| Autor: |
Nikhilesh Cherukuri, Aditya Kiran Brahma |
| Jazyk: |
angličtina |
| Rok vydání: |
2022 |
| Předmět: |
|
| DOI: |
10.5281/zenodo.7047596 |
| Popis: |
Understanding the grammatical errors present in the chat conver- sations is crucial in developing chatbot with high quality data. In food delivery platforms, conversational AI development via chatbot platform is continuously built to understand the context of the cus- tomer conversations and suggest the next utterance by retrieving them from a similar scenario occurred in the past. These sugges- tions if used by agents are sometimes manually edited further based on the relevance in current scenario and suggestion quality. The grammatical quality of suggestions play significant role for the agents to utilize conversational AI assistance and provide better customer resolutions in quick and effective manner. In this paper, we analyse a use case of identifying the frequent grammatical er- rors present in the texts typed by the customer care agents and utilize them to build an automatic grammatical error correction model for the data specific to food delivery conversations. We show that using large pretrained encoder-decoder transformer models and systematic fine-tuning with a smaller downstream task specific data (Grammatical error correction) achieved an overall gain of 15.5 % GLEU score compared to the the baseline approach of using the pretrained model alone. |
| Databáze: |
OpenAIRE |
| Externí odkaz: |
|
