File size: 8,263 Bytes
88d8172
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
import streamlit as st
import pandas as pd
import numpy as np

from nltk.tokenize import sent_tokenize

# Split the text into sentences. Necessary for NLI models
def split_sentences(text):
    return sent_tokenize(text)

###### Prompting
def query_model_prompting(model, text, prompt_with_mask, top_k, targets):
    """Query the prompting model

    :param model: Prompting model object
    :type model: Huggingface pipeline object
    :param text: Event description (context)
    :type text: str
    :param prompt_with_mask: Prompt with a mask
    :type prompt_with_mask: str
    :param top_k: Number of tokens to output
    :type top_k: integer
    :param targets: Restrict the answer to these possible tokens
    :type targets: list
    :return: Results of the prompting model
    :rtype: list of dict
    """
    sequence = text + prompt_with_mask
    output_tokens = model(sequence, top_k=top_k, targets=targets)
    
    return output_tokens

def do_sentence_entailment(sentence, hypothesis, model):
    """Concatenate context and hypothesis then perform entailment

    :param sentence: Event description (context), 1 sentence
    :type sentence: str
    :param hypothesis: Mask filled with a token
    :type hypothesis: str
    :param model: NLI Model
    :type model: Huggingface pipeline
    :return: DataFrame containing the result of the entailment
    :rtype: pandas DataFrame
    """
    text = sentence + '</s></s>' + hypothesis
    res = model(text, return_all_scores=True)
    df_res = pd.DataFrame(res[0])
    df_res['label'] = df_res['label'].apply(lambda x: x.lower())
    df_res.columns = ["Label", "Score"]
    return df_res

def softmax(x):
    """Compute softmax values for each sets of scores in x."""
    return np.exp(x) / np.sum(np.exp(x), axis=0)



######### NLI + PROMPTING
def do_text_entailment(text, hypothesis, model):
    """
    Do entailment for each sentence of the event description as
    model was trained on sentence pair

    :param text: Event Description (context)
    :type text: str
    :param hypothesis: Mask filled with a token
    :type hypothesis: str
    :param model: Model NLI
    :type model: Huggingface pipeline
    :return: List of entailment results for each sentence of the text
    :rtype: list
    """
    text_entailment_results = []
    for i, sentence in enumerate(split_sentences(text)):
        df_score = do_sentence_entailment(sentence, hypothesis, model)
        text_entailment_results.append((sentence, hypothesis, df_score))
    return text_entailment_results

def get_true_entailment(text_entailment_results, nli_limit):
    """
    From the result of each sentence entailment, extract the maximum entailment score and 
    check if it's higher than the entailment threshold. 
    """
    true_hypothesis_list = []
    max_score = 0
    for sentence_entailment in text_entailment_results:
        df_score = sentence_entailment[2]
        score = df_score[df_score["Label"] == 'entailment']["Score"].values.max()
        if score > max_score:
            max_score = score
    if max_score > nli_limit:
        true_hypothesis_list.append((sentence_entailment[1], np.round(max_score,2)))
    return list(set(true_hypothesis_list))

def prompt_to_nli(text, prompt, model_prompting, nli_model, nlp, top_k=10, nli_limit=0.5, remove_lemma=False):
    """
    Apply the PR-ENT pipeline
    
    :param text: Event description
    :type text: str
    :param prompt: Prompt with mask
    :type prompt: str
    :param model_prompting: Prompting Model
    :type model_prompting: Huggingface pipeline
    :param nli_model: NLI Model
    :type nli_model: Huggingface pipeline
    :param top_k: Number of words output by the prompting model
    :type top_k: int
    :param nli_limit: Entailment threshold
    :type nli_limit: float

