Social media generate vast amounts of authentic linguistic data daily, offering insights into users' conceptualizations, language use, and cultural phenomena. However, managing this data manually is challenging due to its volume and evolving algorithms. This study identifies key concepts and specific words that shape Greek consumer identity by analyzing Instagram posts on traditional and local food consumption. Sentiment analysis was conducted using generative AI, and new computational tools were developed for automated data collection and analysis. AI-generated responses were evaluated against human assessments, and they proved to be successful in identifying and categorizing posts and hashtags according to criteria set for the task. Case studies included Mediterranean snails as a sustainable meat alternative and local cheeses from the North-Aegean islands. The methodology can be applied to specialized lexicography, linguistic and cultural studies, and terminology research, such as identifying neologisms and tracking term usage over time.
Product failure in the marketplace happens even if products show high liking scores by consumers
in lab and in-house tests. Thus, alternative methods are being used in sensory and consumer studies
to obtain data from consumers in environments outside the lab, one being social media-based
methods. Online social media networking sites, content communities, online reviews, forums, and
blogs provide a rich and expansive source of qualitative data that can be analyzed in a quantitative
manner [
Manual handling is not feasible when managing substantial amounts of data nor can it be a measure
of accuracy, but it can provide insight into the accuracy of current Natural Language Processing
(NLP) tools [
A wide range of programming languages support reading and processing of collected data, such as Python.
Sentiment analysis is an analysis method also applicable in food research on social media.
Sentiment analysis is the computational study of people's opinions, emotions, and attitudes towards
entities, and topics [
The aim of the present study is to understand how geographical factors, namely the country or place of origin of a food product or a recipe, affect consumers’ response and emotions to foods, using a novel methodology that incorporates NLP models. The main objectives are a) to identify the way consumers feel and talk about traditional and local foods as opposed to relevant novel ones on social media, b) to investigate how sensory attributes, geographical characteristics, nutritiousness, and environmental concerns impact consumers’ food choice, and c) to identify relevant concepts. The methodology developed was evaluated in the case studies of a) snails, a traditional food for the Mediterranean market and a sustainable alternative to meat, and b) local cheeses of the NorthAegean Sea islands.
The present study is original as regards the methodology and tools applied. More specifically, a new methodology has been developed for the mining and handling of data on Instagram for specific language related purposes, combining existing programming and NLP tools in an original way to decrease the need for manual data handling. In addition, ChatGPT is used for automated application of criteria as a substitute for human to save time and ensure repeatability of results, and, to compensate for the lack of geotagging information provided for Instagram posts, a geographical tag (using the prefectures of Greece) based on relevant hashtags has been added to the respective posts.
Instagram was the social media platform chosen for this study because Instagram users interact with
companies more often than on other platforms in Greece, and cooking comes second (together with
health/ fitness) among the most common interests of Greek Instagram users [
For data collection, Apify [
Posts from April 2012 (Instagram platform release) to March 2023 were collected. Posts were merged into a single file and duplicates were removed. Data obtained for each post contained: post id, type of post [image, sidecar (i.e., group of images), video], shortCode (shortened URL of the post), caption, hashtags, number of comments, number of likes, timestamp, and whether it appeared on a professional account or not. The data were further cleaned by removing hashtags that belong to languages other than Greek, Greeklish, or English, nonsensical words, parts of speech that were not meaningful for the present study (pronouns, articles, and prepositions), names of businesses (producers, sellers, restaurants), and the original hashtags that had been used for data collection.
In addition to Apify, a set of Python scripts was developed specifically for this study to automate the cleaning, merging, and filtering of Instagram posts, as well as to preprocess hashtag metadata and detect Greeklish variations. These scripts also allowed seamless interaction with large language models (LLMs), enabling automated application of classification criteria (e.g., food-relatedness and sentiment) and structured storage of model responses. This toolset supported reproducible workflows and minimized manual effort across multiple stages of data handling and analysis.
ChatGPT, a novel Artificial Intelligence application (edition 3.5), was used for the automated application of the food-relatedness criterion, so that only food-related posts would be left for further analysis. ChatGPT was chosen because: a) it can process data in multiple languages, b) it can automate the processing of data in Python, and c) it is low cost. The instruction was: “Process only food related posts”. The linguist of the research team subsequently performed a manual check regarding food-relatedness. This was done to evaluate the degree of agreement between human and machine responses and explore the potential of the machine in substituting manual data management to save time and effort.
