Add XLMRobertaForQuestionAnswering

docs/source/model_doc/xlmroberta.rst

@@ -84,6 +84,13 @@ XLMRobertaForTokenClassification
     :members:
 
 
+XLMRobertaForQuestionAnswering
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.XLMRobertaForQuestionAnswering
+    :members:
+
+
 TFXLMRobertaModel
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 

src/transformers/__init__.py

@@ -298,6 +298,7 @@
         XLMRobertaForMultipleChoice,
         XLMRobertaForSequenceClassification,
         XLMRobertaForTokenClassification,
+        XLMRobertaForQuestionAnswering,
         XLM_ROBERTA_PRETRAINED_MODEL_ARCHIVE_LIST,
     )
     from .modeling_mmbt import ModalEmbeddings, MMBTModel, MMBTForClassification

src/transformers/modeling_auto.py

@@ -121,6 +121,7 @@
 from .modeling_xlm_roberta import (
     XLMRobertaForMaskedLM,
     XLMRobertaForMultipleChoice,
+    XLMRobertaForQuestionAnswering,
     XLMRobertaForSequenceClassification,
     XLMRobertaForTokenClassification,
     XLMRobertaModel,
@@ -230,6 +231,7 @@
         (DistilBertConfig, DistilBertForQuestionAnswering),
         (AlbertConfig, AlbertForQuestionAnswering),
         (LongformerConfig, LongformerForQuestionAnswering),
+        (XLMRobertaConfig, XLMRobertaForQuestionAnswering),
         (RobertaConfig, RobertaForQuestionAnswering),
         (BertConfig, BertForQuestionAnswering),
         (XLNetConfig, XLNetForQuestionAnsweringSimple),

src/transformers/modeling_xlm_roberta.py

@@ -23,6 +23,7 @@
 from .modeling_roberta import (
     RobertaForMaskedLM,
     RobertaForMultipleChoice,
+    RobertaForQuestionAnswering,
     RobertaForSequenceClassification,
     RobertaForTokenClassification,
     RobertaModel,
@@ -120,3 +121,17 @@ class XLMRobertaForTokenClassification(RobertaForTokenClassification):
     """
 
     config_class = XLMRobertaConfig
+
+
+@add_start_docstrings(
+    """XLM-RoBERTa Model with a span classification head on top for extractive question-answering tasks like SQuAD (a
+    linear layers on top of the hidden-states output to compute `span start logits` and `span end logits`).""",
+    XLM_ROBERTA_START_DOCSTRING,
+)
+class XLMRobertaForQuestionAnswering(RobertaForQuestionAnswering):
+    """
+    This class overrides :class:`~transformers.RobertaForQuestionAnswering`. Please check the
+    superclass for the appropriate documentation alongside usage examples.
+    """
+
+    config_class = XLMRobertaConfig