- Firstly, open TravelTweets.py. Replace the consumer_key, consumer_secret, access_token, access_token_secret with your own keys;
- Run $ python TravelTweets.py;
- Set up Google environment and install Google API(https://github.com/googleapis/google-api-python-client);
- Wait for the process done. You should see a new folder named "result" created containing 25 text files titled with city names;
- After the folder is successfully created, run $ python languageprocess.py;
- Then you will get a text file named "result.txt" containing a list of 10 cities.
In this project, we want to know about what are the most welcome cities in the US for a short trip by implenmenting analysis of the tweets. We will give a list of cities ranked from the most positive to the most negative (or least positive) in the overall sentiment of all related text tweet content.
The project can benefit anyone who currently has a plan of traveling or want to get information for certain traveling cities in the US.
It can also provide information to traveling agencies like which cities are most and least welcome to help them design tourism products, for example, they can use this infomation in advertisement or arrange more trips to the most welcome cities.
I, as a user, want to know what are the most welcome cities in the US for a short trip.
I, as a foreigner newly arrived at the US, want to know that which cities I should start with for traveling so that I can get the best experience within the limited time.
I, as a travelling agnecy, want to know the popularity rank of the popular US cities, so that I can arrange tourism products accordingly.
The project will sort out the top 10 most favored US cities for a short trip using google sentiment analysis on the popular and real-time related Tweets.
Firstly, we want to identify several US cities that may be popular traveling destinations for people. We referred to "The 25 Best U.S. Cities to Spend a Weekend" updated by THRILLIST TRAVEL on 05/07/2019 (https://www.thrillist.com/travel/nation/best-us-cities-to-spend-a-weekend-nashville-austin-charleston-providence) and took the 25 cities as our base city pool.
Then, we extracted tweets including "#travel OR #holiday OR #vacation AND the city name" for each city. Considering we are recommending US cities, we only searched for English tweets and extracted the most popular and also real-time tweets of the first 100 pages, and excluded retweets to avoid redundancy. For the code of Tweets part, we leveraged TweetSearch library for the sake of convenience which is also included in this git project. Then we wrote each city's tweets into reparative text files so that we can process the twitters' sentiment and feeling in their tweets city by city.
In the language process program, we read the txt file one by one and use the google NLP api to analyze them. After we got the sentiment scores and magnitudes from google services, we rank this data into a list and sorted these cities by considering both the sentiment scores and magnitudes. Finally we save the result into a txt file and display analysis result in a top-down rank.
The biggest issue is the extracted tweets are quite un-organized and contains a lot of unrelated, trivial or misleading information, for example, some tweets are from traveling agencies or advertisers which will bias the final result. Besides, the total amount of tweets data is too small for a reliable result. To solve this problem, getting some priori knowledge of the tweets publishers and trying to eliminate those irrelevant ones out by setting up filters, also running analysis on the tweets of larger amount and over longer period may be helpful.
In language process module, we decided to classify and calculate the sentiment score all by NlP api. However, we found it is really hard to classify the sentences, so we classify these sentences before input them into google api. We created several files and used the cities as their names. Each file saves all comments about one city. So, we can simply process the commments file by file.
-
What you liked doing? I liked to study all the API functions provided by Twitter, which gave me a lot of initial thinking of the possible usage of Tweets analysis and methods of improve the analysis accuracy.
-
What you could have done better? I could try to combine more Twitter API functions rather than just TweetSearch to get better raw Tweet data.
-
What you will avoid in the future? We should build up runnable code withouth considering too many details and run tests as early as possible to find out problems/issues unexpected. This time, we had to change our methods from running "Entity Analysis" (Identify entities within documents and label them by types such as date, person, locations)on Tweets with "#travel" and identifying all locations included and ranking them by frenquency. However, we found out that the "location" keywords extracted by this analysis tool were not just city names but also contained too many location-related words like "home", "site", etc.. Thus, in the end we have to change our methods and re-wrote code.