xlmr-multilingual-sentence-segmentation

This model is a fine-tuned version of xlm-roberta-base on a corrupted version of the universal dependency datasets. It achieves the following results on the (also corrupted) evaluation set:

• Loss: 0.0074
• Precision: 0.9664
• Recall: 0.9677
• F1: 0.9670

Test set performance

Results

All results here are percentage F1:

Opus100 [2]

Who wins most? XLM-RoBERTa: 56, WtPSplit: 12, Spacy (multilingual): 8

	af	am	ar	az	be	bg	bn	ca	cs	cy	da	de	el	en	eo	es	et	eu	fa	fi	fr	fy	ga	gd	gl	gu	ha	he	hi	hu	hy	id	is	it	ja	ka	kk	km	kn	ko	ku	ky	lt	lv	mg	mk	ml	mn	mr	ms	my	ne	nl	pa	pl	ps	pt	ro	ru	si	sk	sl	sq	sr	sv	ta	te	th	tr	uk	ur	uz	vi	xh	yi	zh
Spacy (multilingual)	42.61	6.69	58.52	73.59	34.78	93.74	38.04	88.76	87.70	26.30	90.52	74.15	89.75	89.25	88.77	90.95	87.26	81.20	55.40	93.28	85.77	21.49	60.61	36.83	88.77	5.59	89.39	92.21	53.33	93.26	24.14	90.13	95.38	86.32	0.20	38.24	42.39	0.10	9.66	51.79	27.64	21.77	76.91	77.02	83.60	93.74	39.09	33.23	86.56	87.39	0.10	6.59	93.65	5.26	92.42	2.41	92.07	91.63	75.95	75.91	92.13	93.00	92.96	95.01	93.52	36.97	64.59	21.64	94.05	89.68	29.17	64.99	90.59	64.89	4.14	0.09
WtPSplit	76.90	59.08	68.08	76.42	71.29	93.97	79.76	89.79	89.36	73.21	90.02	80.74	92.80	91.91	92.24	92.11	84.47	87.24	59.97	91.96	88.53	65.84	79.49	83.33	90.31	70.51	82.43	90.58	66.70	93.00	87.14	89.80	94.77	87.43	41.79	91.26	73.25	69.54	68.98	56.21	79.12	83.94	81.33	82.70	89.33	92.87	80.81	73.26	89.20	88.51	65.54	71.33	92.63	64.11	92.72	62.84	91.05	90.91	84.23	80.32	92.30	92.19	90.32	94.76	92.08	63.48	76.49	68.88	93.30	89.60	52.59	77.79	91.29	80.28	75.70	71.64
XLM-RoBERTa (ours)	83.97	41.59	81.56	81.30	85.68	94.34	84.10	91.80	91.23	78.72	92.64	86.73	93.87	94.50	94.57	93.18	90.19	90.28	74.79	94.06	90.46	81.76	84.33	85.62	92.55	67.26	86.61	91.22	72.69	94.53	89.83	92.24	93.78	89.27	41.43	78.39	89.15	36.60	70.51	82.77	58.14	89.41	89.99	88.25	86.82	92.81	86.14	94.73	93.25	92.44	49.39	66.02	93.60	69.22	93.51	61.86	92.84	93.19	89.47	86.24	92.95	93.46	91.79	94.16	93.93	72.74	81.77	74.49	93.17	92.15	62.92	75.65	93.41	84.89	56.85	77.07

Universal Dependencies [3]

Who wins most? XLM-RoBERTa: 24, WtPSplit: 17 Spacy (multilingual): 13

	af	ar	be	bg	bn	ca	cs	cy	da	de	el	en	es	et	eu	fa	fi	fr	ga	gd	gl	he	hi	hu	hy	id	is	it	ja	jv	kk	ko	la	lt	lv	mr	nl	pl	pt	ro	ru	sk	sl	sq	sr	sv	ta	th	tr	uk	ur	vi	zh
Spacy (multilingual)	98.47	80.38	80.27	93.62	51.85	98.95	89.68	98.89	94.96	88.02	94.16	92.20	98.70	93.77	95.79	99.83	92.88	96.33	96.67	63.04	92.37	94.37	0.32	98.45	11.39	98.01	95.41	92.49	0.37	98.03	96.21	99.80	0.09	93.86	98.52	92.13	92.86	97.02	94.91	98.05	84.31	90.26	98.23	100.00	97.84	94.91	66.67	1.95	97.63	94.16	0.37	96.40	0.40
WtPSplit	98.27	83.00	89.28	98.16	99.12	98.52	92.98	99.26	94.56	96.13	96.94	94.73	97.60	94.09	97.24	97.29	94.69	96.71	86.60	72.17	98.87	95.79	96.78	96.08	96.80	98.41	86.39	95.45	95.84	98.18	96.28	99.11	91.43	97.67	96.42	91.84	93.61	95.92	96.13	81.50	86.28	95.57	96.85	99.17	98.45	95.86	97.54	70.26	96.00	92.08	93.79	92.97	97.25
XLM-RoBERTa (ours)	96.81	78.99	91.60	97.89	99.12	95.99	96.05	97.17	96.62	96.29	94.33	94.76	95.73	96.20	97.37	97.49	96.34	95.70	89.78	84.20	95.72	95.95	97.51	96.24	95.62	97.22	92.93	96.88	94.23	96.29	98.40	97.46	96.35	95.82	96.91	95.92	96.27	97.24	95.83	94.63	91.59	95.88	96.43	98.36	96.83	94.95	95.93	89.26	96.52	94.59	96.20	97.31	95.12

