Question Answering Specifics

The goal of Question Answering is to find the answer to a question given a question and an accompanying context. The predicted answer will be either a span of text from the context or an empty string (indicating the question cannot be answered from the context).

Usage Steps

The process of performing Question Answering in Simple Transformers does not deviate from the standard pattern.

  1. Initialize a QuestionAnsweringModel
  2. Train the model with train_model()
  3. Evaluate the model with eval_model()
  4. Make predictions on (unlabelled) data with predict()

Supported Model Types

New model types are regularly added to the library. Question Answering tasks currently supports the model types given below.

Model Model code for QuestionAnsweringModel
ALBERT albert
BERT bert
DistilBERT distilbert
ELECTRA electra
Longformer longformer
MobileBERT mobilebert
RoBERTa roberta
XLM xlm
XLM-RoBERTa xlmroberta
XLNet xlnet

Tip: The model code is used to specify the model_type in a Simple Transformers model.

Lazy Loading Data

The system memory required to keep a large dataset in memory can be prohibitively large. In such cases, the data can be lazy loaded from disk to minimize memory consumption.

To enable lazy loading, you must set the lazy_loading flag to True in QuestionAnsweringArgs.

model_args = QuestionAnsweringArgs()
model_args.lazy_loading = True

Note: This will typically be slower as the feature conversion is done on the fly. However, the tradeoff between speed and memory consumption should be reasonable.

Tip: See Lazy Loading Data Formats for information on the data formats.

Tip: See Configuring a QuestionAnsweringArgs model for information on configuring the model to read the lazy loading data file correctly.

Tip: You can find a minimal example script in examples/question_answering/