Aspect Level Sentiment Analysis For Intelligent Tourism

In recent years,with the increasing demand of people's travel,tourism gradually occupies an important standing in the social economy.The advent of the information age has injected new vitality into the tourism industry.It is a very forward-looking research direction to apply advanced information technology to the tourism industry to provide personalized and high satisfaction services for tourists.Accommodation is an important part of the tourism.After Intelligent terminals have gradually become an important tool in people's convenient life,tourists are more inclined to reserve hotels through various tourism websites or applications as well as leave text comments after consumption to express the check-in experience.However,this kind of data is often as high as TB even PB level containing complex information,which is difficult for tourists to obtain the effective information they are interested in.This thesis aims to use the deep learning methods to filter useless noise information from the massive text review data and mine the quality score of hotel services and provide suggestions for tourists in choosing hotels.The main work includes the following four parts.1.The web crawler technology is studied and the web page structure of a large tourism website is analyzed.The corresponding data crawling scheme is developed by using Scrapy and Selenium automation tools.Then the data is stored in SQL Server database to build a large hotel comments data set;2.An aspect extraction algorithm based on unsupervised attention mechanism is proposed.The algorithm prepares positive and negative sample data for each aspect category,and uses unsupervised attention mechanism model to run multiple binary classification tasks so that the model can fit the features of each aspect category to allocate higher attention weight for the aspect terms.The aspect term vocabulary obtained from the model is used as the seed vocabulary of the dictionary,and then the vocabulary is expanded in the word vector space obtained by the model.Finally,the effectiveness of this model is validated by the model comparison experiments.3.An interactive attention-based convolutional GRU for aspect level sentiment analysis is presented.In this model,position information and Part of Speech(POS)information are integrated in the word embedding layer,so that words with different distances and different POS categories can get different weights in the model training.Then,CNN and Bi-GRU are combined to effectively extract the complex semantic and sequence features and aspect-specific information in sentences.This model also treats context words and aspect terms equally and constructs an interactive attention mechanism to supervise the generation of attention weight.The model comparison experiments on Sem Eval2014 data set and public hotel comments data set verify the significance of the model.4.According to the detailed information of the hotel and the fine-grained sentiment information,the hotel review data visualization system is developed based on RESTful API.Four modules are designed,including map display,satisfaction of customer,impression of customer and competitive analysis.The interaction between each module provides decision support for customers when choosing hotels.