Documenting languages helps to prevent the extinction of endangered dialects - many of which are otherwise expected to disappear by the end of the century. When documenting oral languages, unsupervised word segmentation (UWS) from speech is a useful, yet challenging, task. It consists in producing time-stamps for slicing utterances into smaller segments corresponding to words, being performed from phonetic transcriptions, or in the absence of these, from the output of unsupervised speech discretization models. These discretization models are trained using raw speech only, producing discrete speech units that can be applied for downstream (text-based) tasks. In this paper we compare five of these models: three Bayesian and two neural approaches, with regards to the exploitability of the produced units for UWS. For the UWS task, we experiment with two models, using as our target language the Mboshi (Bantu C25), an unwritten language from Congo-Brazzaville. Additionally, we report results for Finnish, Hungarian, Romanian and Russian in equally low-resource settings, using only 4 hours of speech. Our results suggest that neural models for speech discretization are difficult to exploit in our setting, and that it might be necessary to adapt them to limit sequence length. We obtain our best UWS results by using Bayesian models that produce high quality, yet compressed, discrete representations of the input speech signal.

We have developed an open source web reader in Iceland for under-resourced languages. The web reader was developed due to the need for a free and good quality web reader for languages which fall outside the scope of commercially available web readers. It relies on a text-to-speech (TTS) pipeline accessed via a cloud service. The web reader was developed using the Icelandic TTS voices Alfur and Dilja, but could be connected to any language which has a TTS pipeline. The design of our web reader focuses on functionality, adaptability and user friendliness. Therefore, the web reader’s feature set heavily overlaps with the minimal features necessary to provide a good web reading experience while still being extensible enough to be adapted to work for other languages, high-resourced and under-resourced. The web reader works well on all the major web browsers and has a Web Content Accessibility Guidelines 2.0 Level AA: Acceptable compliance, meaning that it works well for the largest user groups, people in under-resourced languages with visual impairments and difficulty reading. The code for our web reader is available and published with an Apache 2.0 license at https://github.com/cadia-lvl/WebRICE, which includes a simple demo of the project.

We propose a new approach for phoneme mapping in cross-lingual transfer learning for text-to-speech (TTS) in under-resourced languages (URLs), using phonological features from the PHOIBLE database and a language-independent mapping rule. This approach was validated through our experiment, in which we pre-trained acoustic models in Dutch, Finnish, French, Japanese, and Spanish, and fine-tuned them with 30 minutes of Frisian training data. The experiment showed an improvement in both naturalness and pronunciation accuracy in the synthesized Frisian speech when our mapping approach was used. Since this improvement also depended on the source language, we then experimented on finding a good criterion for selecting source languages. As an alternative to the traditionally used language family criterion, we tested a novel idea of using Angular Similarity of Phoneme Frequencies (ASPF), which measures the similarity between the phoneme systems of two languages. ASPF was empirically confirmed to be more effective than language family as a criterion for source language selection, and also to affect the phoneme mapping’s effectiveness. Thus, a combination of our phoneme mapping approach and the ASPF measure can be beneficially adopted by other studies involving multilingual or cross-lingual TTS for URLs.

While the alignment of audio recordings and text (often termed “forced alignment”) is often treated as a solved problem, in practice the process of adapting an alignment system to a new, under-resourced language comes with significant challenges, requiring experience and expertise that many outside of the speech community lack. This puts otherwise “solvable” problems, like the alignment of Indigenous language audiobooks, out of reach for many real-world Indigenous language organizations. In this paper, we detail ReadAlong Studio, a suite of tools for creating and visualizing aligned audiobooks, including educational features like time-aligned highlighting, playing single words in isolation, and variable-speed playback. It is intended to be accessible to creators without an extensive background in speech or NLP, by automating or making optional many of the specialist steps in an alignment pipeline. It is well documented at a beginner-technologist level, has already been adapted to 30 languages, and can work out-of-the-box on many more languages without adaptation.

Sentiment Analysis (SA) employing code-mixed data from social media helps in getting insights to the data and decision making for various applications. One such application is to analyze users’ emotions from comments of videos on YouTube. Social media comments do not adhere to the grammatical norms of any language and they often comprise a mix of languages and scripts. The lack of annotated code-mixed data for SA in a low-resource language like Tulu makes the SA a challenging task. To address the lack of annotated code-mixed Tulu data for SA, a gold standard trlingual code-mixed Tulu annotated corpus of 7,171 YouTube comments is created. Further, Machine Learning (ML) algorithms are employed as baseline models to evaluate the developed dataset and the performance of the ML algorithms are found to be encouraging.

