Add "no answer" aggregation to Transformersreader (#259)

* Add no answer aggregation

* Change to covariant type annotation

* Remove n_best_per_passage from transformersreader

haystack/reader/base.py

@@ -1,5 +1,7 @@
+import numpy as np
+from scipy.special import expit
 from abc import ABC, abstractmethod
-from typing import List, Optional
+from typing import List, Optional, Sequence
 
 from haystack.database.base import Document
 
@@ -9,3 +11,30 @@ class BaseReader(ABC):
     @abstractmethod
     def predict(self, question: str, documents: List[Document], top_k: Optional[int] = None):
         pass
+
+    @staticmethod
+    def _calc_no_answer(no_ans_gaps: Sequence[float], best_score_answer: float):
+        # "no answer" scores and positive answers scores are difficult to compare, because
+        # + a positive answer score is related to one specific document
+        # - a "no answer" score is related to all input documents
+        # Thus we compute the "no answer" score relative to the best possible answer and adjust it by
+        # the most significant difference between scores.
+        # Most significant difference: a model switching from predicting an answer to "no answer" (or vice versa).
+        # No_ans_gap is a list of this most significant difference per document
+        no_ans_gaps = np.array(no_ans_gaps)
+        max_no_ans_gap = np.max(no_ans_gaps)
+        # all passages "no answer" as top score
+        if (np.sum(no_ans_gaps < 0) == len(no_ans_gaps)):  # type: ignore
+            no_ans_score = best_score_answer - max_no_ans_gap  # max_no_ans_gap is negative, so it increases best pos score
+        else:  # case: at least one passage predicts an answer (positive no_ans_gap)
+            no_ans_score = best_score_answer - max_no_ans_gap
+
+        no_ans_prediction = {"answer": None,
+               "score": no_ans_score,
+               "probability": float(expit(np.asarray(no_ans_score) / 8)),  # just a pseudo prob for now
+               "context": None,
+               "offset_start": 0,
+               "offset_end": 0,
+               "document_id": None,
+               "meta": None,}
+        return no_ans_prediction, max_no_ans_gap

haystack/reader/farm.py

@@ -40,7 +40,7 @@ def __init__(
         context_window_size: int = 150,
         batch_size: int = 50,
         use_gpu: bool = True,
-        no_ans_boost: Optional[int] = None,
+        no_ans_boost: Optional[float] = None,
         top_k_per_candidate: int = 3,
         top_k_per_sample: int = 1,
         num_processes: Optional[int] = None,
@@ -446,32 +446,6 @@ def _check_no_answer(c: QACandidate):
             return False
 
 
-    @staticmethod
-    def _calc_no_answer(no_ans_gaps: List[float], best_score_answer: float):
-        # "no answer" scores and positive answers scores are difficult to compare, because
-        # + a positive answer score is related to one specific document
-        # - a "no answer" score is related to all input documents
-        # Thus we compute the "no answer" score relative to the best possible answer and adjust it by
-        # the most significant difference between scores.
-        # Most significant difference: a model switching from predicting an answer to "no answer" (or vice versa).
-        # No_ans_gap coming from FARM mean how much no_ans_boost should change to switch predictions
-        no_ans_gaps = np.array(no_ans_gaps)
-        max_no_ans_gap = np.max(no_ans_gaps)
-        # all passages "no answer" as top score
-        if (np.sum(no_ans_gaps < 0) == len(no_ans_gaps)):  # type: ignore
-            no_ans_score = best_score_answer - max_no_ans_gap  # max_no_ans_gap is negative, so it increases best pos score
-        else:  # case: at least one passage predicts an answer (positive no_ans_gap)
-            no_ans_score = best_score_answer - max_no_ans_gap
-
-        no_ans_prediction = {"answer": None,
-               "score": no_ans_score,
-               "probability": float(expit(np.asarray(no_ans_score) / 8)),  # just a pseudo prob for now
-               "context": None,
-               "offset_start": 0,
-               "offset_end": 0,
-               "document_id": None}
-        return no_ans_prediction, max_no_ans_gap
-
     def predict_on_texts(self, question: str, texts: List[str], top_k: Optional[int] = None):
         documents = []
         for text in texts:

