script to parse results into tsv

code/contextual_embeddings/configs/bert.jsonnet

@@ -67,6 +67,6 @@
     "num_serialized_models_to_keep": 3,
     "num_epochs": 10,
     "grad_norm": 5.0,
-    "cuda_device": 0
+    "cuda_device": -1
   }
 }

code/ssplit/README.md

+#Requirements 
+
+`ersatz` (`pip install ersatz`)
+
+#Usage
+
+`python tok2conllu.py <file>` 
+
+File must be a `.tok` file.
+
+#Output
+
+`<file>_conllu`: same file but sentences are separated by a line-jump and a commented line with sentence id.
+
+#Ersatz library
+
+Command-line usage:
+
+`ersatz --input <input.txt> --output <output.txt>`
+
+Takes as input any text file and outputs the same text file with sentences separated by a line-break.
+
+

code/ssplit/parse_stanza.py

@@ -65,4 +65,4 @@ def ssplit_stanza(lang, fp_toks, out_dir, treebank=None):
                     f_out.write('\t'.join(row)+'\n')
                     #print('\t'.join(row).encode('utf-8'), file=f_out)
                 f_out.write('\n')
-                #print('', file=f_out)
\ No newline at end of file
+                #print('', file=f_out)

code/ssplit/tok2conllu.py

+import sys
+import pandas as pd
+import os
+
+tab = "\t"
+space = " "
+
+def parse_file(f):
+    """Take a .tok file and turn it into a sequence of token ids and tokens (.tok_seq). Token id precedes token."""
+
+    column_names = ['tok_id', 'tok', '1', '2', '3', '4', '5', '6', '7', 'seg']
+
+    dataframe = pd.read_csv(f, names=column_names, comment="#", sep="\t",skipinitialspace=True)
+    tok_ids = dataframe['tok_id'].values
+    toks = dataframe['tok'].values
+
+    return dataframe, tok_ids, toks
+
+def write_seq_file(f, tok_ids, toks):
+    """Write sequence of token ids and tokens to a .tok_ssplit file"""
+
+    with open(f, "w") as wf:
+        for i, tok in zip(tok_ids, toks):
+            wf.write(str(i) + tab + tok + tab)
+
+def parse_ssplit_file(f):
+    """Take a .tok_ssplit file and return ids of sentence-beginning tokens."""
+
+    with open(f, "r") as rf:
+        sentences = rf.readlines()
+
+    sstart_ids = [0] * len(sentences)
+    for i, sentence in enumerate(sentences):
+        ids_toks = sentence.strip().split(space) 
+        sstart_ids[i] = ids_toks[0]
+
+    return sstart_ids
+
+def make_ssplit(rf, wf, sstart_ids):
+    """Write new file with sentence boundaries"""
+    with open(rf, "r") as f:
+        lines = f.readlines()
+
+    doc_id = None
+    next_sentence = 0 #index of token beginning next sentence in sstart_ids
+    sent_counter = 0
+
+    with open(wf, "w") as f:
+        for line in lines:
+            split = line.strip().split(tab)
+            tok_id = split[0]
+            if tok_id.startswith("#"):
+                doc_id = line
+                sent_counter = 0
+                f.write(line)
+            elif tok_id == sstart_ids[next_sentence]:
+                doc_id_nb = doc_id.strip().split("= ")[1]
+                if sent_counter: newline = "\n" 
+                else: newline = ""
+                sent_counter += 1
+                sent_id = "# sent_id = " + doc_id_nb + "-" + str(sent_counter)
+                f.write(newline + sent_id + "\n")
+                f.write(line)
+                next_sentence += 1
+            else:
+                f.write(line)
+
+def t2c(f):
+    dataframe, tok_ids, toks = parse_file(f)
+    f_seq = f + "_seq"
+    write_seq_file(f_seq, tok_ids, toks)
+    f_ssplit = f + "_ssplit"
+    os.system(f"ersatz --input {f_seq} --output {f_ssplit}")
+    sstart_ids = parse_ssplit_file(f_ssplit)
+    f_conllu = f + "_conllu"
+    make_ssplit(f, f_conllu, sstart_ids)
+    os.system(f"rm {f_seq} {f_ssplit}") #remove temporary files
+
+def main():
+    if len(sys.argv) < 2: 
+        print("usage: python tok2conllu.py <file>")
+        sys.exit()
+    for f in sys.argv[1:]:
+        t2c(f)
+
+if __name__ == '__main__':
+    main()

