Co-occurrence networks of Twitter content after manual or automatic processing. A case-study on “gluten-free”

Highlights

• Co-occurrence networks of tweets after manual coding and just cleaned were compared.

• Cleaning and coding text provided networks with similar structure and terms relevance.

• Most tweets on gluten free mention products: bread, cake, cookie, beer, and pizza.

• Users share how to get gluten-free products (buying or preparing) or eating situations.

Abstract

Gathering information from social networks such as Twitter has emerged to obtain spontaneous and direct opinions of users about a topic. This study focuses on using co-occurrence networks to analyse Twitter information. The objectives were to study the impact of text pre-treatment (codification based in qualitative analysis or just pre-cleaning) and to apply co-occurrence networks for analysing what is said on Twitter about specific topics like “gluten-free”. As such, 16,386 tweets in Spanish containing terms “sin-gluten” and “gluten-free” were collected. A subset of 3000 tweets was used to make co-occurrence networks two ways: i) from the manually coded text and ii) from pre-cleaned text. Results indicate that the co-occurrence network from pre-cleaned text provides meaningful information showing structure and relevance for terms like the network from coded text. The whole set of tweets was used to explore Twitter information on gluten-free, showing users share information about products, occasions, social situations, and places but also product characteristics, sensations, and diet or health issues related to the products. Five product categories, critical for the lack of gluten (bread, cake, cookie, beer, and pizza), occupied most tweets, and according to the related terms, were intended to recommend how to get (buying or cooking) these gluten-free products and to exhibit what (how, when, and where) they prepare and eat. These aspects were different among products, and separated co-occurrence networks allowed better identification.

Introduction

In recent years, an increase in consumer demand has been observed for gluten-free products (Christoph et al., 2018, Missbach et al., 2015, Molina-Rosell, 2013). Research on gluten-free products has focused on strategies dealing with the negative impact a lack of gluten has on the quality properties of these products. Manufacturing gluten-free cereal products is a challenging task for the food industry (Capriles et al., 2016, Houben et al., 2012). Besides, according to Naqash, Gani, Gani, and Masoodi (2017), most approaches include the addition of functional ingredients to the formulation (gluten-free flours, starches, hydrocolloids, proteins, fats, and fibres) or the adoption of alternative processing methods (high pressure, extrusion, and sourdough fermentation) to produce gluten-free products with good sensory quality, especially a texture comparable to those containing gluten (Marston et al., 2016, Matos and Rosell, 2012, O’Shea et al., 2013, Penjumras et al., 2019).

However, according to do Nascimento, Fiates, and Teixeira (2017), consumers concerns for gluten-free products include sensory quality of products and the issues they experience trying to have a “normal life”, especially in a social context. Still, information on the relevance of extrinsic properties of products, context aspects, and individual attitudes and opinions of gluten-free consumers is scarce. This is for the difficulty in finding coeliac participants for consumer studies, accounting for an estimated 1–2% of population (Sapone et al., 2012). Therefore, we believe what is said on social media networks may be a way to obtain opinions of this target group of consumers, allowing us to understand their motivations and interests when consuming gluten-free products.

Among social media platforms, Twitter is one of the most popular and dynamic microblogging services, with 500 million text-based messages, called “tweets”, generated by active users per day (Chae, 2015, Da Silva et al., 2014, Mention, 2018, Vidal et al., 2015). The informal and colloquial nature of tweets, together with the ease and instant access of the platform make its use widespread, giving rise to a huge volume of rapidly generated data (Fried et al., 2015, Moe and Schweidel, 2017). Unlike other opinion gathering methods for consumers (surveys), social media users spontaneously post what they want when they want, avoiding forced biases to express their opinion.

Food represents one of the key themes discussed on Twitter (Platania & Spadoni, 2018) and consequently, tweets are potentially valuable data sources for gaining insight on food-related consumer studies. To date, the exploration of user-generated content on Twitter has been useful to study food-related topics (food in general, influence of food choices, language of food, food chains, health food, different eating situations, and emotional responses to food and beverages) (Chen and Yang, 2014, Fried et al., 2015, He et al., 2013, Platania and Spadoni, 2018, Samoggia et al., 2019, Vidal et al., 2015, Vidal et al., 2016). However, no study has addressed the exploration and interpretation of a topic like gluten-free.

