Downloads: 1
Diff of
/src/features/position_feature.py
[000000]
..
[735bb5]
Switch to unified view
| a | b/src/features/position_feature.py | ||
|---|---|---|---|
| 1 | # Base Dependencies |
||
| 2 | # ---------------- |
||
| 3 | import numpy |
||
| 4 | |||
| 5 | # Local Dependencies |
||
| 6 | # ------------------ |
||
| 7 | from models import RelationCollection |
||
| 8 | |||
| 9 | # 3rd-Party Dependencies |
||
| 10 | # ---------------------- |
||
| 11 | from sklearn.base import BaseEstimator |
||
| 12 | |||
| 13 | # Constants |
||
| 14 | # --------- |
||
| 15 | from constants import ( |
||
| 16 | N2C2_ATTR_ENTITY_CANDIDATES, |
||
| 17 | DDI_ATTR_ENTITY_CANDIDATES, |
||
| 18 | ) |
||
| 19 | |||
| 20 | |||
| 21 | class PositionFeature(BaseEstimator): |
||
| 22 | """ |
||
| 23 | Position Distance |
||
| 24 | |||
| 25 | Computes the position of the entity candidate (drug) with respect to |
||
| 26 | the attribute among the entire entity candidates of the attribute, where |
||
| 27 | the position of medical attribute is set to 0. |
||
| 28 | |||
| 29 | Source: |
||
| 30 | Alimova and Tutubalina (2020) - Multiple features for clinical relation extraction: A machine learning approach |
||
| 31 | """ |
||
| 32 | |||
| 33 | def __init__(self, dataset: str): |
||
| 34 | if dataset == "n2c2": |
||
| 35 | self.attr_entity_candidates = N2C2_ATTR_ENTITY_CANDIDATES |
||
| 36 | |||
| 37 | elif dataset == "ddi": |
||
| 38 | self.attr_entity_candidates = DDI_ATTR_ENTITY_CANDIDATES |
||
| 39 | else: |
||
| 40 | raise ValueError( |
||
| 41 | "only datasets 'n2c2' and 'ddi' are supported, but no '{}'".format( |
||
| 42 | dataset |
||
| 43 | ) |
||
| 44 | ) |
||
| 45 | self.dataset = dataset |
||
| 46 | |||
| 47 | def get_feature_names(self, input_features=None): |
||
| 48 | return ["position_1", "position_2"] |
||
| 49 | |||
| 50 | def create_position_feature(self, collection: RelationCollection) -> numpy.array: |
||
| 51 | features = [] |
||
| 52 | for r in collection.relations: |
||
| 53 | feature = [0] * 2 |
||
| 54 | |||
| 55 | attr, drug = r._ordered_entities |
||
| 56 | candidates = self.attr_entity_candidates[attr.type] |
||
| 57 | |||
| 58 | # count middle entities which could form the same type of relation |
||
| 59 | # i.e., count number of middle entities that are drugs for n2c2 and DDI |
||
| 60 | position = 0 |
||
| 61 | for ent in r.middle_entities: |
||
| 62 | if ent.type in candidates: |
||
| 63 | position += 1 |
||
| 64 | |||
| 65 | ent1 = r.entity1 |
||
| 66 | ent2 = r.entity2 |
||
| 67 | # if the attribute is the first entity, the position is positive |
||
| 68 | if ent1.type == attr.type: |
||
| 69 | feature[0] = 0 |
||
| 70 | feature[1] = position |
||
| 71 | # if the attribute is the second entity, the position is negative |
||
| 72 | elif ent2.type == attr.type: |
||
| 73 | feature[0] = -position |
||
| 74 | feature[1] = 0 |
||
| 75 | else: |
||
| 76 | raise ValueError("none of the entities correspond with the attribute") |
||
| 77 | |||
| 78 | features.append(feature) |
||
| 79 | |||
| 80 | return numpy.array(features) |
||
| 81 | |||
| 82 | def fit(self, x: RelationCollection, y=None): |
||
| 83 | return self |
||
| 84 | |||
| 85 | def transform(self, x: RelationCollection) -> numpy.array: |
||
| 86 | return self.create_position_feature(x) |
||
| 87 | |||
| 88 | def fit_transform(self, x: RelationCollection, y=None) -> numpy.array: |
||
| 89 | return self.create_position_feature(x) |