Character-Based Embedding Models and Reranking Strategies for Understanding Natural Language Meal Descriptions

Character-based embedding models provide robustness for handling misspellings and typos in natural language. In this paper, we explore convolutional neural network based embedding models for handling out-of-vocabulary words in a meal description food ranking task. We demonstrate that character-based models combined with a standard word-based model improves the top-5 recall of USDA database food items from 26.3% to 30.3% on a test set of all USDA foods with typos simulated in 10% of the data. We also propose a new reranking strategy for predicting the top USDA food matches given a meal description, which significantly outperforms our prior method of n-best decoding with a finite state transducer, improving the top-5 recall on the all USDA foods task from 20.7% to 63.8%.