code/utils/get_scores.py

+import sys
+import os
+import pandas as pd
+
+#Must be executed from discut dir
+
+def parse_score(line):
+	split = line.strip().split(": ")
+	return str((round(float(split[-1]), 4) * 100))[:5]
+
+def parse_model(line):
+	split = line.strip().split("_")
+	model_ft = split[1].strip().split("-")
+	if len(model_ft) > 1: model = model_ft 
+	else: model = [model_ft[0], "n/a"]
+	return model
+
+def parse_test(line):
+	split = line.strip().split("_")
+	return split[0]
+
+def write_scores(result_dir, output_file):
+	if not os.path.isdir(result_dir):
+		print(f"directory {result_dir} doesn't exist.")
+
+	_, dirs, _ = next(iter(os.walk(result_dir)))
+	table = []
+	
+	for d in dirs:
+		files = os.listdir(os.path.join(result_dir, d))
+		for fil in files:
+			if fil.endswith('test.scores'):
+				with open(os.path.join(result_dir, d, fil), "r") as r_f:
+					lines = r_f.readlines()
+				precision, recall, fscore = [parse_score(l) for l in lines[-3:]]
+				model = parse_model(d)
+				entry = [model[0], model[1], parse_test(fil), precision, recall, fscore]
+				table.append(entry)
+
+	columns = ['Model', 'Fine-tuning', 'Test', 'Precision', 'Recall', 'F1']
+	df = pd.DataFrame(table, columns = columns)
+	df.to_csv((output_file + ".tsv"), sep="\t", index=False)
+
+def main():
+	if len(sys.argv) != 3:
+		print("usage: python get_scores.py <result_dir> <output_file>"); sys.exit()
+	else:
+		result_dir = sys.argv[1]
+		output_file = sys.argv[2]
+	write_scores(result_dir, output_file)
+
+if __name__ == '__main__':
+	main()
+	

scores_expes/220622_conllu.tsv

+Model	Fine-tuning	Test	Precision	Recall	F1
+fra.sdrt.annodis	n/a	fra.sdrt.annodis	86.85	88.67	87.75
+fra.sdrt.annodis	n/a	deu.rst.pcc	72.96	97.28	83.38
+fra.sdrt.annodis	n/a	eng.sdrt.stac	94.42	89.72	92.01
+fra.sdrt.annodis	n/a	eus.rst.ert	56.98	87.7	69.08
+fra.sdrt.annodis	n/a	fas.rst.prstc	79.5	86.27	82.75
+deu.rst.pcc	n/a	deu.rst.pcc	95.62	96.6	96.11
+deu.rst.pcc	n/a	rus.rst.rrt	69.38	77.88	73.38
+deu.rst.pcc	n/a	zho.rst.sctb	54.76	95.83	69.69
+deu.rst.pcc	n/a	eng.sdrt.stac	98.34	88.17	92.97
+deu.rst.pcc	n/a	eus.rst.ert	77.85	79.32	78.58
+deu.rst.pcc	n/a	fas.rst.prstc	91.57	68.06	78.08
+deu.rst.pcc	n/a	fra.sdrt.annodis	86.00	66.67	75.11
+deu.rst.pcc	n/a	nld.rst.nldt	92.55	88.17	90.3
+deu.rst.pcc	n/a	spa.rst.rststb	71.31	91.3	80.08
+deu.rst.pcc	n/a	por.rst.cstn	82.44	70.59	76.06
+eus.rst.ert	n/a	deu.rst.pcc	87.06	91.5	89.22
+eus.rst.ert	n/a	eng.sdrt.stac	99.05	87.2	92.75
+eus.rst.ert	n/a	eus.rst.ert	90.06	83.24	86.52
+eus.rst.ert	n/a	fas.rst.prstc	85.34	68.66	76.1
+fas.rst.prstc	n/a	deu.rst.pcc	81.55	93.2	86.98
+fas.rst.prstc	n/a	eng.sdrt.stac	98.76	87.46	92.77
+fas.rst.prstc	n/a	eus.rst.ert	74.49	79.32	76.83
+fas.rst.prstc	n/a	fas.rst.prstc	92.38	92.24	92.31
+deu.rst.pcc	deu.rst.pcc	deu.rst.pcc	95.88	94.89	95.38
+deu.rst.pcc	eus.rst.ert	deu.rst.pcc	98.58	94.56	96.53
+deu.rst.pcc	fas.rst.prstc	deu.rst.pcc	94.65	96.26	95.45
+deu.rst.pcc	fra.sdrt.annodis	deu.rst.pcc	96.86	94.56	95.7
+eng.sdrt.stac	deu.rst.pcc	eng.sdrt.stac	94.69	95.67	95.17
+eng.sdrt.stac	eus.rst.ert	eng.sdrt.stac	96.46	93.47	94.94
+eng.sdrt.stac	fas.rst.prstc	eng.sdrt.stac	96.41	94.05	95.22
+eng.sdrt.stac	fra.sdrt.annodis	eng.sdrt.stac	96.82	94.51	95.65
+eus.rst.ert	deu.rst.pcc	eus.rst.ert	90.62	82.3	86.26
+eus.rst.ert	eus.rst.ert	eus.rst.ert	87.31	86.49	86.9
+eus.rst.ert	fas.rst.prstc	eus.rst.ert	89.55	81.08	85.11
+eus.rst.ert	fra.sdrt.annodis	eus.rst.ert	92.53	82.03	86.96
+fas.rst.prstc	deu.rst.pcc	fas.rst.prstc	88.94	93.73	91.28
+fas.rst.prstc	eus.rst.ert	fas.rst.prstc	93.05	91.94	92.49
+fas.rst.prstc	fas.rst.prstc	fas.rst.prstc	93.8	90.3	92.02