In an attempt to anticipate Food Security (FS) crises and overcome the limits of existing early warning systems, predictive models can forecast risk indices by combining heterogeneous data. While using diferent data sources (e.g., satellite imagery, agroclimatic data, food prices) allows to consider various factors that may impact food crises, the explainability of these models remains challenging. In this work, we propose a Food Security indicator solely based on textual data, discerning among diferent triggers and accounting for possible biases in the spatial coverage of news. We evaluate our approach on a corpus of French-language documents from Burkina Faso and highlight its significance, paving the way for more open and explainable data sources for monitoring food insecurity.
CEUR ceur-ws.org
Monitoring food security status at an appropriate spatio-temporal level is crucial to detecting early
deterioration of food availability and accessibility. Significant progress has been achieved by
standardizing methods and indicators to quantify food insecurity [
In this work, we propose an analysis of the explanatory capacity of textual data. We overcome diferent limits of the first version of a Food Security indicator fully based on textual data, namely -
[
Online news are retrieved by scraping web pages from a list of online sources covering the actuality of
Western African countries (e.g. LeFaso.net, Burkina24). We applied two filters to ensure the selection
of a corpus with a relevant spatial and thematic focus. First, we extracted all spatial entities using a
pre-trained transformer-based model dedicated to French textual data, camemBERT [
As food crises are multifactorial, approaches to explicability must make it possible to distinguish
between the diferent drivers (or triggers). To identify them in the text, we adapted an expert-built
vocabulary about food security [
The food insecurity indicator −
, for a category of triggers t in the province p is: − , = 1 −
1 , + 1 , = , + , (1) (2) where
, represents the number of news articles assigned to the province p in which a crisis trigger was detected. , is 1 if at least one trigger alert was detected in the news assigned to the province p, and 0 otherwise. We added this binary variable to account for the decorrelation between some provinces’ food security situation and media coverage, where an approach based solely on the number of articles would risk biasing the results.
The final text-based food insecurity indicator in province p is defined as the weighted average of the food insecurity indicators of each category of triggers: − =
∑=1 −
, ⋅ ∑=1 + { 0 if ∀, ≠ 0 otherwise where is the weight associated with the category of trigger t, is the number of categories of triggers, is a weight, ranging from 0 to 1, that is added if all weights are non-zero for the province. This allows for the consideration of the synergistic efect of the coexistence of these diferent factors. In the preliminary analysis, we considered an equal contribution from all categories of trigger weight ( = 1). The synergy factor W was set to 0.5.
From the 15.844 online news published by the selected sources in 2022, 1773 contained at least one location associated with Burkina Faso. Among them, 309 online news were considered relevant (225 1Training parameters: learning_rate=2e-5, num_train_epochs=5, and weight_decay=0.01 were manually labelled to train the classifier, and 84 were automatically labelled). On the training dataset, the classifier obtained an F-score of 0.87 (weighted mean taking into account the class imbalance). The ifnal dataset of online news that were mapped to at least one province consisted of 195 news.
To enable a comparison of − with the IPC scores, we split the − values into quartiles (Figure 1). The indicator is elevated in several provinces from the Sahelian region, western and eastern provinces and the capital province (Kadiogo). The alert triggers allow us to contextualize the indicator values, especially in the provinces with discrepancies with the IPC value. For instance, the Kompienga province had a high − score but was only considered stressed according to the IPC. The corpus contains several references to local situations, in cities such as Pama and Madjoari, particularly afected by the consequences of civil insecurity, as shown in the examples below: ”Some populations, particularly in the Burkina Faso towns of Pama, Mansila, Kelbo, Madjoari and Djibo, are facing dramatic situations.”
This example highlights the importance of considering the diferent locations mentioned in the articles to enable analysis of local situations. It also points out that a direct comparison between local situation assessment and the IPC, produced on a provincial scale, is not necessarily relevant. (a) − , discretized by quantile (b) IPC (as of 2022-06-01)
and IPC values per province. (a) − discretized by quantile (2022) (b)
In this work, we proposed an indicator that explicitly takes into account the major triggers of food insecurity, enabling the detection of signals even in areas little covered by the press. The resulting maps complement those obtained with household surveys and machine learning-based indicators. This improved indicator contributes to food insecurity early warning and complements existing ones by providing insights into local news.
In the proposed approach, we simplified the temporal aspect by aggregating the data on a yearly scale. However, the driver’s efects on food crises can be long-lasting (e.g. a drought afecting the productive season, conflicts leading to structural production issues, etc.). Assessing the full impact of the triggers may require a better representation of their spatio-temporal footprint, by combining features extracted from the text as well as expert knowledge about the diferent types of features.
During the preparation of this work, the authors used Grammarly for English grammar and spelling check.