Oral corpora for linguistic inquiry are frequently built based on the content of news, radio, and/or TV shows, sometimes also of laboratory recordings. Most of these existing corpora are restricted to languages with a large amount of data available. Furthermore, such corpora are not always accessible under a free open-access license. We propose a crowd-sourced alternative to this gap. Lingua Libre is the participatory linguistic media library hosted by Wikimedia France. It includes recordings from more than 140 languages. These recordings have been provided by more than 750 speakers worldwide, who voluntarily recorded word entries of their native language and made them available under a Creative Commons license. In the present study, we take Polish, a less-resourced language in terms of phonetic data, as an example, and compare our phonetic observations built on the data from Lingua Libre with the phonetic observations found by previous linguistic studies. We observe that the data from Lingua Libre partially matches the phonetic inventory of Polish as described in previous studies, but that the acoustic values are less precise, thus showing both the potential and the limitations of Lingua Libre to be used for phonetic research.

TuLaR (Tupian Language Resources) is a project for collecting, documenting, analyzing, and developing computational and pedagogical material for low-resource Brazilian indigenous languages. It provides valuable data for language research regarding typological, syntactic, morphological, and phonological aspects. Here we present TuLaR’s databases, with special consideration to TuDeT (Tupian Dependency Treebanks), an annotated corpus under development for nine languages of the Tupian family, built upon the Universal Dependencies framework. The annotation within such a framework serves a twofold goal: enriching the linguistic documentation of the Tupian languages due to the rapid and consistent annotation, and providing computational resources for those languages, thanks to the suitability of our framework for developing NLP tools. We likewise present a related lexical database, some tools developed by the project, and examine future goals for our initiative.

In this work, we make the case of quality over quantity when training a MT system for a medium-to-low-resource language pair, namely Catalan-English. We compile our training corpus out of existing resources of varying quality and a new high-quality corpus. We also provide new evaluation translation datasets in three different domains. In the process of building Catalan-English parallel resources, we evaluate the impact of drastically filtering alignments in the resulting MT engines. Our results show that even when resources are limited, as in this case, it is worth filtering for quality. We further explore the cross-lingual transfer learning capabilities of the proposed model for parallel corpus filtering by applying it to other languages. All resources generated in this work are released under open license to encourage the development of language technology in Catalan.

Multilingual sentiment analysis is a process of detecting and classifying sentiment based on textual information written in multiple languages. There has been tremendous research advancement on high-resourced languages such as English. However, progress on under-resourced languages remains underrepresented with limited opportunities for further development of natural language processing (NLP) technologies. Sentiment analysis (SA) for under-resourced language still is a skewed research area. Although, there are some considerable efforts in emerging African countries to develop such resources for under-resourced languages, languages such as indigenous South African languages still suffer from a lack of datasets. To the best of our knowledge, there is currently no dataset dedicated to SA research for South African languages in a multilingual context, i.e. comments are in different languages and may contain code-switching. In this paper, we present the first subset of the multilingual sentiment corpus SAfriSenti for the three most widely spoken languages in South Africa—English, Sepedi (i.e. Northern Sotho), and Setswana. This subset consists of over 40,000 annotated tweets in all the three languages including even 36.6% of code-switched texts. We present data collection, cleaning and annotation strategies that were followed to curate the dataset for these languages. Furthermore, we describe how we developed language-specific sentiment lexicons, morpheme-based sentiment taggers, conduct linguistic analyses and present possible solutions for the challenges of this sentiment dataset. We will release the dataset and sentiment lexicons to the research communities to advance the NLP research of under-resourced languages.

This paper describes our submission to the MT4All Shared Task in unsupervised machine translation from English to Ukrainian, Kazakh and Georgian in the legal domain. In addition to the standard pipeline for unsupervised training (pretraining followed by denoising and back-translation), we used supervised training on a pseudo-parallel corpus retrieved from the provided mono-lingual corpora. Our system scored significantly higher than the baseline hybrid unsupervised MT system.

