Mrp 2020 (#106)

* check for cycles

* add support for 2020 mrp format

* merge changes

* fix

* CR fixes

* remove print

Co-authored-by: Ofir Arviv <t-ofarvi@microsoft.com>

Author: OfirArviv

requirements.txt

@@ -6,3 +6,5 @@ dynet==2.1
 logbook>=1.5.2
 word2number>=1.1
 git+https://github.com/cfmrp/mtool.git#egg=mtool
+networkx==2.4
+matplotlib == 3.2.1

tupa/config.py

@@ -22,9 +22,9 @@
 COMPOUND = "compound"
 
 # Required number of edge labels per framework
-NODE_LABELS_NUM = {"amr": 1000, "dm": 1000, "psd": 1000, "eds": 1000, "ucca": 0}
-NODE_PROPERTY_NUM = {"amr": 1000, "dm": 510, "psd": 1000, "eds": 1000, "ucca": 0}
-EDGE_LABELS_NUM = {"amr": 141, "dm": 59, "psd": 90, "eds": 10, "ucca": 15}
+NODE_LABELS_NUM = {"amr": 1000, "dm": 1000, "psd": 1000, "eds": 1000, "ucca": 0, "ptg": 1000}
+NODE_PROPERTY_NUM = {"amr": 1000, "dm": 510, "psd": 1000, "eds": 1000, "ucca": 0, "ptg": 1000}
+EDGE_LABELS_NUM = {"amr": 141, "dm": 59, "psd": 90, "eds": 10, "ucca": 15, "ptg": 150}
 EDGE_ATTRIBUTE_NUM = {"amr": 0, "dm": 0, "psd": 0, "eds": 0, "ucca": 2}
 NN_ARG_NAMES = set()
 DYNET_ARG_NAMES = set()
@@ -537,20 +537,20 @@ def __str__(self):
 
 
 def requires_node_labels(framework):
-    return framework != "ucca"
+    return framework not in ("ucca", "drg", "ptg")
 
 
 def requires_node_properties(framework):
-    return framework != "ucca"
+    return framework not in ("ucca", "drg")
 
 
 def requires_edge_attributes(framework):
     return framework == "ucca"
 
 
 def requires_anchors(framework):
-    return framework != "amr"
+    return framework not in ("amr", "drg")
 
 
 def requires_tops(framework):
-    return framework in ("ucca", "amr")
+    return framework in ("ucca", "amr", "drg", "ptg")

tupa/constraints/ptg.py

@@ -0,0 +1,6 @@
+from .validation import Constraints
+
+
+class PtgConstraints(Constraints):
+    def __init__(self, **kwargs):
+        super().__init__(multigraph=True, **kwargs)

tupa/constraints/validation.py

@@ -155,12 +155,18 @@ def eds_constraints(**kwargs):
     return EdsConstraints(**kwargs)
 
 
+def ptg_constraints(**kwargs):
+    from .ptg import PtgConstraints
+    return PtgConstraints(**kwargs)
+
+
 CONSTRAINTS = {
     "ucca":   ucca_constraints,
     "amr":    amr_constraints,
     "dm":     sdp_constraints,
     "psd":    sdp_constraints,
     "eds":    eds_constraints,
+    "ptg":    ptg_constraints,
 }
 
 

tupa/oracle.py

@@ -193,8 +193,12 @@ def get_node_label(self, state, node):
         return true_label, raw_true_label
 
     def get_node_property_value(self, state, node):
-        true_property_value = next((k, v) for k, v in node.ref_node.properties.items()
-                                   if k not in (node.properties or ()))
+        try:
+            true_property_value = next((k, v) for k, v in (node.ref_node.properties.items()
+                                                           if node.ref_node.properties else [])
+                                       if k not in (node.properties or ()))
+        except StopIteration:
+            return None
         if self.args.validate_oracle:
             try:
                 state.check_valid_property_value(true_property_value, message=True)

tupa/parse.py

@@ -57,6 +57,7 @@ def tokens_per_second(self):
 