ChatGPT checked the posts and provided a response (i.e., positive, neutral, negative) for each post regarding sentiment, instructed to take into consideration the following components of each post: caption, emojis/ emoticons, hashtag(s). The linguist of the research team subsequently performed a manual check regarding sentiment. As previously described, this was done for comparison purposes. The posts were grouped per sentiment category (positive, neutral, negative) and per hashtag.
Instagram does not offer post-geotagging information due to personal information protection reasons. Since locality was one of the focus points of the study, a region tag was added to those hashtags that were names of a city, island, or area of Greece, using the thirteen prefectures of Greece as reference.
Both ChatGPT and human identified 44% of the posts as food-related and 47% of them as not foodrelated, thus reaching an overall agreement rate of 91%. Beyond-chance agreement was confirmed using kappa test. Additionally, sentiment analysis revealed high agreement between ChatGPT and human evaluations, with 61% of the posts identified as positive and 37% as neutral, resulting in an overall agreement rate of 98%, beyond chance. Quantitative evaluation metrics also confirmed the effectiveness of ChatGPT in both classification tasks. For the food-relatedness classification, ChatGPT achieved a precision of 92.4%, a recall of 88.9%, and an F1-score of 90.6%. In the sentiment classification task, class-wise performance was also high: for the “positive” sentiment class, precision was 99.8%, recall 96.6%, and F1-score 98.2%; for the “neutral” class, precision was 94.6%, recall 99.7%, and F1-score 97.1%. No post was identified as negative by either human or AI. However, ChatGPT’s performance in identifying “Greeklish” was unsatisfactory and inconsistent. More specifically, when instructed to categorize hashtags according to language (Greek, English, Greeklish), in order to subsequently transcribe Greeklish into Greek for further analysis, the application consistently identified Greeklish as English. In an attempt to rectify that, it was provided with a list of hashtags in Greeklish matched to their Greek equivalents and instructed to categorize the original list again based on this input. However, the new output was again erroneous, identifying Greeklish as either Greeklish or English. Every time it was asked to perform the same task the categorization of the same hashtags was different. As a result, the task was performed manually.
Certain key concepts related to the foods under study were identified in the collected posts. Firstly, concepts related specifically to the sensory aspect of eating were taste (the sense), tasting (the act), tastiness (the food quality), and gourmet eating. They were found as hashtags, such as #taste, #winetasting, #eat, #slurp, #wow, #delicious, #tasty, #tastyrecipes, appearing with high frequency which probably means that traditional and local food consumption closely correlates with sensory satisfaction. Secondly, concepts referring to nutritional value (e.g., #protein, #omega3, #organic, #vitamins) and dieting styles (e.g., #keto, #vegan, #glutenfree, #eatclean, #healthylifestyle, #vegetarianfood) were identified but did not appear with high frequency. This may mean that nutritional content is not the focus or goal when consuming or posting about traditional and local foods. Thirdly, hashtags pertaining to the environmental aspect of food production, such as #cleanandgreen, #ethical, sustainable, #bio, #greenfood, were not frequently mentioned. This implies either that the Greek consumer has not cultivated environmental concerns to a significant degree yet, or that this aspect of food production is not relevant to traditional and local food consumption. The fact that the two concepts coappear in posts, however infrequent, suggests that the two are connected, and it was identified that it is important for the food to be locally and organically produced.
The 1836 word-hashtags collected were categorized into twenty semantic categories, using a Greek thesaurus. Hashtags that belonged to more than one category were duplicated (e.g., #tastyrecipes was categorized as post content and assessment). These semantic categories led to drawing conclusions in a more meaningful and organized way. Hashtag grouping into semantic categories was done manually in this case, but GenAI (ChatGPT and Copilot) has been subsequently assessed in food related case studies and was able to group hashtags in a meaningful way. Automated categorization is necessary to minimize time and effort.
We were able to identify key drivers of consumption, and draw the conceptual, sensory, and
emotional profiles of the foods under study. For example, as regards cheeses, intended use,
familiarity, and price are the main drivers of consumption (in the order mentioned). Local cheeses
correlate with concepts such as tradition, family, granny, summer, bread, salad, olive oil, and honey,
while non-Greek cheeses with concepts such as wine, Italy, pasta, fast food, pizza, and mushrooms.