Reliable and maintained indicators of the space of languages on the Internet are required to support appropriate public policies and well-informed linguistic studies. Current sources are scarce and often strongly biased. The model to produce indicators on the presence of languages in the Internet, launched by the Observatory in 2017, has reached a sensible level of maturity and its data products are shared in CC-BY-SA 4.0 license. It reaches now 329 languages (L1 speakers > one million) and all the biases associated with the model have been controlled to an acceptable threshold, giving trust to the data, within an estimated confidence interval of +-20%. Some of the indicators (mainly the percentage of L1+L2 speakers connected to the Internet per language and derivates) rely on Ethnologue Global Dataset #24 for demo-linguistic data and ITU, completed by World Bank, for the percentage of persons connected to the Internet by country. The rest of indicators relies on the previous sources plus a large combination of hundreds of different sources for data related to Web contents per language. This research poster focuses the description of the new linguistic resources created. Methodological considerations are only exposed briefly and will be developed in another paper.

There is a growing interest in building language technologies (LTs) for low resource languages (LRLs). However, there are flaws in the planning, data collection and development phases mostly due to the assumption that LRLs are similar to High Resource Languages (HRLs) but only smaller in size. In our paper, we first provide examples of failed LTs for LRLs and provide the reasons for these failures. Second, we discuss the problematic issues with the data for LRLs. Finally, we provide recommendations for building better LTs for LRLs through insights from sociolinguistics and multilingualism. Our goal is not to solve all problems around LTs for LRLs but to raise awareness about the existing issues, provide recommendations toward possible solutions and encourage collaboration across academic disciplines for developing LTs that actually serve the needs and preferences of the LRL communities.

We describe our work on sentiment analysis for Hausa, where we investigated monolingual and cross-lingual approaches to classify student comments in course evaluations. Furthermore, we propose a novel stemming algorithm to improve accuracy. For studies in this area, we collected a corpus of more than 40,000 comments—the Hausa-English Sentiment Analysis Corpus For Educational Environments (HESAC). Our results demonstrate that the monolingual approaches for Hausa sentiment analysis slightly outperform the cross-lingual systems. Using our stemming algorithm in the pre-processing even improved the best model resulting in 97.4% accuracy on HESAC.

Language model pre-training has significantly impacted NLP and resulted in performance gains on many NLP-related tasks, but comparative study of different approaches on many low-resource languages seems to be missing. This paper attempts to investigate appropriate methods for pretraining a Transformer-based model for the Nepali language. We focus on the language-specific aspects that need to be considered for modeling. Although some language models have been trained for Nepali, the study is far from sufficient. We train three distinct Transformer-based masked language models for Nepali text sequences: distilbert-base (Sanh et al., 2019) for its efficiency and minuteness, deberta-base (P. He et al., 2020) for its capability of modeling the dependency of nearby token pairs and XLM-ROBERTa (Conneau et al., 2020) for its capabilities to handle multilingual downstream tasks. We evaluate and compare these models with other Transformer-based models on a downstream classification task with an aim to suggest an effective strategy for training low-resource language models and their fine-tuning.

We propose a method for identifying monolingual textual segments in multilingual documents. It requires only a minimal number of linguistic resources – word lists and monolingual corpora – and can therefore be adapted to many under-resourced languages. Taking these languages into account when processing multilingual documents in NLP tools is important as it can contribute to the creation of essential textual resources. This language identification task – code switching detection being its most complex form – can also provide added value to various existing data or tools. Our research demonstrates that a language identification module performing well on short texts can be used to efficiently analyse a document through a sliding window. The results obtained for code switching identification – between 87.29% and 97.97% accuracy – are state-of-the-art, which is confirmed by the benchmarks performed on the few available systems that have been used on our test data.

Indonesia has many varieties of ethnic languages, and most come from the same language family, namely Austronesian languages. Coming from that same language family, the words in Indonesian ethnic languages are very similar. However, there is research stating that Indonesian ethnic languages are endangered. Thus, to prevent that, we proposed to create a bilingual dictionary between ethnic languages using a neural network approach to extract transformation rules using character level embedding and the Bi-LSTM method in a sequence-to-sequence model. The model has an encoder and decoder. The encoder functions read the input sequence, character by character, generate context, then extract a summary of the input. The decoder will produce an output sequence where every character in each time-step and the next character that comes out are affected by the previous character. The current case for experiment translation focuses on Minangkabau and Indonesian languages with 13761-word pairs. For evaluating the model’s performance, 5-Fold Cross-Validation is used.

