The Helsinki-NLP team participated in the AmericasNLP 2023 Shared Task with 6 submissions for all 11 language pairs arising from 4 different multilingual systems. We provide a detailed look at the work that went into collecting and preprocessing the data that led to our submissions. We explore various setups for multilingual Neural Machine Translation (NMT), namely knowledge distillation and transfer learning, multilingual NMT including a high-resource language (English), language-specific fine-tuning, and multilingual NMT exclusively using low-resource data. Our multilingual Model B ranks first in 4 out of the 11 language pairs.
This work presents an unsupervised method of selecting filters and threshold values for the OpusFilter parallel corpus cleaning toolbox. The method clusters sentence pairs into noisy and clean categories and uses the features of the noisy cluster center as filtering parameters. Our approach utilizes feature importance analysis to disregard filters that do not differentiate between clean and noisy data. A randomly sampled subset of a given corpus is used for filter selection and ineffective filters are not run for the full corpus. We use a set of automatic evaluation metrics to assess the quality of translation models trained with data filtered by our method and data filtered with OpusFilter’s default parameters. The trained models cover English-German and English-Ukrainian in both directions. The proposed method outperforms the default parameters in all translation directions for almost all evaluation metrics.
This paper introduces Bayesian uncertainty modeling using Stochastic Weight Averaging-Gaussian (SWAG) in Natural Language Understanding (NLU) tasks. We apply the approach to standard tasks in natural language inference (NLI) and demonstrate the effectiveness of the method in terms of prediction accuracy and correlation with human annotation disagreements. We argue that the uncertainty representations in SWAG better reflect subjective interpretation and the natural variation that is also present in human language understanding. The results reveal the importance of uncertainty modeling, an often neglected aspect of neural language modeling, in NLU tasks.
In our submission to the SIGMORPHON 2022 Shared Task on Morpheme Segmentation, we study whether an unsupervised morphological segmentation method, Morfessor, can help in a supervised setting. Previous research has shown the effectiveness of the approach in semisupervised settings with small amounts of labeled data. The current tasks vary in data size: the amount of word-level annotated training data is much larger, but the amount of sentencelevel annotated training data remains small. Our approach is to pre-segment the input data for a neural sequence-to-sequence model with the unsupervised method. As the unsupervised method can be trained with raw text data, we use Wikipedia to increase the amount of training data. In addition, we train multilingual models for the sentence-level task. The results for the Morfessor-enriched features are mixed, showing benefit for all three sentencelevel tasks but only some of the word-level tasks. The multilingual training yields considerable improvements over the monolingual sentence-level models, but it negates the effect of the enriched features.
Cross-language forced alignment is a solution for linguists who create speech corpora for very low-resource languages. However, cross-language is an additional challenge making a complex task, forced alignment, even more difficult. We study how linguists can impart domain expertise to the tasks to increase the performance of automatic forced aligners while keeping the time effort still lower than with manual forced alignment. First, we show that speech recognizers have a clear bias in starting the word later than a human annotator, which results in micro-pauses in the results that do not exist in manual alignments, and study which is the best way to automatically remove these silences. Second, we ask the linguists to simplify the task by splitting long interview audios into shorter lengths by providing some manually aligned segments and evaluating the results of this process. We also study how correlated source language performance is to target language performance, since often it is an easier task to find a better source model than to adapt to the target language.
The University of Helsinki participated in the AmericasNLP shared task for all ten language pairs. Our multilingual NMT models reached the first rank on all language pairs in track 1, and first rank on nine out of ten language pairs in track 2. We focused our efforts on three aspects: (1) the collection of additional data from various sources such as Bibles and political constitutions, (2) the cleaning and filtering of training data with the OpusFilter toolkit, and (3) different multilingual training techniques enabled by the latest version of the OpenNMT-py toolkit to make the most efficient use of the scarce data. This paper describes our efforts in detail.
