tilse implements various models for predicting timelines. In particular, it implements a simple supervised regression baseline, the model by Chieu and Lee (2004), and various models employing optimization of submodular functions.
tilse needs data in a specific format. The command get-and-preprocess-data downloads, processes and converts
existing timeline summarization corpora. Possible options for get-and-preprocess-data are timeline17 and
crisis. The data is stored in a folder with the respective name in the current working directory (processing the
data may take a few hours).
Timelines can be predicted using the command predict-timelines, which takes a JSON configuration file as input (see
below). Evaluation results are printed to standard output, and output (containing predicted timelines, reference
timelines and evaluation results) are stored in the current working directory. Currently, timelines can only predicted
for corpora that have been downloaded and processed via get-and-preprocess-data.
predict-timelines expects a JSON configuration file. The file contains parameters needed for all models, and
model-specific parameters. The config folder in this repository contains configuration files that were used
to obtain the results in the paper.
The following example is a configuration file that describes a model using optimization of submodular functions.
{
"name": "tlsconstraints_timeline17",
"algorithm": "submodular",
"corpus": "/home/mitarb/martschat/timeline17",
"restrict_topics_to": null,
"date_selection": null,
"assess_length": "average_length_in_sentences",
"sentence_representation": "ChieuSentenceRepresentation",
"keyword_mapping": {
"bpoil": ["bp", "oil", "spill"],
"egypt": ["egypt", "egyptian"],
"finan": ["financial", "economic", "crisis"],
"h1n1": ["h1n1", "swine", "flu"],
"haiti": ["haiti", "quake", "earthquake"],
"iraq": ["iraq", "iraqi"],
"libya": ["libya", "libyan"],
"mj": ["michael", "jackson"],
"syria": ["syria", "syrian"],
"yemen": ["yemen"]
},
"rouge_computation": "original",
"properties": {
"constraint": "is_valid_individual_constraints",
"semantic_cluster": "clusters_by_similarity",
"date_cluster": "clusters_by_date",
"coefficients": [1, 1, 0, 0]
}
}
| Name | Type | Valid values | Description |
|---|---|---|---|
| name | str | any | Name of the experiments. Output of the experiment will be stored in a file called name.obj |
| algorithm | str | 'random', 'regression', 'chieu', 'submodular' | Which algorithm to use |
| corpus | str | any path | Location of the corpus to process |
| restrict_corpus_to | list(str) | entries: topics of the corpus | If not null, limit processing to specified topics |
| date_selection | str | functions in tilse.models.date_selection |
Which date selection algorithm to use before predicting timelines. If null, do date selection jointly with summary generation |
| assess_length | str | functions in tilse.models.assess_length |
Which daily summary length assession algorithm to use. Cannot be null |
| sentence_representation | str | classes in tilse.representations_sentence_representations |
Which daily summary length assession algorithm to use. Cannot be null |
| keyword_mapping | map(str, list(str)) | Maps from topic names to list of keywords | If not null, only consider sentences that contain at least one of the specified keywords |
| rouge_computation | str | original, reimplementation |
Which ROUGE implementation to use original uses the original ROUGE toolkit, reimplementation uses a Python reimplementation. The reimplementation is faster, but leads to slightly different results |
If the algorithm is set to 'chieu' or 'submodular', further model-specific parameters need to be given via the 'properties' key.
| Name | Type | Valid values | Description |
|---|---|---|---|
| interest_measure | str | functions in tilse.models.chieu.interest_measures |
Interest measure to use |
| Name | Type | Valid values | Description |
|---|---|---|---|
| constraint | str | functions in tilse.models.submodular.constraints |
Constraint to respect during greedy algorithm |
| semantic_cluster | str | functions in tilse.models.submodular.semantic_cluster_functions |
Cluster computation function on which semantic diversity is based |
| date_cluster | str | functions in tilse.models.submodular.date_cluster_functions |
Cluster computation function on which temporal diversity is based |
| coefficients | list(double) | lists of length 4 | coefficients for coverage, semantic diversity, temporal diversity, date references |