Refactor Coval Scoring code

licenses/3rd_party_licenses.txt

@@ -127,3 +127,36 @@ distributed under the License is distributed on an "AS IS" BASIS,
 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License.
+
+
+coval
+-----
+
+* Files: scorer.py
+
+The implementations of ClusterEvaluator, lea, get_cluster_info, and
+get_markable_assignments are adapted from coval, which is distributed
+under the following license:
+
+The MIT License (MIT)
+
+Copyright 2018 Nafise Sadat Moosavi (ns.moosavi at gmail dot com)
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+

spacy/pipeline/coref.py

@@ -24,7 +24,7 @@
     doc2clusters,
 )
 
-from ..coref_scorer import Evaluator, get_cluster_info, lea
+from ..scorer import ClusterEvaluator, get_cluster_info, lea
 
 
 default_config = """
@@ -314,7 +314,7 @@ def score(self, examples, **kwargs):
         https://api.semanticscholar.org/CorpusID:17606580
         """
 
-        evaluator = Evaluator(lea)
+        evaluator = ClusterEvaluator(lea)
 
         for ex in examples:
             p_clusters = doc2clusters(ex.predicted, self.span_cluster_prefix)

spacy/scorer.py

@@ -1143,3 +1143,144 @@ def _auc(x, y):
         # regular numpy.ndarray instances.
         area = area.dtype.type(area)
     return area
+
+
+def get_cluster_info(predicted_clusters, gold_clusters):
+    p2g = get_markable_assignments(predicted_clusters, gold_clusters)
+    g2p = get_markable_assignments(gold_clusters, predicted_clusters)
+    # this is the data format used as input by the evaluator
+    return (gold_clusters, predicted_clusters, g2p, p2g)
+
+
+def get_markable_assignments(in_clusters, out_clusters):
+    markable_cluster_ids = {}
+    out_dic = {}
+    for cluster_id, cluster in enumerate(out_clusters):
+        for m in cluster:
+            out_dic[m] = cluster_id
+
+    for cluster in in_clusters:
+        for im in cluster:
+            for om in out_dic:
+                if im == om:
+                    markable_cluster_ids[im] = out_dic[om]
+                    break
+
+    return markable_cluster_ids
+
+
+class ClusterEvaluator:
+    def __init__(self, metric, beta=1, keep_aggregated_values=False):
+        self.p_num = 0
+        self.p_den = 0
+        self.r_num = 0
+        self.r_den = 0
+        self.metric = metric
+        self.beta = beta
+        self.keep_aggregated_values = keep_aggregated_values
+
+        if keep_aggregated_values:
+            self.aggregated_p_num = []
+            self.aggregated_p_den = []
+            self.aggregated_r_num = []
+            self.aggregated_r_den = []
+
+    def update(self, coref_info):
+        (
+            key_clusters,
+            sys_clusters,
+            key_mention_sys_cluster,
+            sys_mention_key_cluster,
+        ) = coref_info
+
+        pn, pd = self.metric(sys_clusters, key_clusters, sys_mention_key_cluster)
+        rn, rd = self.metric(key_clusters, sys_clusters, key_mention_sys_cluster)
+        self.p_num += pn
+        self.p_den += pd
+        self.r_num += rn
+        self.r_den += rd
+
+        if self.keep_aggregated_values:
+            self.aggregated_p_num.append(pn)
+            self.aggregated_p_den.append(pd)
+            self.aggregated_r_num.append(rn)
+            self.aggregated_r_den.append(rd)
+
+    def f1(self, p_num, p_den, r_num, r_den, beta=1):
+        p = 0
+        if p_den != 0:
+            p = p_num / float(p_den)
+        r = 0
+        if r_den != 0:
+            r = r_num / float(r_den)
+
+        if p + r == 0:
+            return 0
+
+        return (1 + beta * beta) * p * r / (beta * beta * p + r)
+
+    def get_f1(self):
+        return self.f1(self.p_num, self.p_den, self.r_num, self.r_den, beta=self.beta)
+
+    def get_recall(self):
+        if self.r_num == 0:
+            return 0
+
+        return self.r_num / float(self.r_den)
+
+    def get_precision(self):
+        if self.p_num == 0:
+            return 0
+
+        return self.p_num / float(self.p_den)
+
+    def get_prf(self):
+        return self.get_precision(), self.get_recall(), self.get_f1()
+
+    def get_counts(self):
+        return self.p_num, self.p_den, self.r_num, self.r_den
+
+    def get_aggregated_values(self):
+        return (
+            self.aggregated_p_num,
+            self.aggregated_p_den,
+            self.aggregated_r_num,
+            self.aggregated_r_den,
+        )
+
+
+def lea(input_clusters, output_clusters, mention_to_gold):
+    """
+    LEA is a metric for scoring coref clusters design to avoid pitfals of prior
+    methods. Proposed in Moosavi and Strube 2016.
+
+    https://api.semanticscholar.org/CorpusID:17606580
+    """
+    num, den = 0, 0
+
+    for c in input_clusters:
+        if len(c) == 1:
+            all_links = 1
+            if (
+                c[0] in mention_to_gold
+                and len(output_clusters[mention_to_gold[c[0]]]) == 1
+            ):
+                common_links = 1
+            else:
+                common_links = 0
+        else:
+            common_links = 0
+            all_links = len(c) * (len(c) - 1) / 2.0
+            for i, m in enumerate(c):
+                if m in mention_to_gold:
+                    for m2 in c[i + 1 :]:
+                        if (
+                            m2 in mention_to_gold
+                            and mention_to_gold[m] == mention_to_gold[m2]
+                        ):
+                            common_links += 1
+
+        num += len(c) * common_links / float(all_links)
+        den += len(c)
+
+    return num, den