15 Percent Increase in Conversions: Computer scientists have found a solution to the miscommunication problem. The team examined dialog systems, such as personal assistants and chatbots designed for humans.
When these systems are confronted with idioms or similes, their performance drops by 10-20%, according to the researchers.
A partial solution was also developed by the research team: a simple script that recognizes figurative phrases and replaces them with their literal meaning. As a result, the performance of dialog systems improved by up to 15%.
The findings will be presented at the Empirical Methods in Natural Language Processing (EMNLP) conference in November 2021.
This work has applications not only for personal assistants but also for systems that are designed to summarise information, such as the box at the top of a Google page that summarises search results. Automated systems that need to answer questions would also benefit from this work.
“We want to enable more natural conversations between people and dialog systems,” said Harsh Jhamtani, the paper’s first author.
“Most people don’t know the rules of conversation, so they often use figurative language. This research is about building systems that can understand these phrases and respond in a way that humans will find natural too.”
Jhamtani’s own struggles with figurative language inspired the research. He is a native Hindi speaker who can also communicate in English, India’s second official language. However, he had to learn the many idioms and metaphors used by his colleagues in the United States.
The researchers took conversations from a dataset of 13. They then extracted the conversations that included figurative language and ran the systems through those only.
The performance dropped by between. Researchers also wrote a script that allowed systems to quickly check dictionaries containing translation for such phrases, which is faster and more efficient than re-training them with the complete content of these books.
15 Percent Increase in Conversions, Experiments showed that this method improved performance by up to 15%.
However, before the text can be converted, the researchers must rely on human observers to identify figurative language within the dataset. Further study is needed in this area.
Read More: BotDistrikt: Creating a Successful Digital Business from the Ground Up