Spaces:
Sleeping
Sleeping
| from lingua.concept.linguagraph.standard import LinguaGraphNodePoses | |
| from lingua.structure.gpgraph import GPGraph, GPGAuxNode, GPGPhraseNode, GPGNode | |
| priority_map = { | |
| "AttributePredicator": 80, | |
| "AppositionalPredicator": 80, | |
| "FactualPredicator": 80, | |
| "LogicalPredicator": 80, | |
| "ReferentialPredicator": 80, | |
| "ConjunctionalFunctor": 80, | |
| "ExpressionFunctor": 40, | |
| "GeneralFunctor": 40, | |
| "ListFunctor": 40, | |
| "ModificationalFunctor": 80, | |
| "DeterminerConstant": 80, | |
| "InterjectionConstant": 80 , | |
| "ModificationalConstant": 80, | |
| "NominalConstant": 80, | |
| "OtherConstant": 80, | |
| "PunctuationalConstant": 80, | |
| "SymbolConstant": 80 | |
| } | |
| def min_node_distance(n1: GPGNode, n2: GPGNode): | |
| if isinstance(n1, GPGAuxNode) or isinstance(n2, GPGAuxNode): | |
| return 10000 | |
| assert isinstance(n1, GPGPhraseNode) and isinstance(n2, GPGPhraseNode) | |
| import numpy as np | |
| n1_words = np.array(list(n1.words(with_aux=False))) | |
| n2_words = np.array(list(n2.words(with_aux=False))) | |
| from scipy.spatial.distance import cdist | |
| return np.min(cdist(n1_words[..., np.newaxis], n2_words[..., np.newaxis])) # word index difference between n1 and n2 | |
| def get_modification_priority(graph: GPGraph, node, modification_phrases): | |
| base_priority = 20 | |
| node_words = list(node.words(with_aux=False)) | |
| min_node_pos = min(node_words) | |
| phrase_scores = [] | |
| for phrase in modification_phrases: | |
| min_modification_pos = 10000 | |
| min_dist = 10000 | |
| for c in phrase: | |
| min_dist = min(min_dist, min_node_distance(node, c)) | |
| if isinstance(c, GPGPhraseNode): | |
| c_words = list(c.words(with_aux=False)) | |
| min_modification_pos = min(min_modification_pos, min(c_words)) | |
| # 如果修饰词在中心词之前，给予奖励；否则不奖励 | |
| positional_bonus = 100 if min_modification_pos < min_node_pos else 0 | |
| # 如果距离超过1，每超过1个 token 扣除一定分数 | |
| distance_penalty_factor = 5 | |
| distance_penalty = (min_dist - 1) * distance_penalty_factor if min_dist > 1 else 0 | |
| effective_score = base_priority + positional_bonus - distance_penalty | |
| phrase_scores.append(effective_score) | |
| return phrase_scores | |