haystack/reader/transformers.py

@@ -22,7 +22,7 @@ def __init__(
         tokenizer: str = "distilbert-base-uncased",
         context_window_size: int = 30,
         use_gpu: int = 0,
-        n_best_per_passage: int = 2,
+        top_k_per_candidate: int = 4,
         no_answer: bool = True
     ):
         """
@@ -40,14 +40,17 @@ def __init__(
                             The context usually helps users to understand if the answer really makes sense.
         :param use_gpu: < 0  -> use cpu
                         >= 0 -> ordinal of the gpu to use
-        :param n_best_per_passage: num of best answers to take into account for each passage
+        :param top_k_per_candidate: How many answers to extract for each candidate doc that is coming from the retriever (might be a long text).
+                                           Note: - This is not the number of "final answers" you will receive
+                                           (see `top_k` in TransformersReader.predict() or Finder.get_answers() for that)
+                                         - Can includes no_answer in the sorted list of predictions
         :param no_answer: True -> Hugging Face model could return an "impossible"/"empty" answer (i.e. when there is an unanswerable question)
                         False -> otherwise
 
         """
         self.model = pipeline('question-answering', model=model, tokenizer=tokenizer, device=use_gpu)
         self.context_window_size = context_window_size
-        self.n_best_per_passage = n_best_per_passage
+        self.top_k_per_candidate = top_k_per_candidate
         self.no_answer = no_answer
 
         # TODO context_window_size behaviour different from behavior in FARMReader
@@ -80,27 +83,49 @@ def predict(self, question: str, documents: List[Document], top_k: Optional[int]
         """
         # get top-answers for each candidate passage
         answers = []
+        no_ans_gaps = []
+        best_overall_score = 0
         for doc in documents:
             query = {"context": doc.text, "question": question}
-            predictions = self.model(query, topk=self.n_best_per_passage,handle_impossible_answer=self.no_answer)
+            predictions = self.model(query, topk=self.top_k_per_candidate, handle_impossible_answer=self.no_answer)
             # for single preds (e.g. via top_k=1) transformers returns a dict instead of a list
             if type(predictions) == dict:
                 predictions = [predictions]
             # assemble and format all answers
-            for pred in predictions:
-                context_start = max(0, pred["start"] - self.context_window_size)
-                context_end = min(len(doc.text), pred["end"] + self.context_window_size)
-                answers.append({
-                    "answer": pred["answer"],
-                    "context": doc.text[context_start:context_end],
-                    "offset_start": pred["start"],
-                    "offset_end": pred["end"],
-                    "probability": pred["score"],
-                    "score": None,
-                    "document_id": doc.id,
-                    "meta": doc.meta
-                })
 
+            best_doc_score = 0
+            # because we cannot ensure a "no answer" prediction coming back from transformers we initialize it here with 0
+            no_ans_doc_score = 0
+            # TODO add no answer bias on haystack side after getting "no answer" scores from transformers
+            for pred in predictions:
+                if pred["answer"]:
+                    if pred["score"] > best_doc_score:
+                        best_doc_score = pred["score"]
+                    context_start = max(0, pred["start"] - self.context_window_size)
+                    context_end = min(len(doc.text), pred["end"] + self.context_window_size)
+                    answers.append({
+                        "answer": pred["answer"],
+                        "context": doc.text[context_start:context_end],
+                        "offset_start": pred["start"],
+                        "offset_end": pred["end"],
+                        "probability": pred["score"],
+                        "score": None,
+                        "document_id": doc.id,
+                        "meta": doc.meta
+                    })
+                else:
+                    no_ans_doc_score = pred["score"]
+
+                if best_doc_score > best_overall_score:
+                    best_overall_score = best_doc_score
+
+            no_ans_gaps.append(no_ans_doc_score - best_doc_score)
+
+        # Calculate the score for predicting "no answer", relative to our best positive answer score
+        no_ans_prediction, max_no_ans_gap = self._calc_no_answer(no_ans_gaps, best_overall_score)
+
+        if self.no_answer:
+            answers.append(no_ans_prediction)
         # sort answers by their `probability` and select top-k
         answers = sorted(
             answers, key=lambda k: k["probability"], reverse=True

test/conftest.py

@@ -85,7 +85,7 @@ def no_answer_reader(request):
     if request.param == "transformers":
         return TransformersReader(model="deepset/roberta-base-squad2",
                                   tokenizer="deepset/roberta-base-squad2",
-                                  use_gpu=-1, n_best_per_passage=5)
+                                  use_gpu=-1, top_k_per_candidate=5)
 
 
 @pytest.fixture()