Sensorially, local cheeses are thought to be hard and salty, while non-Greek cheeses are considered
mainly gummy, creamy, fatty, and soft. Local cheeses elicit emotions by bringing forth childhood
memories, as opposed to non-Greek ones that elicit emotions of sensuality [
As regards the snail-related study, when consuming (or posting about) traditional foods, nutrition
or dieting styles seem to be irrelevant, and the emotional, cultural, and regional aspects of the foods
are highlighted more often than taste (in any sense of the word). Traditional food consumption
relates to authenticity, health, simplicity, freshness of ingredients, homemade cooking, hospitality,
respect for the cook, passion (a.k.a. Greek meraki2), and emotions of happiness, love, care, nostalgia,
liveliness, bliss, fun, and comfort, thus depicting a link between traditional food consumption and
community in interaction. Community has always been an integral part of the Greek culture [
The present study was culture oriented and thus worked on identifying and solving problems specific to posts containing Greek, Greeklish, and English hashtags of interest. Post collection was also restricted to Instagram platform. However, the methodology and tools applied can be extended to 2 "Meraki" is a Greek word that describes putting your heart, soul, and creativity into something (e.g., work or an artistic endeavor). It conveys a deep sense of love, care, and personal investment in what one does. other languages and platforms. Challenges in data collection from other platforms, such as Reddit, remain to be identified. Certain tools, like ChatGPT, could be further tested using different settings to deliver more accurate and repeatable results, and new similar applications, like the Greek KriKri, could be assessed for the same tasks. What is more, translation of hashtags into English could be a route to be tested to ensure uniformity in the language of hashtags before further analysis. In that case, a wider variety of NLP tools would be available. An extension of this study, already in progress is the collection and analysis of data from online sources in general (reviews, blogs, magazines, etc.) to investigate the digital identity of agrifood products (sensory attributes, key relevant concepts, consumer attitudes and expectations, etc.).
On the one hand, food producing and marketing companies could use hashtag analysis, in particular the frequency of appearance and co-appearances in posts, to gain insight into how consumers view their products and brand. For example, the link identified between meat alternatives and fasting or gourmet dining could be used to promote relevant products. On the other hand, linguists, and anyone studying linguistic and cultural phenomena (food-related or not) could also use hashtag frequency of appearance and co-appearances in posts. These hashtags, grouped in semantic categories, form a network of related concept frames that depict cultural aspects of food consumption. These semantic groups can also be used to compile domain-specific glossaries (e.g., for traditional food consumption to be used in travel guides, menus, and consumer studies), or to identify concepts that correlate positively or negatively with a phenomenon under study in (cross-) linguistic and (cross-) cultural studies (e.g., to study consumer ethnocentrism). Theoretical and applied linguists can additionally extract lists of consistently misspelled words to gain some understanding of specific spelling difficulties concerning users of the languages under study. Finally, neologisms can be identified and studied regarding frequency and context of appearance, as well as use over time.
The present study explored the use of existing and the development of new LLMs to collect, manage, and analyse linguistic data for specific purposes from a social media platform. ChatGPT was validated as a tool for automated application of criteria as a substitute for human handling to save time and ensure repeatability of results. The use of Python scripts was effective and successful in analyzing data collected from Instagram. Misspelled hashtags in Greek and English, and the use of Greeklish created problems during post collection, hashtag identification, and hashtag analysis. ChatGPT was not able to identify Greeklish in a correct and consistent manner, even following attempts to train it by providing lists of Greek words with their Greeklish counterparts as feedback. Areas for improvement in NLP systems still exist, especially regarding the Greek language. The databases created using this methodology can be further populated, by collecting hashtags relevant to other traditional and local foods and in other languages.
The present study has been funded by the Greek National Development Program 2021-2025 through the General Secretariat for Research and Innovation under the call "Flagship action in sustainable agri-food systems - applied research, development of infrastructure and services for the sustainability of the sector - (Sust.Agri.Food)" [MIS code: 5201774].
During the preparation of this work, the authors used ChatGPT-4 for grammar and spelling checks. The authors have subsequently reviewed and edited the content and take full responsibility for the publication’s final version.