Forced alignment is an effective process to speed up linguistic research. However, most forced aligners are language-dependent, and under-resourced languages rarely have enough resources to train an acoustic model for an aligner. We present a new Finnish grapheme-based forced aligner and demonstrate its performance by aligning multiple Uralic languages and English as an unrelated language. We show that even a simple non-expert created grapheme-to-phoneme mapping can result in useful word alignments.
We consider a low-resource translation task from Finnish into Northern Sámi. Collecting all available parallel data between the languages, we obtain around 30,000 sentence pairs. However, there exists a significantly larger monolingual Northern Sámi corpus, as well as a rule-based machine translation (RBMT) system between the languages. To make the best use of the monolingual data in a neural machine translation (NMT) system, we use the backtranslation approach to create synthetic parallel data from it using both NMT and RBMT systems. Evaluating the results on an in-domain test set and a small out-of-domain set, we find that the RBMT backtranslation outperforms NMT backtranslation clearly for the out-of-domain test set, but also slightly for the in-domain data, for which the NMT backtranslation model provided clearly better BLEU scores than the RBMT. In addition, combining both backtranslated data sets improves the RBMT approach only for the in-domain test set. This suggests that the RBMT system provides general-domain knowledge that cannot be found from the relative small parallel training data.
Character-based Neural Network Language Models (NNLM) have the advantage of smaller vocabulary and thus faster training times in comparison to NNLMs based on multi-character units. However, in low-resource scenarios, both the character and multi-character NNLMs suffer from data sparsity. In such scenarios, cross-lingual transfer has improved multi-character NNLM performance by allowing information transfer from a source to the target language. In the same vein, we propose to use cross-lingual transfer for character NNLMs applied to low-resource Automatic Speech Recognition (ASR). However, applying cross-lingual transfer to character NNLMs is not as straightforward. We observe that relatedness of the source language plays an important role in cross-lingual pretraining of character NNLMs. We evaluate this aspect on ASR tasks for two target languages: Finnish (with English and Estonian as source) and Swedish (with Danish, Norwegian, and English as source). Prior work has observed no difference between using the related or unrelated language for multi-character NNLMs. We, however, show that for character-based NNLMs, only pretraining with a related language improves the ASR performance, and using an unrelated language may deteriorate it. We also observe that the benefits are larger when there is much lesser target data than source data.
Contextualized word representations encode rich information about syntax and semantics, alongside specificities of each context of use. While contextual variation does not always reflect actual meaning shifts, it can still reduce the similarity of embeddings for word instances having the same meaning. We explore the imprint of two specific linguistic alternations, namely passivization and negation, on the representations generated by neural models trained with two different objectives: masked language modeling and translation. Our exploration methodology is inspired by an approach previously proposed for removing societal biases from word vectors. We show that passivization and negation leave their traces on the representations, and that neutralizing this information leads to more similar embeddings for words that should preserve their meaning in the transformation. We also find clear differences in how the respective features generalize across datasets.
This paper introduces OpusTools, a package for downloading and processing parallel corpora included in the OPUS corpus collection. The package implements tools for accessing compressed data in their archived release format and make it possible to easily convert between common formats. OpusTools also includes tools for language identification and data filtering as well as tools for importing data from various sources into the OPUS format. We show the use of these tools in parallel corpus creation and data diagnostics. The latter is especially useful for the identification of potential problems and errors in the extensive data set. Using these tools, we can now monitor the validity of data sets and improve the overall quality and consitency of the data collection.
Data-driven segmentation of words into subword units has been used in various natural language processing applications such as automatic speech recognition and statistical machine translation for almost 20 years. Recently it has became more widely adopted, as models based on deep neural networks often benefit from subword units even for morphologically simpler languages. In this paper, we discuss and compare training algorithms for a unigram subword model, based on the Expectation Maximization algorithm and lexicon pruning. Using English, Finnish, North Sami, and Turkish data sets, we show that this approach is able to find better solutions to the optimization problem defined by the Morfessor Baseline model than its original recursive training algorithm. The improved optimization also leads to higher morphological segmentation accuracy when compared to a linguistic gold standard. We publish implementations of the new algorithms in the widely-used Morfessor software package.
