TAA4EUDCP

An official implementation for the paper "Tree-Averaging Algorithms for Ensemble-Based Unsupervised Discontinuous Constituency Parsing."

Usage

The ensemble function takes multiple parameters to control its behavior, providing flexibility in its application. Below is a guide on how to use this function effectively.

Function Signature

def ensemble(references, beta=1, weights=None, MAX_CANDID=float('inf'), MAX_LEN=float('inf'),
             lowerbound=True, guarantee=True, return_candid_count=False, MitM=True,
             parallel=True, return_times=False, progress_bar=True, DP=False, binary=False):

Parameters

• references: A list of references (trees) to be combined.
• beta: To indicate f_beta as the ensemble objective, default is 1.
• binary: A boolean indicating whether to make the results binary. Default is False.
• weights: A list of weights corresponding to each reference. If not provided, all weights are set to 1.
• MAX_CANDID: The maximum number of candidates to consider. Default is infinity.
• MAX_LEN: The maximum length of the sequence to consider. Default is infinity.
• lowerbound: A boolean indicating whether to apply a lower bound during the pruning step. Default is True.
• guarantee: A boolean indicating whether to exclude guaranteed spans from the search. Default is True.
• parallel: A boolean indicating whether to execute in parallel. Default is True.
• return_times: A boolean indicating whether to return the execution times for each sentence. Default is False.
• return_candid_count: A boolean indicating whether to return the count of candidates. Default is False.
• DP: A boolean indicating whether to use dynamic programming. Otherwise, follows the general search. Default is False.
• MitM: A boolean indicating whether to use the meet-in-the-middle approach. Default is True.
• progress_bar: A boolean indicating whether to display a progress bar. Default is True.

Returns

The function returns the average tree representations. If return_times is set to True, it also returns the execution times for each sentence.

Example Usage

from evaluate import f1
from individuals.utils import read_individuals
from library.ensemble import ensemble

# Reading individual references and their weights and the gold standard for evaluation
individuals, weights = zip(*[
    ('individuals/lassy/TN_LCFRS2024-01-27-00_54_23', 5.95),
    ('individuals/lassy/TN_LCFRS2024-01-27-00_58_43', 5.82),
    ('individuals/lassy/TN_LCFRS2024-01-27-00_55_50', 5.94),
    ('individuals/lassy/TN_LCFRS2024-01-27-00_55_01', 5.52),
    ('individuals/lassy/TN_LCFRS2024-01-27-00_56_08', 1.38),
])
individuals, gold = read_individuals('test', individuals)

# Generating the average tree
avgs = ensemble(individuals, weights=weights)

# Evaluating the results
print(f1(avgs, gold))