--- a
+++ b/tests/readme.md
@@ -0,0 +1,20 @@
+# Testing the algorithm
+
+Various tests for the components of the spaCy pipeline.
+
+We decided to design tests entity-wise, meaning that we only check the validity
+of the computed modality on a set of entities. This design choice is motivated by
+the fact that :
+
+1. That's what we actually care about. We want our pipeline to detect negation,
+   family context, patient history and hypothesis relative to a given entity.
+
+2. Deciding on the span of an annotation (negated, hypothesis, etc) is tricky.
+   Consider the example : `"Le patient n'est pas malade."`. Should the negated span
+   correspond to `["est", "malade"]`, `["malade"]`, `["n'", "est", "pas", "malade", "."]` ?
+
+3. Depending on the design of the algorithm, the span might be off, even though it
+   can correctly assign polarity to a given entity (but considered that the punctuation
+   was negated as well).
+   By relaxing the need to infer the correct span, we avoid giving an unfair disadvantage
+   to an otherwise great algorithm.