deepset

/

roberta-base-squad2

like 697

Here is a rudimentary test for finding answers at the end of the context input. It appears to start having trouble beyond around 100,000 characters.

from transformers import pipeline, AutoModelForQuestionAnswering, AutoTokenizer
from pathlib import Path

tokenizer = AutoTokenizer('deepset/roberta-base-squad2')
model = AutoModelForQuestionAnswering.from_pretrained('deepset/roberta-base-squad2')
nlp = pipeline('question-answering', tokenizer=tokenizer, model=model)

# the text is the concatenation of the cleaned wikipedia articles for Germany and Greece
# a litle less than 167,000 characters.
text = Path('text/germany-greece.txt').read_text()
question = 'What is my name?'
base_ix = 80000
ix_step = 1000

for step in range(85):
    ctx_len = t[:base_ix + step * ix_step].rfind('.') + 1
    qa_input = {'context': t[:ctx_len] + ' My name is James.', 'question': q}
    answer = nlp(qa_input)
    print(f'{ctx_len:>6}:', answer['answer'], round(answer['score'], 3))

Output:

 59995: James 0.378
 60923: James 0.24
 61954: James 0.037
 62912: James 0.654
 63888: James 0.753
 64923: James 0.48
 65924: James 0.832
 66671: James 0.89
 67914: James 0.92
 68915: James 0.869
 69932: James 0.756
 70956: James 0.52
 71981: James 0.89
 72944: James 0.887
 73993: James 0.544
 74968: James 0.864
 75882: James 0.902
 76958: James 0.389
 77962: James 0.916
 78964: James 0.939
 79963: James 0.854
 80969: James 0.796
 81861: James 0.893
 82886: James 0.51
 83990: James 0.817
 84894: James 0.807
 85833: James 0.714
 86918: James 0.363
 87988: James 0.865
 88963: James 0.525
 89803: James 0.609
 90990: James 0.686
 91963: James 0.662
 92873: James 0.658
 93963: James 0.916
 94998: James 0.575
 95846: James 0.768
 96986: James 0.813
 97753: James 0.895
 98780: James 0.847
 99812: James 0.71
100818: James 0.887
101808: James 0.889
102918: James 0.648
103956: James 0.729
104633: James 0.81
105863: James 0.887
106906: James 0.789
107749: Ellinikos Stratos, ES 0.471
108949: James 0.88
109977: James 0.89
110954: James 0.802
111948: Ellinikos Stratos, ES 0.471
112951: Ellinikos Stratos, ES 0.471
113877: James 0.612
114935: James 0.617
115979: James 0.849
116755: James 0.906
117825: James 0.907
118983: James 0.927
119839: James 0.913
120955: James 0.728
121934: Ellinikos Stratos, ES 0.471
122876: James 0.535
123960: Ellinikos Stratos, ES 0.471
124970: James 0.918
125002: James 0.921
126989: James 0.899
127993: Ellinikos Stratos, ES 0.471
128939: James 0.676
129958: James 0.923
130987: James 0.782
131757: James 0.752
132956: James 0.836
133853: Ellinikos Stratos, ES 0.471
134975: Ellinikos Stratos, ES 0.471
135899: Ellinikos Stratos, ES 0.471
136932: Ellinikos Stratos, ES 0.471
137930: James 0.684
138860: Ellinikos Stratos, ES 0.471
139932: James 0.475
140903: James 0.896
141886: James 0.765
142992: James 0.54
143991: James 0.502