Deep Semantic Representations for Language Translation

How would you translate the word "samosa" for someone who has never experienced this food item? One way would be to map the word onto names of similar foods that you think the person does know, and then explain ways in which the foods are similar and different. We are developing a system called TLC that will learn to translate food terms in this way. An understanding of food will be created in two ways. First, the system will acquire a set of data structures that capture the similarities between basic food ingredients such as sugar, eggs, or chicken. This level of representation will model aspects of the human taste and olfactory systems. Second, TLC will compile a massive collection of recipes from the Internet into a structured database. With these two sources of knowledge, the system will be able to compare foods with potentially complex underlying structures and act as a language translator. Although TLC is focused on the domain of food, the underlying data representations and algorithms can be applied to numerous other terminology translation problems.