Machine translation has been researched using deep neural networks in recent years. These networks require lots of data to learn abstract representations of the input stored in continuous vectors. Dialect translation has become more important since the advent of social media. In particular, when dialect speakers and standard language speakers no longer understand each other, machine translation is of rising concern. Usually, dialect translation is a typical low-resourced language setting facing data scarcity problems. Additionally, spelling inconsistencies due to varying pronunciations and the lack of spelling rules complicate translation. This paper presents the best-performing approaches to handle these problems for Alemannic dialects. The results show that back-translation and conditioning on dialectal manifestations achieve the most remarkable enhancement over the baseline. Using back-translation, a significant gain of +4.5 over the strong transformer baseline of 37.3 BLEU points is accomplished. Differentiating between several Alemannic dialects instead of treating Alemannic as one dialect leads to substantial improvements: Multi-dialectal translation surpasses the baseline on the dialectal test sets. However, training individual models outperforms the multi-dialectal approach. There, improvements range from 7.5 to 10.6 BLEU points over the baseline depending on the dialect.

In this work, we build a Question Answering (QA) classification dataset from a social media platform, namely the Telegram public channel called @AskAnythingEthiopia. The channel has more than 78k subscribers and has existed since May 31, 2019. The platform allows asking questions that belong to various domains, like politics, economics, health, education, and so on. Since the questions are posed in a mixed-code, we apply different strategies to pre-process the dataset. Questions are posted in Amharic, English, or Amharic but in a Latin script. As part of the pre-processing tools, we build a Latin to Ethiopic Script transliteration tool. We collect 8k Amharic and 24K transliterated questions and develop deep learning-based questions answering classifiers that attain as high as an F-score of 57.29 in 20 different question classes or categories. The datasets and pre-processing scripts are open-sourced to facilitate further research on the Amharic community-based question answering.

Automatic morphology induction is important for computational processing of natural language. In resource-scarce languages in particular, it offers the possibility of supplementing data-driven strategies of Natural Language Processing with morphological rules that may cater for out-of-vocabulary words. Unfortunately, popular approaches to unsupervised morphology induction do not work for some of the most productive morphological processes of the Yorùbá language. To the best of our knowledge, the automatic induction of such morphological processes as full and partial reduplication, infixation, interfixation, compounding and other morphological processes, particularly those based on the affixation of stem-derived morphemes have not been adequately addressed in the literature. This study proposes a method for the automatic detection of stem-derived morphemes in Yorùbá. Words in a Yorùbá lexicon of 14,670 word-tokens were clustered around “word-labels”. A word-label is a textual proxy of the patterns imposed on words by the morphological processes through which they were formed. Results confirm a conjectured significant difference between the predicted and observed probabilities of word-labels motivated by stem-derived morphemes. This difference was used as basis for automatic identification of words formed by the affixation of stem-derived morphemes. Keywords: Unsupervised Morphology Induction, Recurrent Partials, Recurrent Patterns, Stem-derived Morphemes, Word-labels.

Finite-state approaches to morphological analysis have been shown to improve the performance of natural language processing systems for polysynthetic languages, in-which words are generally composed of many morphemes, for tasks such as language modelling (Schwartz et al., 2020). However, finite-state morphological analyzers are expensive to construct and require expert knowledge of a language’s structure. Currently, there is no broad-coverage finite-state model of morphology for Wolastoqey, also known as Passamaquoddy-Maliseet, an endangered low-resource Algonquian language. As this is the case, in this paper, we investigate using two unsupervised models, MorphAGram and Morfessor, to obtain morphological segmentations for Wolastoqey. We train MorphAGram and Morfessor models on a small corpus of Wolastoqey words and evaluate using two an notated datasets. Our results indicate that MorphAGram outperforms Morfessor for morphological segmentation of Wolastoqey.

This paper presents baseline classification models for subjectivity detection, sentiment analysis, emotion analysis, sarcasm detection, and irony detection. All models are trained on user-generated content gathered from newswires and social networking services, in three different languages: English —a high-resourced language, Maltese —a low-resourced language, and Maltese-English —a code-switched language. Traditional supervised algorithms namely, Support Vector Machines, Naïve Bayes, Logistic Regression, Decision Trees, and Random Forest, are used to build a baseline for each classification task, namely subjectivity, sentiment polarity, emotion, sarcasm, and irony. Baseline models are established at a monolingual (English) level and at a code-switched level (Maltese-English). Results obtained from all the classification models are presented.

