Automated Text Analysis Based on Skip-Gram Model for Food Evaluation in Predicting Consumer Acceptance

Abstract

The purpose of this paper is to evaluate food taste, smell, and characteristics from consumers' online reviews. Several studies in food sensory evaluation have been presented for consumer acceptance. However, these studies need taste descriptive word lexicon, and they are not suitable for analyzing large number of evaluators to predict consumer acceptance. In this paper, an automated text analysis method for food evaluation is presented to analyze and compare recently introduced two jjampong ramen types (mixed seafood noodles). To avoid building a sensory word lexicon, consumers' reviews are collected from SNS. Then, by training word embedding model with acquired reviews, words in the large amount of review text are converted into vectors. Based on these words represented as vectors, inference is performed to evaluate taste and smell of two jjampong ramen types. Finally, the reliability and merits of the proposed food evaluation method are confirmed by a comparison with the results from an actual consumer preference taste evaluation.

Figure 1

An example of one-hot representation method.

Figure 2

Overview of the proposed framework for automated text analysis in online food reviews.

Figure 3

Continuous skip-gram model [7].

Figure 4

Example of hierarchical softmax.

Figure 5

Difference in smell between two jjampong ramen types.

Figure 6

Difference in taste between two jjampong ramen types.

Figure 7

Cluster structure of correlations between typical characteristics and noun words in jjampong ramen A.

Figure 8

Cluster structure of correlations between typical characteristics and noun words in jjampong ramen B.

Figure 9

Perception maps of relation between typical characteristics and taste.

Algorithm 1

Algorithm for taste and smell analysis.

Algorithm 2

Algorithm for relationship analysis.