Back to search

Agricultural University of Athens

Ανάλυση μεγάλου όγκου δεδομένων αγροτικών ψηφιακών αγορών

Abstract

dc:description

Nowadays, a lot of companies aim to tap into social media networking in order to maximize their profit by endorsing their products or services and for improving their brands’ names. The development of Web 2.0 has permitted Internet users to post, share and exchange their own self-generated opinions or thoughts on various topics on different websites. A large amount of data containing useful information concerning the consumers’ preferences is generated from a variety of sources such as reviews, posts, microblogs or online digital markets. In the case of Food and Beverage (F&B) sector, more and more review websites are established globally (Yelp, TripAdvisor, etc.) and most of them allow users to digitally make their own orders for delivery, or take away goods (just eat, etc.), as well as to digitally evaluate about the products or services that they have consumed. The produced information (evaluations) which is generated rapidly can be large and generally modifies consumers’ behavior. It is also notable that the e-evaluations could be characterized as mixed-documents, because users evaluate different companies’ functions at the same time. However, in most of the cases, the involved stakeholders (companies or customers) cannot mine and analyze in real time the useful information that is created due to the lack of resources and humans’ physical or mental restrictions.In bibliography, there are various methodologies to face this problem, with the sentiment analysis being the preferred one. It is presented as the automatic detection of positive or negative expressions in user generated content. Also, in research there are two approaches: the machine learning approach and the semantic orientation approach and three levels of analysis: the document-level, the sentence-level and the aspect-level. A lot of frameworks have already proposed that use machine learning techniques, while the semantic orientation approach has not received the same interest yet. However, some previous researches concluded that the semantic orientation approach may be more efficient when implemented for classifying users’ opinions and for determining the polarity and the strength of opinions at the same time. Finally, while a lot has been written and researched about sentiment analysis in various domains and languages too, the F&B sector in the Greek language has drawn limited researchers’ attention. This study faces this problem by making a suggestion of a new hybrid framework of sentiment analysis using the semantic orientation approach. A thorough analysis of 91,504 customers’ reviews (9,150 reviews in the training set) collected from an e-ordering platform of some F&B companies distributed in almost all prefectures of Greece is occurred. After, the implementation of some well-known models and metrics in the training set (BOW, tf-idf and POS) resulted in the creation of an opinion dictionary concerning the F&B domain. The dictionary consists of three sub-dictionaries that correspond to three companies’ functions (quality of food, customer service and image of the company). Further, after the implementation of the PMI metric in the training set, the relationships between the parts of the speech and their positions in mixed customers’ e-valuations are identified and used for constructing the proposed patterns of tags. The patterns of tags are then used for detecting and quantifying the customers’ opinions per function in a review. The main innovation of this framework is based on the form of the analysis, first at sentence and after at aspect-level. Also, a new hybrid technique for constructing opinion dictionaries is proposed, using all the known techniques that are presented in bibliography (manually-technique, corpus-based technique and dictionary-based technique). Finally, since the extraction of useful information from the user generated content must usually be held in real-time in order to facilitate the stakeholders to make smart decisions on-time, a web-based system that able to mine and analyze customers’ evaluations in almost real-time, is introduced. For this, two modules will be designed and implemented. The first is a web scraping module for mining the reviews from all e-ordering platforms and the second is a module for implementing the proposed hybrid framework.Based on the confusion matrix of the designed sets, the well-known performance metrics of accuracy, precision, recall, and F-score were computed. The results showed a remarkable high performance on the sentiment classification and the aspect-level analysis respectively. Specifically, the proposed hybrid framework showed an average accuracy of 98.45% in the training set concerning the sentiment classification. Moreover, it showed an average accuracy of 93.61% in the annotated set concerning the aspect-analysis. Then, for the purposes of assessing the framework in a big set of customers’ reviews, it tested in the data set and it showed an extremely high average accuracy of 90.69%. However, in all designed sets were calculated lower values in negative predictions, mainly due to the identification difficulties of negation and sarcasm, in natural language texts. In more detail, the framework showed an average recall of 85.99% in negative predictions against of 99.43% in positive predictions in the training set, an average recall of 81.49% in negative predictions against of 96.43% in positive predictions in the annotated set and an average recall of 70.81% in negative predictions against of 95.79% in positive predictions in the data set. It is noteworthy that the performance of negative prediction remains at a high level, comparable to other similar researches. The average high values of the metrics lead us to conclude that the vast majority of aspects that described the examined functions in the F&B sector with the semantic orientation terms (adjectives) that con-currently occur, were detected. To conclude, a new technique for managing the instances that cannot be detected and evaluated by the system is proposed.

Degree

thesis:*
Grantor dc:publisher
Agricultural University of Athens
Year dc:date
2020

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Liapakis, Anastasios
  • Λιαπάκης, Αναστάσιος

Subjects

dc:subject × 20

Rights

dc:rights
Statement dc:rights
  • BY_NC_ND
Language dc:language
gre

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/48307

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Liapakis, Anastasios; Λιαπάκης, Αναστάσιος. Ανάλυση μεγάλου όγκου δεδομένων αγροτικών ψηφιακών αγορών. Agricultural University of Athens, 2020. http://hdl.handle.net/10442/hedi/48307