Ersatz [4]

Who wins most? XLM-RoBERTa: 10, WtPSplit: 8, Spacy (multilingual): 4

	ar	cs	de	en	es	et	fi	fr	gu	hi	ja	kk	km	lt	lv	pl	ps	ro	ru	ta	tr	zh
Spacy (multilingual)	91.26	96.46	93.89	94.40	97.31	97.15	94.99	96.43	4.44	18.41	0.18	97.11	0.08	93.53	98.73	93.69	94.44	94.87	93.45	68.65	95.39	0.10
WtPSplit	89.45	93.41	95.93	97.16	98.74	95.84	97.10	97.61	90.62	94.87	82.14	95.94	82.89	96.74	97.22	95.16	86.99	97.55	97.82	94.76	93.53	89.02
XLM-RoBERTa (ours)	79.78	96.94	97.02	96.10	97.06	96.80	97.67	96.33	93.73	95.34	77.54	97.28	78.94	96.13	96.45	96.71	92.33	96.24	97.15	95.94	95.76	90.11

German--English code-switching [5]

	de
Spacy (multilingual)	79.55
WtPSplit	77.41
XLM-RoBERTa (ours)	85.78

[1] Where’s the Point? Self-Supervised Multilingual Punctuation-Agnostic Sentence Segmentation (Minixhofer et al., ACL 2023)

[2] Improving Massively Multilingual Neural Machine Translation and Zero-Shot Translation (Zhang et al., ACL 2020)

[3] Universal Dependencies (de Marneffe et al., CL 2021)

[4] A unified approach to sentence segmentation of punctuated text in many languages (Wicks & Post, ACL-IJCNLP 2021)

[5] The Denglisch Corpus of German-English Code-Switching (Osmelak & Wintner, SIGTYP 2023)

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 64
• eval_batch_size: 64
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 5

Training results

Training Loss	Epoch	Step	Validation Loss	Precision	Recall	F1
No log	0.2	100	0.0125	0.9320	0.9487	0.9403
No log	0.4	200	0.0099	0.9547	0.9513	0.9530
No log	0.6	300	0.0092	0.9616	0.9506	0.9561
No log	0.81	400	0.0083	0.9584	0.9618	0.9601
0.0212	1.01	500	0.0082	0.9551	0.9642	0.9596
0.0212	1.21	600	0.0084	0.9630	0.9614	0.9622
0.0212	1.41	700	0.0079	0.9606	0.9648	0.9627
0.0212	1.61	800	0.0077	0.9609	0.9661	0.9635
0.0212	1.81	900	0.0076	0.9623	0.9649	0.9636
0.0067	2.02	1000	0.0077	0.9598	0.9689	0.9643
0.0067	2.22	1100	0.0075	0.9614	0.9680	0.9647
0.0067	2.42	1200	0.0073	0.9626	0.9682	0.9654
0.0067	2.62	1300	0.0075	0.9617	0.9692	0.9654
0.0067	2.82	1400	0.0073	0.9658	0.9648	0.9653
0.0054	3.02	1500	0.0076	0.9656	0.9663	0.9660
0.0054	3.23	1600	0.0073	0.9625	0.9703	0.9664
0.0054	3.43	1700	0.0073	0.9658	0.9659	0.9658
0.0054	3.63	1800	0.0073	0.9626	0.9707	0.9666
0.0054	3.83	1900	0.0073	0.9659	0.9677	0.9668
0.0046	4.03	2000	0.0075	0.9671	0.9659	0.9665
0.0046	4.23	2100	0.0075	0.9654	0.9687	0.9671
0.0046	4.44	2200	0.0075	0.9662	0.9676	0.9669
0.0046	4.64	2300	0.0074	0.9657	0.9684	0.9670
0.0046	4.84	2400	0.0074	0.9664	0.9678	0.9671

Framework versions

• Transformers 4.39.1
• Pytorch 2.2.1+cu121
• Datasets 2.18.0
• Tokenizers 0.15.2