Overcoming the Long Tail Problem: A Case Study on CO2-Footprint Estimation of Recipes using Information Retrieval

Overcoming the Long Tail Problem: A Case Study on CO2-Footprint Estimation of Recipes using Information Retrieval Melanie Geiger1,2, Martin Braschler2 1 Universit´e de Neuchˆatel, Neuchˆatel, Switzerland 2 Zurich University of Applied Sciences, Winterthur, Switzerland {imhf, bram}@zhaw.ch Abstract We propose approaches that use information retrieval methods for the automatic calculation of CO2-footprints of cooking recipes. A particular challenge is the “long tail problem” that arises with the large diversity of possible ingredients. The proposed approaches are generalizable to other use cases in which a numerical value for semi-structured items has to be calculated, for example, the calculation of the insurance value of a property based on a real estate listing. Our first approach, ingredient matching, calculates the CO2-footprint based on the ingredient descriptions that are matched to food products in a language resource and therefore suffers from the long tail problem. On the other hand, our second approach directly uses the recipe to estimate the CO2-value based on its closest neighbor using an adapted version of the BM25 weighting scheme. Furthermore, we combine these two approaches in order to achieve a more reliable estimate. Our experiments show that the automatically calculated CO2-value estimates lie within an acceptable range compared to the manually calculated values. Therefore, the costs of the calculation of the CO2-footprints can be reduced dramatically by using the automatic approaches. This helps to make the information available to a large audience in order to increase the awareness and transparency of the environmental impact of food consumption. Keywords: BM25 weighting scheme adaptation, cooking recipe retrieval, CO2-footprint estimation 1. Introduction One easily measurable quantitative quality criterion of a language resource (LR) is its coverage. However, achieving a high coverage usually requires a lot of human effort. One of the reasons is that often the frequencies of potential LR entries; e.g. words in human language or food products in cooking recipes (M¨uller et al., 2012), arrange themselves according to the so-called Zipf’s law (Zipf, 1949), mean- ing that most entries relate to entities that occur very in- frequently. Therefore, LRs are most likely never complete. This long tail problem is relevant for most applications that rely on LRs, however, it is particularly severe for informa- tion retrieval (IR) applications that not only use the LR to enhance their effectiveness (e.g. expanding queries with synonyms) but directly use the LR entries to compile their output. To illustrate, consider a new class of retrieval applications that require the calculation of a single numerical value from a semi-structured item that consists of a list of textual ele- ments. For example, such a semi-structured item may be a cooking recipe in which the elements are the instruction lines, such as “100g carrots, sliced” and “1 pizza dough”, or a real estate listing in which the elements are the compo- nents, such as “Bedrooms: 4” and “Heating: Oil-Fired Cen- tral Heating”. In those examples the numerical values to be calculated can be the nutrition value or the CO2-footprint of a recipe or for the real estate example …

Content truncated to 500 words