Contextual word-representations became a standard in modern natural language processing systems. These models use subword tokenization to handle large vocabularies and unknown words. Word-level usage of such systems requires a way of pooling multiple subwords that correspond to a single word. In this paper we investigate how the choice of subword pooling affects the downstream performance on three tasks: morphological probing, POS tagging and NER, in 9 typologically diverse languages. We compare these in two massively multilingual models, mBERT and XLM-RoBERTa. For morphological tasks, the widely used ‘choose the first subword’ is the worst strategy and the best results are obtained by using attention over the subwords. For POS tagging both of these strategies perform poorly and the best choice is to use a small LSTM over the subwords. The same strategy works best for NER and we show that mBERT is better than XLM-RoBERTa in all 9 languages. We publicly release all code, data and the full result tables at https://github.com/juditacs/subword-choice .

This year's iteration of the SIGMORPHON Shared Task on morphological reinflection focuses on typological diversity and cross-lingual variation of morphosyntactic features. In terms of the task, we enrich UniMorph with new data for 32 languages from 13 language families, with most of them being under-resourced: Kunwinjku, Classical Syriac, Arabic (Modern Standard, Egyptian, Gulf), Hebrew, Amharic, Aymara, Magahi, Braj, Kurdish (Central, Northern, Southern), Polish, Karelian, Livvi, Ludic, Veps, Võro, Evenki, Xibe, Tuvan, Sakha, Turkish, Indonesian, Kodi, Seneca, Asháninka, Yanesha, Chukchi, Itelmen, Eibela. We evaluate six systems on the new data and conduct an extensive error analysis of the systems' predictions. Transformer-based models generally demonstrate superior performance on the majority of languages, achieving >90% accuracy on 65% of them. The languages on which systems yielded low accuracy are mainly under-resourced, with a limited amount of data. Most errors made by the systems are due to allomorphy, honorificity, and form variation. In addition, we observe that systems especially struggle to inflect multiword lemmas. The systems also produce misspelled forms or end up in repetitive loops (e.g., RNN-based models). Finally, we report a large drop in systems' performance on previously unseen lemmas.

We present the BME submission for the SIGMORPHON 2021 Task 0 Part 1, Generalization Across Typologically Diverse Languages shared task. We use an LSTM encoder-decoder model with three step training that is first trained on all languages, then fine-tuned on each language family and finally fine-tuned on individual languages. We use a different type of data augmentation technique in the first two steps. Our system outperformed the only other submission. Although it remains worse than the Transformer baseline released by the organizers, our model is simpler and our data augmentation techniques are easily applicable to new languages. We perform ablation studies and show that the augmentation techniques and the three training steps often help but sometimes have a negative effect. Our code is publicly available.

We present a system for mapping Universal Dependency structures to raw text which learns to restore word order by training an Interpreted Regular Tree Grammar (IRTG) that establishes a mapping between string and graph operations. The reinflection step is handled by a standard sequence-to-sequence architecture with a biLSTM encoder and an LSTM decoder with attention. We modify our 2019 system (Kovács et al., 2019) with a new grammar induction mechanism that allows IRTG rules to operate on lemmata in addition to part-of-speech tags and ensures that each word and its dependents are reordered using the most specific set of learned patterns. We also introduce a hierarchical approach to word order restoration that independently determines the word order of each clause in a sentence before arranging them with respect to the main clause, thereby improving overall readability and also making the IRTG parsing task tractable. We participated in the 2020 Surface Realization Shared task, subtrack T1a (shallow, closed). Human evaluation shows we achieve significant improvements on two of the three out-of-domain datasets compared to the 2019 system we modified. Both components of our system are available on GitHub under an MIT license.

We study a typical intermediary task to Machine Translation, the alignment of NPs in the bitext. After arguing that the task remains relevant even in an end-to-end paradigm, we present simple, dictionary- and word vector-based baselines and a BERT-based system. Our results make clear that even state of the art systems relying on the best end-to-end methods can be improved by bringing in old-fashioned methods such as stopword removal, lemmatization, and dictionaries

The Surface Realization Shared Task involves mapping Universal Dependency graphs to raw text, i.e. restoring word order and inflection from a graph of typed, directed dependencies between lemmas. Interpreted Regular Tree Grammars (IRTGs) encode the correspondence between generations in multiple algebras, and have previously been used for semantic parsing from raw text. Our system induces an IRTG for simultaneously building pairs of surface forms and UD graphs in the SRST training data, then prunes this grammar for each UD graph in the test data for efficient parsing and generation of the surface ordering of lemmas. For the inflection step we use a standard sequence-to-sequence model with a biLSTM encoder and an LSTM decoder with attention. Both components of our system are available on GitHub under an MIT license.

We describe a method for expanding existing dictionaries in several languages by discovering previously non-existent links between translations. We call this method triangulation and we present and compare several variations of it. We assess precision manually, and recall by comparing the extracted dictionaries with independently obtained basic vocabulary sets. We featurize the translation candidates and train a maximum entropy classifier to identify correct translations in the noisy data.