Enriching word embeddings with food knowledge for ingredient retrieval

Smart assistants and recommender systems must deal with lots of information coming from different sources and having different formats. This is more frequent in text data, which presents increased variability and complexity, and is rather common for conversational assistants or chatbots. Moreover, this issue is very evident in the food and nutrition lexicon, where the semantics present increased variability, namely due to hypernyms and hyponyms. This work describes the creation of a set of word embeddings based on the incorporation of information from a food thesaurus - LanguaL -through retrofitting. The ingredients were classified according to three different facet label groups. Retrofitted embeddings seem to properly encode food-specific knowledge, as shown by an increase on accuracy as compared to generic embeddings (+23%, +10% and +31% per group). Moreover, a weighing mechanism based on TF-IDF was applied to embedding creation before retrofitting, also bringing an increase on accuracy (+5%, +9% and +5% per group). Finally, the approach has been tested with human users in an ingredient retrieval exercise, showing very positive evaluation (77.3% of the volunteer testers preferred this method over a string-based matching algorithm).