ash-shar / Verified-Summarization

Going Beyond Content Richness: Verified Information Aware Summarization of Crisis-Related Microblogs

Verified Summarization

If this code is helpful in your research, please cite the following publication

Ashish Sharma, Koustav Rudra, and Niloy Ganguly. "Going Beyond Content Richness: Verified Information Aware Summarization of Crisis-Related Microblogs." Proceedings of the 28th ACM International Conference on Information and Knowledge Management. ACM CIKM, 2019.

Initial Steps

Dataset

• We use the dataset created by Zubiaga et al, 2017 which can be downloaded from here.

POS TAGGING

• Use GATE Twitter Part-of-Speech Tagger

• Command:

$ java -jar twitie_tag.jar models/gate-EN-twitter.model $input_file > $output_file

• $input_file: File with each line containing a tweet (only text - space separated words)
• $output_file: space separated _ for each tweet

Processing Corpus File

Command:

$ python create-corpus-file.py

Configure the following input variables inside the code:

• src_data_path: Folder containing files of Source Tweets. In every file, a line contains - DateTime, TweetId, UserId, Tweet Text, Rumor Tag (tab-separated)
• rep_data_path: Folder containing files of Reply Tweets. In every file, a line contains - DateTime, TweetId, UserId, Tweet Text, Source Tweet Id (tab-separated)
• src_pos_tag_path: Folder containing Source POS Tags files. In every file, a line contains - TweetId, Output of Twitie (space separated _) (space-separated)
• rep_pos_tag_path: Folder containing Reply POS Tags files. In every file, a line contains - TweetId, Output of Twitie (space separated _) (space-separated)
• tentative_path: File containing LIWC list of tentative words
• certain_path: File containing LIWC list of certainty words
• negate_path: File containing LIWC list of negative words
• question_path: File containing list of question words

Configure the following output variables inside the code:

• corpus_path: Corpus file (will be input to the topic model). Each line in the file contains: TweetId, Content-Words (space-separated), Expression-Words(Space-Separated), TweetType(S/R), Time(0-1)

Content-Expression Topic Model (CETM)

Compile:

$ g++ -std=c++11 topic-model.cpp -o model

Run:

$ ./model K T iter

where:

K:    Number of Content Word Topics
T:    Number of Expression Word Topics
iter: Number of iterations to run the model for

Configure the following input variables inside the code:

• corpus_path: Corpus file created in the previous step (preprocessing corpus file).

Configure the following output variables inside the code:

• destpath: Folder where all the output files will be stored

Description of files inside the destination folder is as follows:

• c-vocab-mapping.txt: Content words to indices mapping.
• e-vocab-mapping.txt: Expression words to indices mapping.
• behavior-mapping.txt: Tweet Type to indices mapping.
• topic-priors.txt: Prior probability of content topics.
• expression-priors.txt: Prior probability of expression topics.
• c-topic-word-distribution.txt: Content Topic to Word Distribution.
• e-topic-word-distribution.txt: Expression Topic to Word Distribution.
• topic-behavior-distribution.txt: Topic to Behavior Distribution.
• table-assignment-status.txt: Status of Data points seating.
• top-c-topic-words.txt: Top 20 words in each content-word topic.
• top-e-topic-words.txt: Top 20 words in each expression-word topic.
• e-topic-time-alpha.txt: Expression-Topic-Time Alpha values.
• e-topic-time-beta.txt: Expression-Topic-Time Beta values.
• c-topic-time-alpha.txt: Content-Topic-Time Alpha values.
• c-topic-time-beta.txt: Content-Topic-Time Beta values.

Computing Tweet Posteriors

Command:

$ python compute-posteriors.py

Configure the following input variables inside the code:

• basepath: Folder created by topic-model.cpp
• CORPUS_PATH: Corpus file created.

Configure the following output variables inside the code:

• POSTERIOR_PATH: File where posteriors (probability vectors) for each tweet will be stored.

Verified Tweet Detection using Tree LSTM

Generating Trees

Command:

$ python generate-trees.py

Configure the following input variables inside the code:

• datapath: The original dataset folder (download from here)
• feature_path: File containing input feature vectors for all tweets in the dataset. The file contains two tab-separated columns - tweet_id, features
• output_path: Path of the folder where you want the generated trees to be stored

Each tree is stored as a dictionary. A sample tree and the corresponding stored dictionary is shown below:

tree = {
        'f': [0.234, .... , ], 'l': [0, 1], 'c': [
            {'f': [0.109, ... , ], 'l': [0, 1], 'c': []},
            {'f': [0.712, ... , ], 'l': [0, 1], 'c': [
                {'f': [0.352, ... , ], 'l': [0, 1], 'c': []}
            ]},
        ],
    }

Here, f is the input feature vector for each node of the tree, l is the true label of the root of the tree stored as a 2-dimensional one-hot vector (dim-1: verified, dim-2: unverified), and c is the list of children of a node.

Training and Testing Tree-LSTM

Command:

$ python train-Tree-LSTM.py

Configure the following input variables inside the code:

• tree_path: Path to the folder containing generate trees (output_path of the last step).
• IN_FEATURES: Size of the input feature vectors
• NUM_ITERATIONS: Number of iterations for training
• BATCH_SIZE: Batch size for training
• test_set: Disaster events on which you want to test.

VERISUMM

Code coming up soon!!