Keyphrase Dataset Annotation via TagMe

Setting Up

Set up a json file in keyphrase-annotation/configuration.json:

{
	"tagme-token": "TagMe authorization token",
	"keyphrase-data": "Path to keyphrase_data.zip"
}

where the tagme-token can be retrieved by following Installation and setup section available here.

Installing Dependencies

If you won't install dependencies system-wide use some virtualization tool, like Virtualenv. Then just install dependencies with:

pip install -r keyphrase-annotation/requirements.txt

Running

You can annotate a single dataset with:

python keyphrase-annotation annotate dataset_name output_dir

where dataset_name can be:

• duc for the DUC-2001 dataset.
• icsi-asr or icsi-ht for ICSI ASR_Output or Human_Transcript datasets, respectively.
• inspec-train, inspec-val or inspec-test for Inspec training, validation or test datasets, respectively.
• nus for the NUSkeyphraseCorpus dataset.

You can also annotate all datasets at once by typing:

python keyphrase-annotation annotate_all output_dir

and the annotations for each dataset will be saved in the corresponding folder inside output_dir (for example, the annotations of the DUC-2001 dataset will be saved into output_dir/duc).

Output

For each document a json file with TagMe annotations will be generated in the specified folder. Each json file has the following structure:

{
	"tagme":
		[
			{
                "wiki_title": str,
                "wiki_id": str,
                "annotations":
                    [
                        {
                            "begin":    int,
                            "end":      int,
                            "score":    float,
                            "spot":     str
                        }
                    ]
            }
        ]
}

The tagme field contains a list of Wikipedia Entities (each entity is uniquely identified by its wiki_title/wiki_id) annotated in the input document. For each annotated entity, annotations provides information where the corresponding entity has been annotated in the document. score is the TagMe coherence score between that annotation and the others in the surrounding text.