 class GraphParser(AbstractParser):
     """ Parser for a single graph, has a state and optionally an oracle """
+
     def __init__(self, graph, *args, target=None, **kwargs):
         """
         :param graph: gold Graph to get the correct nodes and edges from (in training), or just to get id from (in test)
@@ -74,20 +75,54 @@ def __init__(self, graph, *args, target=None, **kwargs):
             assert self.lang, "Attribute 'lang' is required per passage when using multilingual BERT"
         self.state_hash_history = set()
         self.state = self.oracle = None
-        if self.framework == "amr" and self.alignment:  # Copy alignments to anchors, updating graph
+        if self.framework in ("amr", "drg", "ptg") and self.alignment:  # Copy alignments to anchors, updating graph
             for alignment_node in self.alignment.nodes:
                 node = self.graph.find_node(alignment_node.id)
                 if node is None:
                     self.config.log("graph %s: invalid alignment node %s" % (self.graph.id, alignment_node.id))
                     continue
                 if node.anchors is None:
                     node.anchors = []
-                for conllu_node_id in (alignment_node.label or []) + list(chain(*alignment_node.values or [])):
-                    conllu_node = self.conllu.find_node(conllu_node_id)
-                    if conllu_node is None:
-                        raise ValueError("Alignments incompatible with tokenization: token %s "
-                                         "not found in graph %s" % (conllu_node_id, self.graph.id))
-                    node.anchors += conllu_node.anchors
+
+                conllu_node_id_list = None
+                alignment_node_anchor_char_range_list = None
+                if self.alignment.framework == "alignment":
+                    conllu_node_id_list = (alignment_node.label or []) + list(chain(*alignment_node.values or []))
+                elif self.alignment.framework == "anchoring" and self.framework in ("amr", "ptg"):
+                    conllu_node_id_list = set([alignment_dict["#"] for alignment_dict in
+                                               (alignment_node.anchors or [])
+                                               + ([anchor for anchor_list in (alignment_node.anchorings or []) for anchor in anchor_list])])
+                elif self.alignment.framework == "anchoring" and self.framework == "drg":
+                    alignment_node_anchor_char_range_list = [(int(alignment_dict["from"]),(int(alignment_dict["to"]))) for alignment_dict in
+                                                                 (alignment_node.anchors or [])
+                                                                 + ([anchor for anchor_list in (alignment_node.anchorings or []) for anchor in anchor_list])]
+                    assert all([len(conllu_node.anchors) == 1 for conllu_node in self.conllu.nodes])
+                    anchors_to_conllu_node = {(int(conllu_node.anchors[0]["from"]), int(conllu_node.anchors[0]["to"])):
+                                                  conllu_node
+                                              for conllu_node in self.conllu.nodes}
+                else:
+                    raise ValueError(f'Unknown alignments framework: {alignment_node.framework}')
+
+                if conllu_node_id_list is not None:
+                    assert self.framework in ("amr", "ptg")
+                    for conllu_node_id in conllu_node_id_list:
+                        conllu_node = self.conllu.find_node(conllu_node_id + 1)
+
+                        if conllu_node is None:
+                            raise ValueError("Alignments incompatible with tokenization: token %s "
+                                             "not found in graph %s" % (conllu_node_id, self.graph.id))
+                        node.anchors += conllu_node.anchors
+
+                elif alignment_node_anchor_char_range_list is not None:
+                    for alignment_node_char_range in alignment_node_anchor_char_range_list:
+                        for conllu_anchor_range in anchors_to_conllu_node:
+                            if alignment_node_char_range[0] <= conllu_anchor_range[0] \
+                                    and alignment_node_char_range[1] >= conllu_anchor_range[1]:
+                                conllu_node = anchors_to_conllu_node[conllu_anchor_range]
+                                if conllu_node is None:
+                                    raise ValueError("Alignments incompatible with tokenization: token %s "
+                                                     "not found in graph %s" % (conllu_anchor_range, self.graph.id))
+                                node.anchors += conllu_node.anchors
 
     def init(self):
         self.config.set_framework(self.framework)
@@ -320,6 +355,7 @@ def num_tokens(self, _):
 
 class BatchParser(AbstractParser):
     """ Parser for a single training iteration or single pass over dev/test graphs """
+
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
         self.seen_per_framework = defaultdict(int)
@@ -335,14 +371,11 @@ def parse(self, graphs, display=True, write=False, accuracies=None):
             if conllu is None:
                 self.config.print("skipped '%s', no companion conllu data found" % graph.id)
                 continue
-            alignment = self.alignment.get(graph.id)
+            alignment = self.alignment.get(graph.id) if self.alignment else None
             for target in graph.targets() or [graph.framework]:
                 if not self.training and target not in self.model.classifier.labels:
                     self.config.print("skipped target '%s' for '%s': did not train on it" % (target, graph.id), level=1)
                     continue
-                if target == "amr" and alignment is None:
-                    self.config.print("skipped target 'amr' for '%s': no companion alignment found" % graph.id, level=1)
-                    continue
                 parser = GraphParser(
                     graph, self.config, self.model, self.training, conllu=conllu, alignment=alignment, target=target)
                 if self.config.args.verbose and display:
@@ -403,6 +436,7 @@ def time_per_graph(self):
 
