Add frequency-based noun pruning to DisCoCircReader

lambeq/experimental/discocirc/coref_resolver.py

@@ -78,7 +78,8 @@ def dict_from_corefs(self,
                        for scrf in scoref]
 
             for scoref in scorefs:
-                corefd[scoref] = scorefs[0]
+                if scoref not in corefd:
+                    corefd[scoref] = scorefs[0]
 
         return corefd
 
@@ -87,14 +88,21 @@ class SpacyCoreferenceResolver(CoreferenceResolver):
     """Corefence resolution and tokenisation based on spaCy."""
 
     def __init__(self):
-        self.nlp = spacy.load('en_coreference_web_trf',
-                              exclude=('span_resolver', 'span_cleaner'))
+        # Create basic tokenisation pipeline, for POS
+        self.nlp = spacy.load('en_core_web_sm')
+
+        # Add coreference resolver pipe stage
+        coref_stage = spacy.load('en_coreference_web_trf',
+                                 exclude=('span_resolver', 'span_cleaner'))
+        self.nlp.add_pipe('transformer', source=coref_stage)
+        self.nlp.add_pipe('coref', source=coref_stage)
 
     def tokenise_and_coref(self, text):
         text = self._clean_text(text)
         doc = self.nlp(text)
         coreferences = []
 
+        # Add all coreference instances
         for cluster in doc.spans.values():
             sent_clusters = [[] for _ in doc.sents]
             for span in cluster:
@@ -104,4 +112,14 @@ def tokenise_and_coref(self, text):
                         break
             coreferences.append(sent_clusters)
 
+        # Add trivial coreferences for all nouns, determined by spacy POS
+        spacy_noun_pos = {'NOUN', 'PROPN', 'PRON'}
+
+        for i, sent in enumerate(doc.sents):
+            for tok in sent:
+                if tok.pos_ in spacy_noun_pos:
+                    sent_clusters = [[] for _ in doc.sents]
+                    sent_clusters[i] = [tok.i - sent.start]
+                    coreferences.append(sent_clusters)
+
         return [[str(w) for w in s] for s in doc.sents], coreferences

lambeq/experimental/discocirc/reader.py

@@ -290,7 +290,7 @@ def text2circuit(self,
                      sandwich: bool = False,
                      break_cycles: bool = True,
                      pruned_nouns: Iterable[str] = (),
-                     min_noun_freq: int = 0,
+                     min_noun_freq: int = 1,
                      rewrite_rules: (
                          Iterable[TreeRewriteRule | str] | None
                      ) = ('determiner', 'auxiliary'),
@@ -313,6 +313,9 @@ def text2circuit(self,
             If any of the nouns in this list are present in the diagram,
             the corresponding state and wire are removed from the
             diagram.
+        min_noun_freq: int, default: 1
+            Mininum number of times a noun needs to be referenced to
+            appear in the circuit.
         rewrite_rules : list of `TreeRewriteRule` or str
             List of rewrite rules to apply to the pregroup tree
             before conversion to a circuit.
@@ -329,10 +332,16 @@ def text2circuit(self,
         """
 
         sentences, corefs = self.coref_resolver.tokenise_and_coref(text)
+        corefd = self.coref_resolver.dict_from_corefs(corefs)
 
-        pruned_ids = self._prune_indices(sentences, corefs, pruned_nouns)
+        noun_counts = Counter(corefd.values())
+        freq_pruned_ids = [nid for nid, count in noun_counts.items()
+                           if count < min_noun_freq]
 
-        corefd = self.coref_resolver.dict_from_corefs(corefs)
+        pruned_nouns = set(pruned_nouns).union(
+                        {sentences[i][j] for (i, j) in freq_pruned_ids})
+
+        pruned_ids = self._prune_indices(sentences, corefs, pruned_nouns)
 
         rewriter = TreeRewriter(rewrite_rules)
 

setup.cfg

@@ -88,6 +88,7 @@ experimental =
     numpy == 1.26.4
     spacy ~= 3.4.0
     spacy-experimental ~= 0.6.4
+    en-core-web-sm
     en-coreference-web-trf @ https://github.com/explosion/spacy-experimental/releases/download/v0.6.1/en_coreference_web_trf-3.4.0a2-py3-none-any.whl
 
 [options.entry_points]