Improving Event Coreference Resolution by Modeling Correlations between Event Coreference Chains and Document Topic Structures

STEPS

1. Run https://github.com/prafulla77/TAC-KBP-2017-Participation to extract event mentions and obtain pairwise score for coreference relations.
2. Run https://github.com/PrincetonML/SIF to generate pairwise sentence similarity scores.
3. Extract subevents using the patterns described in paper (a sequence of two or more verb event mentions that modify their parent verb with 'conj:and' or 'ccomp' dependency relations in a sentence are extracted as subevents).
4. Run new_test_ILP.py.

Files format

• Coreference Scores:

  OrderedDict([FileName:OrderedDict([Event_Key:OrderedDict([prev_event_key_1: score_1, prev_key_2:score_2)])...])

• Get Subevents using get_subevents.py

• For sentence similarity: Use:

def _get_sim(emb1, emb2):
    inn = (emb1 * emb2).sum(axis=1)
    emb1norm = np.sqrt((emb1 * emb1).sum(axis=1))
    emb2norm = np.sqrt((emb2 * emb2).sum(axis=1))
    scores = inn / emb1norm / emb2norm
    return scores

where emb_1 and emb_2 are generated based on PrincetonML/SIF Project

Format: Dict{Filename:{previoussentence__currentsentence: score}}