    :return: Results of the pipeline
    :rtype: list
    """
    prompt_masked = prompt.format(model_prompting.tokenizer.mask_token)
    label = []
    output_prompting = query_model_prompting(model_prompting, text, prompt_masked, top_k, targets=None)
    if remove_lemma:
        output_prompting = filter_prompt_output_by_lemma(prompt, output_prompting, nlp)
    for token in output_prompting:
        hypothesis = prompt.format(token['token_str'])
        text_entailment_results = do_text_entailment(text, hypothesis, nli_model)
        true_hypothesis_list = get_true_entailment(text_entailment_results, nli_limit)
        if len(true_hypothesis_list) > 0:
            label.append(((token['token_str'], token['score']), true_hypothesis_list[0]))
    return label


def display_nli_pr_results_as_list(title, list_results):
    """
        Display the list of entailment results as a streamlit choice list
    """
    st.markdown(
        """
    <style>
    span[data-baseweb="tag"] {
    background-color: red !important;
    }
    </style>
    """,
        unsafe_allow_html=True,
    )
    prompt_list = st.multiselect(
        title,
        list_results 
        ,
        list_results, key='results_mix')


##### QA
def question_answering(model, text, questions_list, to_print=True):
    """
    Apply question answering model

    :param model: QA Model
    :type model: Huggingface pipeline
    :param text: Event description (context)
    :type text: str
    :param question: Question to answer
    :type question: str
    :return: Tuple containing the answer and the confidence score
    :rtype: tuple
    """
    for question in questions_list:
        QA_input = {
        'question': question,
        'context': text}
        res = model(QA_input, handle_impossible_answer=False)

        if to_print:
            st.write("Question: {}".format(question))
            st.write("Answer: {}".format(res["answer"]))

    return res["answer"], res["score"]


### Prompt + NLI + QA

def get_who_what_whom_qa(text, tokens, model_qa):
    who_what_whom = []
    if not tokens:
        res_dict = {"Actor":'', "Action":'', "Target": ''}
        st.write("No entailed tokens.")

    else:
        for token in tokens:
            # res_dict = {"who":'', "did_what":token, "to_whom": '', "qa_score": []}
            res_dict = {"Actor":'', "Action":token, "Target": ''}

            if token[-3:] == 'ing':
                perp,score_p = question_answering(model_qa, text, ["Who was {}?".format(token)], to_print=False)
            else:
                perp,score_p = question_answering(model_qa, text, ["Who {} people?".format(token)], to_print=False)
            if perp:
                res_dict["Actor"] = perp + ' [' + str(np.round(score_p*100,1)) + '%]'
            else:
                res_dict["Actor"] = 'N/A' + ' [' + str(np.round(score_p*100,1)) + '%]'

            victim,score_v = question_answering(model_qa, text, ["Who was {}?".format(token)], to_print=False)         

            if victim:
                res_dict["Target"] = victim + ' [' + str(np.round(score_v*100,1)) + '%]'
            else:
                res_dict["Target"] = 'N/A' + ' [' + str(np.round(score_v*100,1)) + '%]'
            
            who_what_whom.append(res_dict)

    return who_what_whom

def remove_similar_lemma_from_list(prompt, list_words, nlp):
    ## Compute a dictionnary with the lemma for all tokens
    ## If there is a duplicate lemma then the dictionnary value will be a list of the corresponding tokens
    lemma_dict = {}
    for each in list_words:
        mask_filled = nlp(prompt.strip('.').format(each))
        lemma_dict.setdefault([x.lemma_ for x in mask_filled][-1],[]).append(each)

    ## Get back the list of tokens
    ## If multiple tokens available then take the shortest one
    new_token_list = []
    for key in lemma_dict.keys():
        if len(lemma_dict[key]) >= 1:
            new_token_list.append(min(lemma_dict[key], key=len))
        else:
            raise ValueError("Lemma dict has 0 corresponding words")
    return new_token_list

def filter_prompt_output_by_lemma(prompt, output_prompting, nlp):
    """
        Remove all similar lemmas from the prompt output (e.g. "protest", "protests")
    """
    list_words = [x['token_str'] for x in output_prompting]
    new_token_list = remove_similar_lemma_from_list(prompt, list_words, nlp)
    return [x for x in output_prompting if x['token_str'] in new_token_list]