Question Answering Model for SQuAD dataset¶
Task definition¶
Question Answering on SQuAD dataset is a task to find an answer on question in a given context (e.g, paragraph from Wikipedia), where the answer to each question is a segment of the context:
Context:
In meteorology, precipitation is any product of the condensation of atmospheric water vapor that falls under gravity. The main forms of precipitation include drizzle, rain, sleet, snow, graupel and hail… Precipitation forms as smaller droplets coalesce via collision with other rain drops or ice crystals within a cloud. Short, intense periods of rain in scattered locations are called “showers”.
Question:
Where do water droplets collide with ice crystals to form precipitation?
Answer:
within a cloud
Datasets, which follow this task format:
Stanford Question Answering Dataset (SQuAD) (EN)
SDSJ Task B (RU)
Models¶
SQuAD model in DeepPavlov is based on BERT. The model predicts answer start and end position in a given context. Their performance is compared in pretrained models section of this documentation.
BERT¶
Pretrained BERT can be used for Question Answering on SQuAD dataset just by applying two linear transformations to BERT outputs for each subtoken. First/second linear transformation is used for prediction of probability that current subtoken is start/end position of an answer.
BERT for SQuAD model documentation on PyTorch torch_transformers_squad:TorchTransformersSquad
.
Configuration¶
Default configs could be found in deeppavlov/configs/squad/ folder.
Prerequisites¶
Before using the model make sure that all required packages are installed running the command:
python -m deeppavlov install squad_bert
By running this command we will install requirements for deeppavlov/configs/squad/squad_bert.json.
Model usage from Python¶
from deeppavlov import build_model
model = build_model('squad_bert', download=True)
model(['DeepPavlov is library for NLP and dialog systems.'], ['What is DeepPavlov?'])
Model usage from CLI¶
Training¶
Warning: training with default config requires about 10Gb on GPU. Run following command to train the model:
python -m deeppavlov train squad_bert
Interact mode¶
Interact mode provides command line interface to already trained model.
To run model in interact mode run the following command:
python -m deeppavlov interact squad_bert
Model will ask you to type in context and question.
Pretrained models:¶
SQuAD¶
We have all pretrained model available to download:
python -m deeppavlov download squad_bert
It achieves ~88 F-1 score and ~80 EM on SQuAD-v1.1 dev set.
In the following table you can find comparison with published results. Results of the most recent competitive solutions could be found on SQuAD Leadearboad.
Model (single model) |
EM (dev) |
F-1 (dev) |
---|---|---|
81.49 |
88.86 |
|
– |
85.6 |
|
75.1 |
83.8 |
|
75.3 |
83.6 |
|
71.1 |
79.5 |
|
67.7 |
77.3 |
SQuAD with contexts without correct answers¶
In the case when answer is not necessary present in given context we have squad_noans with pretrained model. This model outputs empty string in case if there is no answer in context. squad_noans was trained on SQuAD2.0 dataset.
Special trainable no_answer token is added to output of self-attention layer and it makes model able to select no_answer token in cases, when answer is not present in given context.
We got 57.88 EM and 65.91 F-1 on ground truth Wikipedia article (we used the same Wiki dump as DrQA):
Model config |
EM (dev) |
F-1 (dev) |
|
---|---|---|---|
75.54 |
83.56 |
||
59.14 |
67.34 |
||
49.7 |
– |
Pretrained model is available and can be downloaded (~2.5Gb):
python -m deeppavlov download qa_squad2_bert
SDSJ Task B¶
Pretrained model is available and can be downloaded:
python -m deeppavlov download squad_ru_bert
Link to SDSJ Task B dataset: http://files.deeppavlov.ai/datasets/sber_squad-v1.1.tar.gz
Model config |
EM (dev) |
F-1 (dev) |
---|---|---|
66.21 |
84.71 |