A Smart Assisting System is widely used to help users make a better decision based on a combination of user interest, location awareness and the ability to access information from various information sources (such as user schedules, traffic congestion, etc.). Language is the most natural and effective way for communication of user and the system. Natural Language Generation (NLG) is a critical component of smart assisting systems, and it has a significant impact on usability and perceived quality. A classic problem in NLG involves taking structured data as input and producing text that adequately and fluently describes this data as output. The form of NLG is typically considered to require addressing two separate challenges: what-to-say: to select essential data from an unorganized data; and how-to-say: generating natural sentences from the selected data.
In terms of what-to-say, a system needs to learn how to be context-aware, which means that the system should assist differently depending on user intention, current location, or time. This requires time-sensitivity, geolocational awareness, and integration with other real-time metrics. In this thesis, we introduce a strategy for selecting relevant information from various data sources.
In terms of how-to-say, most of the current state-of-the-art NLG approaches belong to a family of the encoder-decoder model. The model encodes a meaning representation into a fixed-length vector by the encoder and generates suitable description using the decoder module. Here, decoder is the crucial part because it has to learn all diction, grammar, and other related linguistic knowledge. Besides, these difficulties become apparent when the model is generating a long sentence with a complex structure. To address the issue, we propose a hierarchical decoder, which incorporates grammar knowledge to produce more fluent description.
Experiments show that using additional linguistic knowledge in our proposed framework improves performance over a vanilla encoder-decoder NLG system. Finally, through our two-fold strategies of optimizing both what-to-say and how-to-say, we can realize a more fluent NLG system, which in the end improves the engagement of a smart assisting system.