This paper describes the joint participation of University of Helsinki and Aalto University to two shared tasks of WMT 2020: the news translation between Inuktitut and English and the low-resource translation between German and Upper Sorbian. For both tasks, our efforts concentrate on efficient use of monolingual and related bilingual corpora with scheduled multi-task learning as well as an optimized subword segmentation with sampling. Our submission obtained the highest score for Upper Sorbian -> German and was ranked second for German -> Upper Sorbian according to BLEU scores. For English–Inuktitut, we reached ranks 8 and 10 out of 11 according to BLEU scores.
This paper introduces OpusFilter, a flexible and modular toolbox for filtering parallel corpora. It implements a number of components based on heuristic filters, language identification libraries, character-based language models, and word alignment tools, and it can easily be extended with custom filters. Bitext segments can be ranked according to their quality or domain match using single features or a logistic regression model that can be trained without manually labeled training data. We demonstrate the effectiveness of OpusFilter on the example of a Finnish-English news translation task based on noisy web-crawled training data. Applying our tool leads to improved translation quality while significantly reducing the size of the training data, also clearly outperforming an alternative ranking given in the crawled data set. Furthermore, we show the ability of OpusFilter to perform data selection for domain adaptation.
In this paper we present the University of Helsinki submissions to the WMT 2019 shared news translation task in three language pairs: English-German, English-Finnish and Finnish-English. This year we focused first on cleaning and filtering the training data using multiple data-filtering approaches, resulting in much smaller and cleaner training sets. For English-German we trained both sentence-level transformer models as well as compared different document-level translation approaches. For Finnish-English and English-Finnish we focused on different segmentation approaches and we also included a rule-based system for English-Finnish.
This paper describes the University of Helsinki Language Technology group’s participation in the WMT 2019 similar language translation task. We trained neural machine translation models for the language pairs Czech <-> Polish and Spanish <-> Portuguese. Our experiments focused on different subword segmentation methods, and in particular on the comparison of a cognate-aware segmentation method, Cognate Morfessor, with character segmentation and unsupervised segmentation methods for which the data from different languages were simply concatenated. We did not observe major benefits from cognate-aware segmentation methods, but further research may be needed to explore larger parts of the parameter space. Character-level models proved to be competitive for translation between Spanish and Portuguese, but they are slower in training and decoding.
This article describes the Aalto University entry to the WMT18 News Translation Shared Task. We participate in the multilingual subtrack with a system trained under the constrained condition to translate from English to both Finnish and Estonian. The system is based on the Transformer model. We focus on improving the consistency of morphological segmentation for words that are similar orthographically, semantically, and distributionally; such words include etymological cognates, loan words, and proper names. For this, we introduce Cognate Morfessor, a multilingual variant of the Morfessor method. We show that our approach improves the translation quality particularly for Estonian, which has less resources for training the translation model.
There are many accurate methods for language identification of long text samples, but identification of very short strings still presents a challenge. This paper studies a language identification task, in which the test samples have only 5-21 characters. We compare two distinct methods that are well suited for this task: a naive Bayes classifier based on character n-gram models, and the ranking method by Cavnar and Trenkle (1994). For the n-gram models, we test several standard smoothing techniques, including the current state-of-the-art, the modified Kneser-Ney interpolation. Experiments are conducted with 281 languages using the Universal Declaration of Human Rights. Advanced language model smoothing techniques improve the identification accuracy and the respective classifiers outperform the ranking method. The higher accuracy is obtained at the cost of larger models and slower classification speed. However, there are several methods to reduce the size of an n-gram model, and our experiments with model pruning show that it provides an easy way to balance the size and the identification accuracy. We also compare the results to the language identifier in Google AJAX Language API, using a subset of 50 languages.