Cargando…
Optimized Transformer Models for FAQ Answering
Informational chatbots provide a highly effective medium for improving operational efficiency in answering customer queries for any enterprise. Chatbots are also preferred by users/customers since unlike other alternatives like calling customer care or browsing over FAQ pages, chatbots provide insta...
Autores principales: | , , , , |
---|---|
Formato: | Online Artículo Texto |
Lenguaje: | English |
Publicado: |
2020
|
Materias: | |
Acceso en línea: | https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7206240/ http://dx.doi.org/10.1007/978-3-030-47426-3_19 |
Sumario: | Informational chatbots provide a highly effective medium for improving operational efficiency in answering customer queries for any enterprise. Chatbots are also preferred by users/customers since unlike other alternatives like calling customer care or browsing over FAQ pages, chatbots provide instant responses, are easy to use, are less invasive and are always available. In this paper, we discuss the problem of FAQ answering which is central to designing a retrieval-based informational chatbot. Given a set of FAQ pages s for an enterprise, and a user query, we need to find the best matching question-answer pairs from s. Building such a semantic ranking system that works well across domains for large QA databases with low runtime and model size is challenging. Previous work based on feature engineering or recurrent neural models either provides low accuracy or incurs high runtime costs. We experiment with multiple transformer based deep learning models, and also propose a novel MT-DNN (Multi-task Deep Neural Network)-based architecture, which we call Masked MT-DNN (or MMT-DNN). MMT-DNN significantly outperforms other state-of-the-art transformer models for the FAQ answering task. Further, we propose an improved knowledge distillation component to achieve [Formula: see text]2.4x reduction in model-size and [Formula: see text]7x reduction in runtime while maintaining similar accuracy. On a small benchmark dataset from SemEval 2017 CQA Task 3, we show that our approach provides an NDCG@1 of 83.1. On another large dataset of [Formula: see text]281K instances corresponding to [Formula: see text]30K queries from diverse domains, our distilled 174 MB model provides an NDCG@1 of 75.08 with a CPU runtime of mere 31 ms establishing a new state-of-the-art for FAQ answering. |
---|