Food & beverage intelligence for market research support

Abstract

Artifical Intelligence (AI) tools are being used more and more in business nowadays to support users in many ways. Food & Beverage (F&B) industry is continuously seeking to expand its product portfolios to enhance the products and services it offers. In order to be able to grow this portfolio, market research about F&B brands is necessary. Such market research typically takes place through specific fairs, reading specialised publications, or through access to internet web resources. It takes a lot of time and effort to stay current with all the resources mentioned. This is where artificial intelligence AI can support the previously indicated efforts through Natural Language Processing (NLP). We undertake this task by making use of transformers-based models that perform attribute value extraction to detect brand names, products and restaurant chains within the text. The extracted attribute values are presented to a different model that can classify the relationships as "owns", "produces", "makes" and "sells" relation types between the respective attribute values in the text. Two datasets, one for the attribute value extraction model and one for the relations classification model were created based on a new F&B related corpus called Food and Beverage dataset (FBdset) (F&B dataset). The newly annotated corpus is made up of 4,000 articles, 9,787 sentences and 143,012 words. Total Attribute values 15,375 (PROD: 4,313, CORPBRAND: 1,748, PRODBRAND: 4,596, RESTCHAIN: 4,718) together with 13,381 relations (SELLS: 7,645, PRODUCES: 728, MAKES: 4,034, OWNS: 974). Food and Beverage dataset (FBdset) (F&B dataset) attribute value extraction total is 7,640 (PROD: 2,884, CORPBRAND: 844, PRODBRAND: 2,212, RESTCHAIN: 1,700 ). The attribute value extraction model achieved an F1 score of 73% whilst the relations classification model achieved an F1 score of 89.5%. We have performed Information Extraction (IE) related to the Food & Beverage (F&B) industry using attribute value extraction and relations classification models. A new corpus that served as the foundation for the relation classification and attribute value extraction models’ datasets was created.