In the digital era, messaging applications have become an integral part of daily life. One increasingly popular feature is Smart Reply, which enables users to respond to messages quickly without typing. However, this feature faces challenges when the generated responses tend to be repetitive or overly similar, providing little meaningful variety. In the context of Bahasa Indonesia, this challenge becomes more complex due to differences in grammar and cultural nuances that are not always effectively captured by technologies originally designed for other languages.
Our recent research offers a solution with an innovative method to enhance the diversity of smart replies in Bahasa Indonesia. This method combines text classification with post-processing to generate responses that are more varied and relevant. A key advantage of this approach is its ability to address the unique challenges of Bahasa Indonesia, including handling sensitive information such as phone numbers and addresses.
The developed method uses a two-step approach. The first step generates initial responses using a Long Short-Term Memory (LSTM) model, which analyzes incoming messages and predicts the most relevant replies. However, since these initial responses are often uniform, the second step becomes crucial. In this step, the system refines the responses using Term Frequency-Inverse Document Frequency (TF-IDF) to identify keywords and K-means clustering to group responses based on their meaning. From each cluster, only the best response is selected, ensuring more diverse options.
Additionally, this method includes a step for removing sensitive information. If a reply contains phone numbers or addresses, the system automatically deletes it and replaces it with a safer option. This step not only enhances relevance but also ensures user privacy is maintained.
Testing results show that this method significantly improves the diversity of smart replies for real conversations in Bahasa Indonesia. For instance, for a message like “sudah di depan rumah” (“I’m in front of the house”), conventional methods tend to produce similar replies such as “saya segera ke sana” (“I’m on my way”). In contrast, our system can offer more varied responses, such as “sebentar ya” (“Just a moment”) or “di mana tepatnya?” (“Where exactly?”). This approach is also effective in handling common replies like “terima kasih” (“thank you”), where only the best variation is selected to prevent repetition.
This research also has broader implications in supporting the achievement of the Sustainable Development Goals (SDGs), particularly SDG 9 (Industry, Innovation, and Infrastructure), by providing smarter and more inclusive technology.
Keywords: Digital Era, LSTM, Bahasa Indonesia
Reference:
Haidar, M.A., Bahar, S., Susyanto, N. (2024). Enhancing the diversity of Smart Reply suggestions: A novel approach combining text classification and post-processing techniques for real conversations in Bahasa Indonesia. Mathematical Modelling of Engineering Problems, Vol. 11, No. 9, pp. 2566-2572. https://doi.org/10.18280/mmep.110927
Author: Nanang Susyanto