def parse_batch_preprocessed_data(self, preprocessed_input_examples, set_wiki, normalize_mod):
order,idsBatch,srcBatch, src_charBatch, sourceBatch,srcBertBatch,srcBertIndexBatch,data_iterator = self.feature_to_torch(preprocessed_input_examples)
probBatch, src_enc = self.model(srcBatch, src_charBatch, rel=False, bertBatch=srcBertBatch, bertIndexBatch = srcBertIndexBatch)
amr_pred_seq,concept_batches,aligns_raw,dependent_mark_batch = self.decoder.probAndSourceToConcepts(sourceBatch,srcBatch, src_charBatch, probBatch,getsense = True )
amr_pred_seq = [ [(uni.cat,uni.le,uni.aux,uni.sense,uni) for uni in seq ] for seq in amr_pred_seq ]
rel_batch,aligns = rel_to_batch(concept_batches,aligns_raw,data_iterator,self.dicts)
rel_prob,roots = self.model((rel_batch,srcBatch, src_charBatch, src_enc,aligns), rel=True, bertBatch=srcBertBatch)
graphs,rel_triples = self.decoder.relProbAndConToGraph(concept_batches, sourceBatch, rel_prob,roots,(dependent_mark_batch,aligns_raw),True,set_wiki,normalizeMod=normalize_mod)
batch_out = [0]*len(graphs)
batch_mrp_graphs = [0]*len(graphs)
for i,data in enumerate(zip(idsBatch, sourceBatch,amr_pred_seq,concept_batches,rel_triples,graphs)):
example_id, source,amr_pred,concept, rel_triple,graph= data
mrp_graph, predicated_graph = graph_to_mrpGraph(example_id, graph, normalizeMod = True, flavor=2, framework="amr", sentence=" ".join(source[0]))
out = []
out.append( "# ::id "+ example_id +"\n")
out.append( "# ::tok "+" ".join(source[0])+"\n")
out.append( "# ::lemma "+" ".join(source[1])+"\n")
out.append( "# ::pos "+" ".join(source[2])+"\n")
out.append( "# ::ner "+" ".join(source[3])+"\n")
out.append( self.decoder.nodes_jamr(graph))
out.append( self.decoder.edges_jamr(graph))
out.append( predicated_graph)
batch_out[order[i]] = "".join(out)+"\n"
batch_mrp_graphs[order[i]] = mrp_graph
return batch_mrp_graphs, batch_out
def parse_batch_preprocessed_data(self, preprocessed_input_examplesi,
set_wiki, normalize_mod):
"""
only mrp outputs, no txt outputs
"""
order, idsBatch, srcBatch, src_charBatch, sourceBatch, srcBertBatch, data_iterator = self.feature_to_torch(
preprocessed_input_examples)
probBatch, src_enc = self.model(srcBatch,
src_charBatch,
rel=False,
bertBatch=srcBertBatch,
bertIndexBatch=srcBertIndexBatch)
concepts_pred_seq, concept_batches, aligns_raw, dependent_mark_batch = self.decoder.probAndSourceToConcepts(
sourceBatch, srcBatch, src_charBatch, probBatch, getsense=True)
concepts_pred_seq = [[(uni.cat, uni.le, uni.aux, uni.sense, uni)
for uni in seq] for seq in concepts_pred_seq]
rel_batch, aligns = rel_to_batch(concept_batches, aligns_raw,
data_iterator, self.dicts)
rel_prob, roots = self.model(
(rel_batch, srcBatch, src_charBatch, src_enc, aligns),
rel=True,
bertBatch=srcBertBatch,
bertIndexBatch=srcBertIndexBatch)
graphs, rel_triples = self.decoderrelProbAndConToGraph(
concept_batches,
rel_prob,
roots, (dependent_mark_batch, aligns),
True,
set_wiki,
normalizeMod=normalize_mod)
batch_out = [0] * len(graphs)
batch_mrp_graphs = [0] * len(graphs)
for i, data in enumerate(
zip(idsBatch, sourceBatch, concepts_pred_seq, concept_batches,
rel_triples, graphs)):
example_id, source, amr_pred, concept, rel_triple, graph = data
mrp_graph, catd_graph = self.decoder.graph_to_mrpGraph(
example_id,
graph,
normalizeMod=opt.normalize_mod,
flavor=0,
framework="ucca",
sentence=" ".join(source[0]))
batch_mrp_graphs[order[i]] = mrp_graph
return batch_mrp_graphs, batch_out
def parse_batch(self, src_text_batch_or_data_batch):
if not self.parse_from_processed:
all_data = [
self.feature_extractor.preprocess(src_text)
for src_text in src_text_batch_or_data_batch
]
else:
all_data = src_text_batch_or_data_batch
order, srcBatch, sourceBatch, data_iterator = self.feature_to_torch(