This paper presents a work-in-progress report of an open-source speech technology project for indigenous Sami languages. A less detailed description of this work has been presented in a more general paper about the whole GiellaLT language infrastructure, submitted to the LREC 2022 main conference. At this stage, we have designed and collected a text corpus specifically for developing speech technology applications, namely Text-to-speech (TTS) and Automatic speech recognition (ASR) for the Lule and North Sami languages. We have also piloted and experimented with different speech synthesis technologies using a miniature speech corpus as well as developed tools for effective processing of large spoken corpora. Additionally, we discuss effective and mindful use of the speech corpus and also possibilities to use found/archive materials for training an ASR model for these languages.

This paper reports on experiments for cross-lingual transfer using the anchor-based approach of Schuster et al. (2019) for English and a low-resourced language, namely Hindi. For the sake of comparison, we also evaluate the approach on three very different higher-resourced languages, viz. Dutch, Russian and Chinese. Initially designed for ELMo embeddings, we analyze the approach for the more recent BERT family of transformers for a variety of tasks, both mono and cross-lingual. The results largely prove that like most other cross-lingual transfer approaches, the static anchor approach is underwhelming for the low-resource language, while performing adequately for the higher resourced ones. We attempt to provide insights into both the quality of the anchors, and the performance for low-shot cross-lingual transfer to better understand this performance gap. We make the extracted anchors and the modified train and test sets available for future research at https://github.com/pranaydeeps/Vyaapak

This poster presents the first publicly available treebank of Yakut, a Turkic language spoken in Russia, and a morphological analyzer for this language. The treebank was annotated following the Universal Dependencies (UD) framework and the mor- phological analyzer can directly access and use its data. Yakut is an under-represented language whose prominence can be raised by making reliably annotated data and NLP tools that could process it freely accessible. The publication of both the treebank and the analyzer serves this purpose with the prospect of evolving into a benchmark for the development of NLP online tools for other languages of the Turkic family in the future.

Spell checkers are an integrated feature of most software applications handling text inputs. When we write an email or compile a report on a desktop or a smartphone editor, a spell checker could be activated that assists us to write more correctly. However, this assistance does not exist for all languages equally. The Kurdish language, which still is considered a less-resourced language, currently lacks spell checkers for its various dialects. We present a trigram language model for the Sorani dialect of the Kurdish language that is created using educational text. We also showcase a spell checker for the Sorani dialect of Kurdish that can assist in writing texts in the Persian/Arabic script. The spell checker was developed as a testing environment for the language model. Primarily, we use the probabilistic method and our trigram language model with Stupid Backoff smoothing for the spell checking algorithm. Our spell checker has been trained on the KTC (Kurdish Textbook Corpus) dataset. Hence the system aims at assisting spell checking in the related context. We test our approach by developing a text processing environment that checks for spelling errors on a word and context basis. It suggests a list of corrections for misspelled words. The developed spell checker shows 88.54% accuracy on the texts in the related context and it has an F1 score of 43.33%, and the correct suggestion has an 85% chance of being in the top three positions of the corrections.

Semantic relatedness between words is one of the core concepts in natural language processing, thus making semantic evaluation an important task. In this paper, we present a semantic model evaluation dataset: SimRelUz - a collection of similarity and relatedness scores of word pairs for the low-resource Uzbek language. The dataset consists of more than a thousand pairs of words carefully selected based on their morphological features, occurrence frequency, semantic relation, as well as annotated by eleven native Uzbek speakers from different age groups and gender. We also paid attention to the problem of dealing with rare words and out-of-vocabulary words to thoroughly evaluate the robustness of semantic models.

Machine Translation (MT)-empowered chatbots are not established yet, however, we see an amazing future breaking language barriers and enabling conversation in multiple languages without time-consuming language model building and training, particularly for under-resourced languages. In this paper we focus on the under-resourced Luxembourgish language. This article describes the experiments we have done with a dataset containing administrative questions that we have manually created to offer BERT QA capabilities to a multilingual chatbot. The chatbot supports visual dialog flow diagram creation (through an interface called BotStudio) in which a dialog node manages the user question at a specific step. Dialog nodes can be matched to the user’s question by using a BERT classification model which labels the question with a dialog node label.