Despite its simplicity, bag-of-n-grams sen- tence representation has been found to excel in some NLP tasks. However, it has not re- ceived much attention in recent years and fur- ther analysis on its properties is necessary. We propose a framework to investigate the amount and type of information captured in a general- purposed bag-of-n-grams sentence represen- tation. We first use sentence reconstruction as a tool to obtain bag-of-n-grams representa- tion that contains general information of the sentence. We then run prediction tasks (sen- tence length, word content, phrase content and word order) using the obtained representation to look into the specific type of information captured in the representation. Our analysis demonstrates that bag-of-n-grams representa- tion does contain sentence structure level in- formation. However, incorporating n-grams with higher order n empirically helps little with encoding more information in general, except for phrase content information.

Analysis of Bag-of-n-grams Representation's Properties Based on Textual Reconstruction