 class Parser(AbstractParser):
     """ Main class to implement transition-based meaning representation parser """
+
     def __init__(self, model_file=None, config=None, training=None, conllu=None, alignment=None):
         super().__init__(config=config or Config(), model=Model(model_file or config.args.model),
                          training=config.args.train if training is None else training,
@@ -646,7 +680,7 @@ def read_graphs_with_progress_bar(file_handle_or_graphs):
     if isinstance(file_handle_or_graphs, IOBase):
         graphs, _ = read_graphs(
             tqdm(file_handle_or_graphs, desc="Reading " + getattr(file_handle_or_graphs, "name", "input"),
-                 unit=" graphs"), format="mrp")
+                 unit=" graphs"), format="mrp", robust=True)
         return graphs
     return file_handle_or_graphs
 

tupa/states/ref_graph.py

@@ -1,10 +1,14 @@
+import sys
+from typing import List, Tuple
+
 from .anchors import expand_anchors
 from .edge import StateEdge
 from .node import StateNode
 from ..constraints.amr import NAME
 from ..constraints.validation import ROOT_ID, ROOT_LAB, ANCHOR_LAB
 from ..recategorization import resolve, compress_name
 
+import networkx as nx
 
 class RefGraph:
     def __init__(self, graph, conllu, framework):
@@ -27,6 +31,7 @@ def __init__(self, graph, conllu, framework):
         offset = len(conllu.nodes) + 1
         self.non_virtual_nodes = []
         self.edges = []
+        have_anchors = False
         for graph_node in graph.nodes:
             node_id = graph_node.id + offset
             id2node[node_id] = node = \
@@ -43,7 +48,21 @@ def __init__(self, graph, conllu, framework):
                     anchor_terminals = [min(self.terminals, key=lambda terminal: min(
                         x - y for x in terminal.anchors for y in node.anchors))]  # Must have anchors, get closest one
                 for terminal in anchor_terminals:
+                    have_anchors = True
                     self.edges.append(StateEdge(node, terminal, ANCHOR_LAB).add())
+
+        if not have_anchors:
+            print(f'framework {graph.framework} graph id {graph.id} have no anchors', file=sys.stderr)
+
+        cycle = find_cycle(graph)
+        while len(cycle) > 0:
+            edge_list = list(graph.edges)
+            first_edge_idx = \
+            [i for i, edge in enumerate(graph.edges) if edge.src == cycle[0][0] and edge.tgt == cycle[0][1]][0]
+            del edge_list[first_edge_idx]
+            graph.edges = set(edge_list)
+            cycle = find_cycle(graph)
+
         for edge in graph.edges:
             if edge.src != edge.tgt:  # Drop self-loops as the parser currently does not support them
                 self.edges.append(StateEdge(id2node[edge.src + offset],
@@ -55,4 +74,30 @@ def __init__(self, graph, conllu, framework):
                     node.properties = compress_name(node.properties)
                 node.properties = {prop: resolve(node, value, introduce_placeholders=True)
                                    for prop, value in node.properties.items()}
+
             node.label = resolve(node, node.label, introduce_placeholders=True)  # Must be after properties in case NAME
+
+
+def find_cycle(graph, plot_graph=False) -> List[Tuple[int, int]]:
+    edges_tuple = [(e.src, e.tgt) for e in graph.edges]
+    nx_graph = nx.DiGraph()
+    nx_graph.add_edges_from(edges_tuple)
+    try:
+        cycle = nx.find_cycle(nx_graph)
+    except nx.exception.NetworkXNoCycle as e:
+        cycle = []
+
+    if plot_graph:
+        import matplotlib.pyplot as plt
+        nx.draw(nx_graph, with_labels=True, font_weight='bold')
+        plt.show()
+
+    return cycle
+
+
+def is_directed_acyclic_graph(graph) -> bool:
+    edges_tuple = list(map(lambda x: (x.src, x.tgt), graph.edges))
+    nx_graph = nx.DiGraph()
+    nx_graph.add_edges_from(edges_tuple)
+
+    assert nx.is_directed_acyclic_graph(nx_graph)