• A 6 layer, 32 node wide CNN is created with the node weights initialized with the sentence level representations.
• This network is trained on the basis of end-member uniqueness for each class, to define the boundary of each class of complaints.

Model Evaluation / Testing

• The results are evaluated using a reverse TF-IDF matrix, mapping it with the original class matrix.
• The dot product of these matrices determines the correctness of the model:
  1. a dense matrix represents high variance and hence low accuracy
  2. a sparse matrix represents a low variance and high accuracy.

References

• SQUAD 2.0 pre-training was referenced from the sequel paper on SQUAD 2.0 dataset and training strategy.

Rajpurkar, Pranav, Robin Jia, and Percy Liang. “Know What You Don’t Know: Unanswerable Questions for SQuAD.” Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), 2018. https://doi.org/10.18653/v1/p18-2124.

• Also, a shoutout to supreetkt for providing a foundation for consumer complaints classification on text based responses!