Adding a new align_to_words param to qa pipeline. (huggingface#18010)

* Adding a new `align_to_words` param to qa pipeline.

* Update src/transformers/pipelines/question_answering.py

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Import protection.

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

src/transformers/pipelines/question_answering.py

@@ -8,7 +8,14 @@
 from ..data import SquadExample, SquadFeatures, squad_convert_examples_to_features
 from ..modelcard import ModelCard
 from ..tokenization_utils import PreTrainedTokenizer
-from ..utils import PaddingStrategy, add_end_docstrings, is_tf_available, is_torch_available, logging
+from ..utils import (
+    PaddingStrategy,
+    add_end_docstrings,
+    is_tf_available,
+    is_tokenizers_available,
+    is_torch_available,
+    logging,
+)
 from .base import PIPELINE_INIT_ARGS, ArgumentHandler, ChunkPipeline
 
 
@@ -18,6 +25,9 @@
     from ..modeling_tf_utils import TFPreTrainedModel
     from ..modeling_utils import PreTrainedModel
 
+    if is_tokenizers_available():
+        import tokenizers
+
 if is_tf_available():
     import tensorflow as tf
 
@@ -180,6 +190,7 @@ def _sanitize_parameters(
         max_seq_len=None,
         max_question_len=None,
         handle_impossible_answer=None,
+        align_to_words=None,
         **kwargs
     ):
         # Set defaults values
@@ -208,6 +219,8 @@ def _sanitize_parameters(
             postprocess_params["max_answer_len"] = max_answer_len
         if handle_impossible_answer is not None:
             postprocess_params["handle_impossible_answer"] = handle_impossible_answer
+        if align_to_words is not None:
+            postprocess_params["align_to_words"] = align_to_words
         return preprocess_params, {}, postprocess_params
 
     def __call__(self, *args, **kwargs):
@@ -243,6 +256,9 @@ def __call__(self, *args, **kwargs):
                 The maximum length of the question after tokenization. It will be truncated if needed.
             handle_impossible_answer (`bool`, *optional*, defaults to `False`):
                 Whether or not we accept impossible as an answer.
+            align_to_words (`bool`, *optional*, defaults to `True`):
+                Attempts to align the answer to real words. Improves quality on space separated langages. Might hurt on
+                non-space-separated languages (like Japanese or Chinese)
 
         Return:
             A `dict` or a list of `dict`: Each result comes as a dictionary with the following keys:
@@ -386,6 +402,7 @@ def postprocess(
         top_k=1,
         handle_impossible_answer=False,
         max_answer_len=15,
+        align_to_words=True,
     ):
         min_null_score = 1000000  # large and positive
         answers = []
@@ -464,15 +481,8 @@ def postprocess(
                 for s, e, score in zip(starts, ends, scores):
                     s = s - offset
                     e = e - offset
-                    try:
-                        start_word = enc.token_to_word(s)
-                        end_word = enc.token_to_word(e)
-                        start_index = enc.word_to_chars(start_word, sequence_index=sequence_index)[0]
-                        end_index = enc.word_to_chars(end_word, sequence_index=sequence_index)[1]
-                    except Exception:
-                        # Some tokenizers don't really handle words. Keep to offsets then.
-                        start_index = enc.offsets[s][0]
-                        end_index = enc.offsets[e][1]
+
+                    start_index, end_index = self.get_indices(enc, s, e, sequence_index, align_to_words)
 
                     answers.append(
                         {
@@ -490,6 +500,24 @@ def postprocess(
             return answers[0]
         return answers
 
+    def get_indices(
+        self, enc: "tokenizers.Encoding", s: int, e: int, sequence_index: int, align_to_words: bool
+    ) -> Tuple[int, int]:
+        if align_to_words:
+            try:
+                start_word = enc.token_to_word(s)
+                end_word = enc.token_to_word(e)
+                start_index = enc.word_to_chars(start_word, sequence_index=sequence_index)[0]
+                end_index = enc.word_to_chars(end_word, sequence_index=sequence_index)[1]
+            except Exception:
+                # Some tokenizers don't really handle words. Keep to offsets then.
+                start_index = enc.offsets[s][0]
+                end_index = enc.offsets[e][1]
+        else:
+            start_index = enc.offsets[s][0]
+            end_index = enc.offsets[e][1]
+        return start_index, end_index
+
     def decode(
         self, start: np.ndarray, end: np.ndarray, topk: int, max_answer_len: int, undesired_tokens: np.ndarray
     ) -> Tuple:

tests/pipelines/test_pipelines_question_answering.py

@@ -171,6 +171,29 @@ def ensure_large_logits_postprocess(
 
         self.assertEqual(nested_simplify(outputs), {"score": 0.028, "start": 0, "end": 11, "answer": "HuggingFace"})
 
+    @slow
+    @require_torch
+    def test_small_model_japanese(self):
+        question_answerer = pipeline(
+            "question-answering",
+            model="KoichiYasuoka/deberta-base-japanese-aozora-ud-head",
+        )
+        output = question_answerer(question="国語", context="全学年にわたって小学校の国語の教科書に挿し絵が用いられている")
+
+        # Wrong answer, the whole text is identified as one "word" since the tokenizer does not include
+        # a pretokenizer
+        self.assertEqual(
+            nested_simplify(output),
+            {"score": 1.0, "start": 0, "end": 30, "answer": "全学年にわたって小学校の国語の教科書に挿し絵が用いられている"},
+        )
+
+        # Disable word alignment
+        output = question_answerer(question="国語", context="全学年にわたって小学校の国語の教科書に挿し絵が用いられている", align_to_words=False)
+        self.assertEqual(
+            nested_simplify(output),
+            {"score": 1.0, "start": 15, "end": 18, "answer": "教科書"},
+        )
+
     @slow
     @require_torch
     def test_small_model_long_context_cls_slow(self):