Consumer Price Indices (CPIs) are one of the major statistics produced by Statistical Offices, and of crucial importance to Central Banks. To calculate CPIs, statistical offices collect a large amount of individual prices of goods and services. Nowadays prices of many consumer goods can be obtained online, enabling a much more detailed measurement of inflation rates. One major challenge is to classify the variety of products, from different shops and languages into the given statistical schema consisting of a complex multi-level classification hierarchy - the European Classification of Individual Consumption according to Purpose (ECOICOP) for European countries, since there is no model, mapping or labelled data available. We focus in our analysis on food, beverage and tobacco which account for 74 of the 258 ECOICOP categories and 19 % of the Euro Area inflation basket. In this paper we build a classifier on web scraped, hand-labeled product data from German retailers and test the transfer to French data using cross lingual word embedding. We compare its performance against a classifier trained on the single languages and a classifier with both languages trained jointly. Furthermore, we propose a pipeline to effectively create a data set with balanced labels using transferred predictions and active learning. In addition we test how much data it takes to build a single language classifier from scratch an if there are benefits from multilingual training. Our proposed system reduces the time to complete the task by about two thirds and is already used to support the analysis of inflation.
Paper: https://aclanthology.org/2020.ecomnlp-1.3/
