Commit 4a640e2a authored by Nico Schnurr's avatar Nico Schnurr
Browse files

Change probing file

parent d05ee937
Loading
Loading
Loading
Loading
+52 −56
Changes for probing.py: 52 added lines, 56 removed lines.
Original line number Diff line number Diff line
from STSB_Dataset import STSB_Dataset
import pandas as pd
import spacy
import random


df = STSB_Dataset('data/stsbenchmark/sts-test.csv').as_dataframe
@@ -14,60 +13,75 @@ sentences2 = df['sentence2'].tolist()
sentences1_nlp = [nlp(sent) for sent in sentences1]
sentences2_nlp = [nlp(sent) for sent in sentences2]

sent1_roots = []
sent2_roots = []

sent1_root_children = []
sent2_root_children = []
dependency_hierarchy1 = []
dependency_hierarchy2 = []

for sent1 in sentences1_nlp:

    root1 = [(token.text, token, token.pos_) for token in sent1 if token.dep_ == "ROOT"]
    sent1_roots.append((root1[0][0], root1[0][2]))
    root_children1 = [(child.text, child.pos_) for child in root1[0][1].children]
    sent1_root_children.append(root_children1)
    sentence_tokens = [token for token in sent1]
    dep_hierarchy1 = []
    dep_hierarchy1.append([token for token in sent1 if token.dep_ == "ROOT"])
    sentence_tokens.remove(dep_hierarchy1[0][0])

    while len(sentence_tokens) != 0:
        children = []
        for i in range(len(dep_hierarchy1[-1])):
            index_children = [t for t in sent1 if t.head.text == dep_hierarchy1[-1][i].text and t in sentence_tokens]
            for elem in index_children:
                children.append(elem)
                sentence_tokens.remove(elem)
        dep_hierarchy1.append(children)

    dependency_hierarchy1.append(dep_hierarchy1)


for sent2 in sentences2_nlp:

    root2 = [(token.text, token, token.pos_) for token in sent2 if token.dep_ == "ROOT"]
    sent2_roots.append((root2[0][0], root2[0][2]))
    root_children2 = [(child.text, child.pos_) for child in root2[0][1].children]
    sent2_root_children.append(root_children2)
    sentence_tokens = [token for token in sent2]
    dep_hierarchy2 = []
    dep_hierarchy2.append([token for token in sent2 if token.dep_ == "ROOT"])
    sentence_tokens.remove(dep_hierarchy2[0][0])

    while len(sentence_tokens) != 0:
        children = []
        for i in range(len(dep_hierarchy2[-1])):
            index_children = [t for t in sent2 if t.head.text == dep_hierarchy2[-1][i].text and t in sentence_tokens]
            for elem in index_children:
                children.append(elem)
                sentence_tokens.remove(elem)
        dep_hierarchy2.append(children)

    dependency_hierarchy2.append(dep_hierarchy2)

assert len(sent1_roots) == len(sent2_roots)
assert len(sent1_root_children) == len(sent2_root_children)


def mask_token(pos_tag):
    del_rows = []     #rows to be deleted
    
    for i in range(len(sent1_roots)):
        if sent1_roots[i][1] == pos_tag:
            df.iloc[i, 0] = df.iloc[i, 0].replace(sent1_roots[i][0], "[MASK]")
        else:
            for e in range(len(sent1_root_children[i])):
                if pos_tag in sent1_root_children[i][e]:
                    df.iloc[i, 0] = df.iloc[i, 0].replace(sent1_root_children[i][e][0], "[MASK]")
    for e in range(len(dependency_hierarchy1)):
        mask_token = None
        for layer in dependency_hierarchy1[e]:
            if mask_token != None:
                break
                elif e == len(sent1_root_children[i]) - 1:
                    del_rows.append(i)

    for i in range(len(sent2_roots)):
        if sent1_roots[i][1] == pos_tag:
            df.iloc[i, 1] = df.iloc[i, 1].replace(sent2_roots[i][0], "[MASK]")
        else:
            for e in range(len(sent2_root_children[i])):
                if pos_tag in sent2_root_children[i][e]:
                    df.iloc[i, 1] = df.iloc[i, 1].replace(sent2_root_children[i][e][0], "[MASK]")
            for i in range(len(layer)):
                if layer[i].pos_ == pos_tag:
                    mask_token = layer[i].text
                    df.iloc[e, 0] = df.iloc[e, 0].replace(mask_token, "[MASK]")
                    break
                elif e == len(sent2_root_children[i]) - 1:
                    if i not in del_rows:
                        del_rows.append(i)

    df.drop(del_rows, inplace=True)
    for e in range(len(dependency_hierarchy2)):
        mask_token = None
        for layer in dependency_hierarchy2[e]:
            if mask_token != None:
                break
            for i in range(len(layer)):
                if layer[i].pos_ == pos_tag:
                    mask_token = layer[i].text
                    df.iloc[e, 1] = df.iloc[e, 1].replace(mask_token, "[MASK]")
                    break
    return df
            


def mask_first(pos_tag):

    i = -1
@@ -102,21 +116,3 @@ def mask_first(pos_tag):
    return df

    
    

            





#print(sent1_roots)

#print(sent2_roots)

#print(sent1_root_children)

#print(sent2_root_children)

# mask_token("VERB")
print(mask_first("NUM"))