The goal is to let intelligent agents interpret and learn high-level user intents which span multiple mobile apps, e.g., to plan a dinner we may need to use Yelp -> Maps -> SMS, etc.

There are several ways to train app embeddings. You can use doc2vec on app descriptions to project each app into a semantic space. Alternatively, you can collect stream of app invocations from people's smart phones and treat it as a corpus of words and apply word2vec.

In sequence_labeling directory you will find following:

1. train, test, dev splits for app sequences train.apps.int, test.apps.int, dev.apps.int. The numeric ids correspond to labels provided apps.csv file.
2. B/I/O tagging information for the app sequences train.labels.int, test.labels.int, dev.labels.int. The numeric ids correspond to labels provided in labels.csv file.;
3. CRFSuite sequence labeling models for these sequences.

1. App invocation sequences collected from 19 users' Android phones (R1.csv);
2. Clean app sequences (apps irrelevant to the intents removed) with user intents annotated by participants (R2.csv);
3. Speech commands (both manual transcripts and Google ASR 1-best hypotheses) at app level to re-enact part of intents in 2 (R3.csv).

Please cite following work if you use this dataset in your research work.

@CONFERENCE {sunSLT2016,
    author    = "Ming Sun, Aasish Pappu, Yun-Nung Chen, Alexander I Rudnicky",
    title     = "Weakly Supervised User Intent Detection for Multi-Domain Dialogues",
    booktitle = "IEEE Workshop on Spoken Language Technology",
    year      = "2016",
    publisher = "IEEE"
}

You can find a video demo here: https://youtu.be/FvQto8pP1OU

Creative Commons License 1.0

For any questions/suggestions contact: mings@cs.cmu.edu, aasishp@gmail.com