Different approaches have been made to analyse tweets; automatic word counting is the simplest method of gathering information from users. Calculating the frequency or occurrence of mentions for an individual word, is simple and rapid for summarising the text according to the terms that are frequently mentioned. Nevertheless, the frequency of occurrence of individual words has several important limitations. It may not represent the meaning of the word isolated in the dataset and can lead to misleading conclusions because of the loss of the words’ context (Hsieh and Shannon, 2005, Vidal et al., 2015, Zhao et al., 2013). Therefore, previous qualitative analysis of tweet contents, with individual reading, was proposed to analyse tweets in the context of which the words are mentioned. Thus, the content was classified into themes and sub-themes related to the specific topic (Nguyen et al., 2019, Platania and Spadoni, 2018, Samoggia et al., 2019, Vidal et al., 2015). Although implementing manual content analysis can be tedious and time consuming for the large amounts of text to be read, it proved to be successful at gaining better interpretation of Twitter content (He et al., 2013, He et al., 2017, Vidal et al., 2015). As an automatic alternative, text analysis based on machine learning algorithms has been used to extract meaningful information from the textual data, recording themes already established or commonly studied (Constantinides and Holleschovsky, 2016, Sengupta and Ghosh, 2020, van Zoonen and van Der Meer, 2016). However, for the correct performance of these models, machine learning algorithms usually require a large external source of coded dataset to analyse the text units (Vidal, Ares, & Jaeger, 2018). Thus, the development and adjustment of the algorithms for new topics is complex or require added information.

Co-occurrence networks have been proposed as an approach to facilitate the understanding and visualisation of the structure of different text items and their content. Co-occurrence networks graphically represent the relevance of terms and the relatedness among them, identifying and displaying patterns of co-occurrence within the text (Ruiz and Barnett, 2015, Su and Lee, 2010). Although broadly applied in studies of bibliometric analysis to identify and visualise the existing connections among data (Skaf et al., 2020, van Eck and Waltman, 2018, Wen et al., 2017), co-occurrence networks can also be used for exploring connections of terms in different text documents.

Co-occurrence networks can be obtained by specific software as VOSviewer and Gephi or by using the Python programming language. In the VOSviewer software used in this study, the construction of a map comprises three steps: i) A similarity matrix (association strength as a measure of similarity) is obtained from a co-occurrence matrix (van Eck and Waltman, 2007, van Eck et al., 2006). The similarity between two terms is calculated as the ratio: the number of co-occurrences of two terms i and j divided by the product of the total number of co-occurrences of i and the total number of co-occurrences of j. ii) The visualisation of similarities (VOS) mapping technique constructs a two-dimensional map in which the items are located in such a way that the distance between any pair of items reflects their similarity. The base for doing so is minimising a weighted sum of the squared Euclidean distances between all pairs of items. The higher the similarity between two items, the higher the weight of their squared distance in the sum. iii) The obtained map is translated, rotated, and reflected to obtain consistent results (always the same map) regardless of the different solutions that can be reached in the optimisation process.

In the obtained network, the size of the label representing a term is proportional to its frequency of appearance in the text (occurrence). The thickness of the line connecting two terms indicates how often they co-occur within the same text unit. The distance between two terms offers an approximate indication of the relatedness of the terms (Cunillera and Guilera, 2018, Marinho et al., 2017, Sharma et al., 2018, van Eck et al., 2010). A dataset of 70 text documents describing flowers has been created to illustrate the explanation (Fig. 1). The table in the figure includes the occurrences and co-occurrences of the seven terms. Each term is represented in the network as a circle of size proportional to the number of occurrences; for example, the labels and circles of terms red colour and pink colour are the largest and the smallest because they are the most and least mentioned terms, respectively. Distribution of terms on the map responds to the relationships between items. Spring was a general term, co-mentioned with many terms (red, rose, poppy, and jasmine) and thus, appears located in the centre. The links with the four terms have the same thickness because the number of co-occurrences is the same (five). Terms that do not show co-occurrences among them are separated in the extremes, and close to the terms with higher co-occurrence. In the top-left appears the term poppy related to red, while in the bottom right, the term jasmine relates to fragrance. Rose shows a strong link with red and fragrance, and appears in the bottom-left. The term pink links to rose but does not show co-occurrence with any other term, appearing separated on the bottom-left extreme of the network.

In this study, we propose using co-occurrence networks as a tool for analysing terms in tweets to give more structured information than word counting. Using of raw text directly from tweets would make the analysis almost automatic, however, a previous qualitative analysis of tweets and the corresponding coding of text could be necessary to provide the relevance of ideas expressed in many different terms to avoid misunderstanding of the text.

Therefore, the first aim was to study how pre-processing of tweet text (coding through qualitative analysis or just pre-cleaning) influences co-occurrence networks to determine if the process can be automated without losing relevant information. The second aim was to analyse tweets about “gluten-free” to gather information about the aspects that are relevant for this specific group of consumers, in general and in relation to specific products.