We introduce Stanza, an open-source Python natural language processing toolkit supporting 66 human languages Compared to existing widely used toolkits, Stanza features a language-agnostic fully neural pipeline for text analysis, including tokenization, multi-word token expansion, lemmatization, part-of-speech and morphological feature tagging, dependency parsing, and named entity recognition We have trained Stanza on a total of 112 datasets, including the Universal Dependencies treebanks and other multilingual corpora, and show that the same neural architecture generalizes well and achieves competitive performance on all languages tested Additionally, Stanza includes a native Python interface to the widely used Java Stanford CoreNLP software, which further extends its functionality to cover other tasks such as coreference resolution and relation extraction Source code, documentation, and pretrained models for 66 languages are available at https://stanfordnlpgithubio/stanza/

/pdf/stanza-a-python-natural-language-processing-toolkit-for-many-10jne5sqxl.pdf

Stanza: A Python Natural Language Processing Toolkit for Many Human Languages

In this paper, we show that Multilingual BERT (M-BERT), released by Devlin et al. (2018) as a single language model pre-trained from monolingual corpora in 104 languages, is surprisingly good at zero-shot cross-lingual model transfer, in which task-specific annotations in one language are used to fine-tune the model for evaluation in another language. To understand why, we present a large number of probing experiments, showing that transfer is possible even to languages in different scripts, that transfer works best between typologically similar languages, that monolingual corpora can train models for code-switching, and that the model can find translation pairs. From these results, we can conclude that M-BERT does create multilingual representations, but that these representations exhibit systematic deficiencies affecting certain language pairs.

/pdf/how-multilingual-is-multilingual-bert-2mnhw6bxhb.pdf

How multilingual is Multilingual BERT

We present an update to UDPipe 1.0 (Straka et al., 2016), a trainable pipeline which performs sentence segmentation,
tokenization, POS tagging, lemmatization and dependency parsing.
We provide models for all 50 languages of UD 2.0, and
furthermore, the pipeline can be trained easily using data in CoNLL-U format. For the purpose of the CoNLL 2017 Shared Task: Multilingual Parsing from Raw Text to Universal Dependencies, the updated UDPipe 1.1 was used as one of the baseline systems, finishing as the 13th system of 33 participants. A further improved UDPipe 1.2 participated in the shared task, placing as the 8th best system, while achieving low running times and moder-
ately sized models.
The tool is available under open-source Mozilla Public Licence (MPL) and provides bindings for C++, Python (through
ufal.udpipe PyPI package), Perl (through UFAL::UDPipe CPAN
package), Java and C#.

/pdf/tokenizing-pos-tagging-lemmatizing-and-parsing-ud-2-0-with-4ohfqa4ope.pdf

Tokenizing, POS Tagging, Lemmatizing and Parsing UD 2.0 with UDPipe

This paper describes the neural dependency parser submitted by Stanford to the CoNLL 2017 Shared Task on parsing Universal Dependencies. Our system uses relatively simple LSTM networks to produce part of speech tags and labeled dependency parses from segmented and tokenized sequences of words. In order to address the rare word problem that abounds in languages with complex morphology, we include a character-based word representation that uses an LSTM to produce embeddings from sequences of characters. Our system was ranked first according to all five relevant metrics for the system: UPOS tagging (93.09%), XPOS tagging (82.27%), unlabeled attachment score (81.30%), labeled attachment score (76.30%), and content word labeled attachment score (72.57%).

/pdf/stanford-s-graph-based-neural-dependency-parser-at-the-conll-218j23u97w.pdf

Stanford’s Graph-based Neural Dependency Parser at the CoNLL 2017 Shared Task

The Conference on Computational Natural Language Learning (CoNLL) features a shared task, in which participants train and test their learning systems on the same data sets. In 2017, the task was devoted to learning dependency parsers for a large number of languages, in a real-world setting without any gold-standard annotation on input. All test sets followed a unified annotation scheme, namely that of Universal Dependencies. In this paper, we define the task and evaluation methodology, describe how the data sets were prepared, report and analyze the main results, and provide a brief categorization of the different approaches of the participating systems.

/pdf/conll-2017-shared-task-multilingual-parsing-from-raw-text-to-45vgerk8my.pdf

CoNLL 2017 Shared Task: Multilingual Parsing from Raw Text to Universal Dependencies

There has been substantial recent interest in annotation schemes that can be applied consistently to many languages. Building on several recent efforts to unify morphological and syntactic annotation, the Universal Dependencies (UD) project seeks to introduce a cross-linguistically applicable part-of-speech tagset, feature inventory, and set of dependency relations as well as a large number of uniformly annotated treebanks. We present Universal Dependencies for Finnish, one of the ten languages in the recent first release of UD project treebank data. We detail the mapping of previously introduced annotation to the UD standard, describing specific challenges and their resolution. We additionally present parsing experiments comparing the performance of a stateof-the-art parser trained on a languagespecific annotation schema to performance on the corresponding UD annotation. The results show improvement compared to the source annotation, indicating that the conversion is accurate and supporting the feasibility of UD as a parsing target. The introduced tools and resources are available under open licenses from http://bionlp.utu.fi/ud-finnish.html.

https://www.ep.liu.se/ecp/109/021/ecp15109021.pdf

Universal Dependencies for Finnish

This paper presents the first results on detecting informality, machine and human translations in the Finnish Internet Parsebank, a project developing a large-scale, web-based corpus with full morphological and syntactic analyses. The paper aims at classifying the Parsebank according to these criteria, as well as studying the linguistic characteristics of the classes. The features used include both lexical and morpho-syntactic properties, such as syntactic n-grams. The results are practically applicable, with an AUC range of 85‐85% for the human, 98% for the machine translated texts and 73% for the informal texts. While word-based classification performs well for the indomain experiments, delexicalized methods with morpho-syntactic features prove to be more tolerant to variation caused by genre or source language. In addition, the results show that the features used in the classification provide interesting pointers for further, more detailed studies on the linguistic characteristics of these texts.

/pdf/towards-the-classification-of-the-finnish-internet-parsebank-yz4d7elcyn.pdf

Anna Missilä

Papers

CoNLL 2017 Shared Task: Multilingual Parsing from Raw Text to Universal Dependencies

Universal Dependencies for Finnish

Towards the Classification of the Finnish Internet Parsebank: Detecting Translations and Informality