Fig 1.
An example of neural NER with sequence labeling.
Fig 2.
The internal structure of a one-layer transformer.
(a) Standard Transformer. (b) Transformer with Adapters.
Fig 3.
The pretrained transformer-based language model coupling with adapters and PGN.
Table 1.
Data statistics, where the number of sentences and entities are reported, and Devel indicates the development set.
Table 2.
Main results of single-source cross-lingual NER, where lavg indicates the averaged performance for each target language, and avg denotes the overall average F-scores of all source-target pairs.
Table 3.
Main results of multi-source cross-lingual NER, where all other languages except the target language itself are exploited as the source languages.
Table 4.
Comparisons with previous studies.
Fig 4.
An example of word alignment visualization between a German sentence and its English translation, where the solid arrows are gold-standard being all correctly predicted by XLM, and the dashed arrows are incorrectly aligned by mBERT, and the others are the same for the two types of word representations.
Table 5.
The comparisons between the fine-tuning and feature-based adapter exploration methods, where XLM is used as the input language model.
Fig 5.
The similarity heatmap of language embeddings for different language pairs, where deeper color indicates higher similarity.
Table 6.
A case study, where the text with underlines indicates errors.