all_data)
probBatch = self.model(srcBatch)
amr_pred_seq, concept_batches, aligns_raw, dependent_mark_batch = self.decoder.probAndSourceToAmr(
sourceBatch, srcBatch, probBatch, getsense=True)
amr_pred_seq = [[(uni.cat, uni.le, uni.aux, uni.sense, uni)
for uni in seq] for seq in amr_pred_seq]
rel_batch, aligns = rel_to_batch(concept_batches, aligns_raw,
data_iterator, self.dicts)
rel_prob, roots = self.model((rel_batch, srcBatch, aligns), rel=True)
graphs, rel_triples = self.decoder.relProbAndConToGraph(
concept_batches, rel_prob, roots,
(dependent_mark_batch, aligns_raw), True, True)
batch_out = [0] * len(graphs)
for i, data in enumerate(
zip(sourceBatch, amr_pred_seq, concept_batches, rel_triples,
graphs)):
source, amr_pred, concept, rel_triple, graph = data
predicated_graph = graph_to_amr(graph)
out = []
out.append("# ::tok " + " ".join(source[0]) + "\n")
out.append("# ::lemma " + " ".join(source[1]) + "\n")
out.append("# ::pos " + " ".join(source[2]) + "\n")
out.append("# ::ner " + " ".join(source[3]) + "\n")
out.append(self.decoder.nodes_jamr(graph))
out.append(self.decoder.edges_jamr(graph))
out.append(predicated_graph)
batch_out[order[i]] = "".join(out) + "\n"
return batch_out
def generate_graph(model,AmrDecoder, data_set,dicts,file):
concept_scores = concept_score_initial(dicts)
rel_scores = rel_scores_initial()
model.eval()
AmrDecoder.eval()
output = []
gold_file = []
for batchIdx in range(len(data_set)):
order,srcBatch,_,_,_,_,_,gold_roots,sourceBatch =data_set[batchIdx]
probBatch = model(srcBatch )
amr_pred_seq,concept_batches,aligns_raw,dependent_mark_batch = AmrDecoder.probAndSourceToAmr(sourceBatch,srcBatch,probBatch,getsense = opt.get_sense )
amr_pred_seq = [ [(uni.cat,uni.le,uni.aux,uni.sense,uni) for uni in seq ] for seq in amr_pred_seq ]
rel_batch,aligns = rel_to_batch(concept_batches,aligns_raw,data_set,dicts)
rel_prob,roots = model((rel_batch,srcBatch,aligns),rel=True)
graphs,rel_triples = AmrDecoder.relProbAndConToGraph(concept_batches,rel_prob,roots,(dependent_mark_batch,aligns_raw),opt.get_sense,opt.get_wiki)
batch_out = [0]*len(graphs)
for score_h in rel_scores:
if score_h.second_filter:
t,p,tp = score_h.T_P_TP_Batch(rel_triples,list(zip(*sourceBatch))[5],second_filter_material = (concept_batches,list(zip(*sourceBatch))[4]))
else:
t,p,tp = score_h.T_P_TP_Batch(rel_triples,list(zip(*sourceBatch))[5])
for score_h in concept_scores:
t,p,tp = score_h.T_P_TP_Batch(concept_batches,list(zip(*sourceBatch))[4])
for i,data in enumerate(zip( sourceBatch,amr_pred_seq,concept_batches,rel_triples,graphs)):
source,amr_pred,concept, rel_triple,graph= data
predicated_graph = graph_to_amr(graph)
out = []
out.append( "# ::tok "+" ".join(source[0])+"\n")
out.append( "# ::lem "+" ".join(source[1])+"\n")
out.append( "# ::pos "+" ".join(source[2])+"\n")
out.append( "# ::ner "+" ".join(source[3])+"\n")
out.append( "# ::predicated "+" ".join([str(re_cat[-1]) for re_cat in amr_pred])+"\n")
out.append( "# ::transformed final predication "+" ".join([str(c) for c in concept])+"\n")
out.append( AmrDecoder.nodes_jamr(graph))
out.append( AmrDecoder.edges_jamr(graph))
out.append( predicated_graph)
batch_out[order[i]] = "".join(out)+"\n"
output += batch_out
t_p_tp = list(map(lambda a,b:a+b, concept_scores[1].t_p_tp,rel_scores[1].t_p_tp))
total_out = "Smatch"+"\nT,P,TP: "+ " ".join([str(i) for i in t_p_tp])+"\nPrecesion,Recall,F1: "+ " ".join([str(i)for i in P_R_F1(*t_p_tp)])
print(total_out)
for score_h in rel_scores:
print("")
print(score_h)
file = file.replace(".pickle",".txt")
with open(file+ opt.output, 'w+') as the_file:
for data in output:
the_file.write(data+'\n')
print(file+ opt.output+" written.")
return concept_scores,rel_scores,output