The TF-IDF (Havrlant and Kreinovich, 2017) is used as a weighting criterion to measure the word vector represented by the document to calculate the importance of terms in the text. The TF-IDF algorithm is used to calculate the frequency of the candidate product feature words, reflecting the importance of the feature words. His main idea is that if a word appears in a text with high TF frequency and rarely in other texts, the word or phrase is considered to have adequate representation and is suitable for classification. Google introduced BERT in 2018, which uses an attention mechanism to identify all other word-related contexts present in a text sequence. The structure of SG is shown in Figure 2A. Sentence generator translates the source text into different languages by using Google Translate, which supports 109 languages, for i consecutive times, with the language randomly selected, and finally translates into the language of the source text. However, high i values lead to trade-offs in generating sentences that are significantly different from the source sentences. Beginning in 1942, the Soviet high command exerted better-coordinated political and military control over the partisans, whose operational significance increased from 1943. Although totaling as many as 700,000 and causing some 35,000 Axis casualties, their main impact was in the rear areas, where they disrupted communications, carried out raids and sabotage, gathered intelligence, and told the local population that the Soviet regime was coming back.
If you can get over the cognitive dissonance, it may just be the perfect travel guitar for you. Finally, contact the travel agents directly with any questions or concerns that you may have before making your final decision. Compare plans. If you already have a credit card, be sure that you’re making a good move before you swap cards. When the length of comments is too large, RNN will not be able to connect relevant information; LSTM cannot be used for parallel computing, which consumes a lot of time and space for experiments; Transformer lacks modeling of the time dimension, making the output of each position are very similar, which will eventually lead to poor performance in location classification. The specific schematic diagram is shown in Figure 1. The algorithmic process of Step 1, Step 2, and Step 3 will be described in detail. Figure 1. Sentiment analysis flowchart. Figure 2. Structure of (A) SG, (B) SD, and (C) BERT. Then, the BERT is used to classify the user sentiments in the dataset created in this paper. The work in this paper can be divided into three steps, (1) data preprocessing, (2) statistical analysis, and (3) sentiment analysis.
Considering these advantages of BERT, this paper choose BERT as the sentiment analysis model for this paper. And based on this result, this paper summarizes 10 categories and labels each comment based on these 10 categories. In present 10 comment categories about travel, the model performs sentiment analysis prediction for each comment and classifies them into Positive, Neutral, and Negative. In this paper, BERT language model is used to perform category recognition sentence-level sentiment classification (Gao et al., 2019) and sentiment analysis (Vaswani et al., 2017; Devlin et al., 2019) on the travel comments. Therefore, in this paper, an iterative translation-based text enhancement method was used (Lee et al., 2021) to enrich the original dataset and expand the 2,000 reviews contained in the dataset to 4,000 reviews, so as to improve the reliability and generalization of the experiment. 2014), and Kim and Hwang (2020) have proved that the emotion or tone conveyed in the text directly affects the usefulness of text communication or eWOM communication. There have been some studies using various deep learning methods for sentiment analysis of reviews. Although there has been a long-term development in computer and other engineering disciplines to combine deep learning methods and textual information retrieval techniques for sentiment analysis (Zhao et al., 2017; Yiwen et al., 2022), its application in social sciences including tourism, linguistics, and communication is still scarce.
The current typical neural network learning methods are convolutional neural network (CNN) (Wang et al., 2022), recurrent neural network (RNN) (Al-Smadi et al., 2018), long-short-term memory network (LSTM) (Priyadarshini and Cotton, 2021), and Transformer (Naseem et al., 2020), etc. The Bidirectional Encoder Representation from Transformers (BERT) (Devlin et al., 2019) is different from previous models in that it is a deep, bi-directional, unsupervised language representation model that can be resumed on top of the latest pre-trained context-sensitive language representation work. Given the proliferation of reviews on online travel sites and the resulting consumer impact (TripAdvisor, 2019), many scholars have made efforts to explore the relationship between online travel reviews and consumer behavior, and to what extent reviews influence the consumer’s decisions and choices (Hlee et al., 2018; Liu et al., 2019; Nguyen and Tong, 2022). Some studies tend to evaluate the content quality of online reviews. The SD structure consists of a merged CNN and a parallel structure of BiLSTM (Liu and Guo, 2019), and the input of CNN-BiLSTM can be one or two sentences. Among these, what makes a review “useful” is the universal central research question (Korfiatis et al., 2012; Li et al., 2021; Liu and Hu, 2021). Yin et al.
