Multi-turn Dialogue Systems in Outbound Telemarketing

To optimize the present retrieval-based multi-turn dialogue model and help the company sponsor improve the outbound sales pitch.

Sorted out the evolution of dialogue systems, retrieval dialogue systems, and single/multi-turn dialogue systems and studied how the models were improved step by step.
Chose Deep Attention Matching (DAM) as the basic model and reimplemented it in Python.
Analyzed datasets provided by partner companies to learn existing problems and used SimBERT for data augmentation for problem-solving.
Used Term Frequency-Inverse Document Frequency (TFIDF) to map all the responses in the company datasets to standard responses and added rule-matching to improve the efficiency.
Applied the improved DAM on the company datasets, trained the algorithm, and evaluated the model performance.
Achieved a better model performance using SimBERT for data augmentation and got the best results of the model trained with the 9-turn dialogue, with an accuracy rate of 0.573 in the R10@1 test, 0.735 in R10@2, and 0.901 in R10@5.