Visual Emotional Latent Dirichlet Allocation (VELDA)

This is the implementation based on the following paper:

Tao Chen, Hany M. SalahEldeen, Xiangnan He, Min-Yen Kan and Dongyuan Lu (2015). VELDA: Relating an Image Tweet’s Text and Images. In Proceedings of the 29th AAAI Conference on Artificial Intelligence (AAAI'15), Austin, USA.

We additionally release three datasets used in the paper:

• Weibo image tweets
• Twitter image tweets
• Wikipedia Picture of the Day (POTD) dataset.

As restricted by the Weibo and Twitter policy, we are only able to release the Post IDs. For POTD, we release the dataset with extracted features.

Please cite our AAAI'15 paper if you use our code or dataset. Thanks!

Author: Tao Chen (http://www.cs.jhu.edu/~taochen)

Usage

• Change the configuration in config.init

• Run java command:

  java Main.java config.init

  If you are training with a large dataset, please allocate more memory to JVM, e.g.,

  java -Xmx2g Main.java config.init

  This code invokes the pipeline of training, testing and evaluation. The evaluation is conducted in a text-based image retrieval task. Please see the paper for detailed description.

Dataset (input) format

• Training set consists of three files:
  • p_train_text.txt
  • p_train_visual.txt
  • p_train_emotion.txt

Each line contains the textual/visual/emotional word IDs (seperated by a whitespace) for a particular document. That is, the nth lines correspond to the nth document in the dataset. Note word IDs are continuous integers ( [0, vocabulary size-1])

• Test set consists of three files:
  • p_test_text.txt
  • p_test_visual.txt
  • p_test_emotion.txt

The file formats are similar to the training set.

Please see the features of POTD dataset for an example.

Output

• Training (model)

  File	Description	Dimension
  p_zt_t.txt	textual topic word distribution	(K+E) * T
  p_zv_t.txt	image-visual topic word distribution	K * C
  p_ze_t.txt	image-emotion topic word distribution	E * S
  p_t_r.txt	textual word relevance distribution	T * 2
  p_d_zv.txt	document visual-topic distribution	D * K
  p_d_ze.txt	document emotion-topic distribution	D * E

  Notation:

  • K/E is the number of visual/emotional topics.
  • T/C/S is the vocabulary size of textual/visual/emotional words.
  • D is the number of training documents.

• Testing

  File	Description	Dimension
  p_d_zv.txt	document visual-topic distribution	D * K
  p_d_ze.txt	document emotion-topic distribution	D * E

  Notation:

  • K/E is the number of visual/emotional topics.
  • D is the number of testing documents.

• Evaluation

  • result.csv: error rate against different top percentage of retrieved results
  • top_results.csv: each line contains the top five results (image ID and score), along with ground truth and a random image for the query text.

  Note the image IDs are continuous integers ([0, D-1]).