Cross-lingual word embeddings are an increasingly important reseource in cross-lingual methods for NLP, particularly for their role in transfer learning and unsupervised machine translation, purportedly opening up the opportunity for NLP applications for low-resource languages.  However, most research in this area implicitly expects the availablility of vast monolingual corpora for training embeddings, a scenario which is not realistic for many of the world's languages.  Moreover, much of the reporting of the performance of cross-lingual word embeddings is based on a fairly narrow set of mostly European language pairs.  Our study examines the performance of cross-lingual alignment across a more diverse set of language pairs; controls for the effect of the corpus size on which the monolingual embedding spaces are trained; and studies the impact of spectral graph properties of the embedding spsace on alignment.  Through our experiments on a more diverse set of language pairs, we find that performance in bilingual lexicon induction is generally poor in heterogeneous pairs, and that even using a gold or heuristically derived dictionary has little impact on the performance on these pairs of languages.  We also find that the performance for these languages only increases slowly with corpus size.  Finally, we find a moderate correlation between the isospectral difference of the source and target embeddings and the performance of bilingual lexicon induction.  We infer that methods other than cross-lingual alignment may be more appropriate in the case of both low resource languages and heterogeneous language pairs.

京都大学 / 0048 / 新制・課程博士 / 博士(情報学) / 甲第21210号 / 情博第663号 / 新制||情||114(附属図書館) / 京都大学大学院情報学研究科知能情報学専攻 / (主査)教授 黒橋 禎夫, 教授 河原 達也, 教授 森 信介 / 学位規則第4条第1項該当 / Doctor of Informatics / Kyoto University / DFAM

No description available.

Nya neurala modeller har lett till stora framsteg inom maskinöversättning, men fungerar fortfarande sämre på språk som saknar stora mängder parallella data, så kallade lågresursspråk. Gawarbati är ett litet, hotat lågresursspråk där endast 5000 parallella meningar finns tillgängligt. Denna uppsats använder överföringsinlärning och hyperparametrar optimerade för små datamängder för att undersöka möjligheter och begränsningar för neural maskinöversättning från gawarbati till engelska. Genom att använda överföringsinlärning där en föräldramodell först tränades på hindi-engelska förbättrades översättningar med 1.8 BLEU och 1.3 chrF. Hyperparametrar optimerade för små datamängder ökade BLEU med 0.6 men minskade chrF med 1. Att kombinera överföringsinlärning och hyperparametrar optimerade för små datamängder försämrade resultatet med 0.5 BLEU och 2.2 chrF. De neurala modellerna jämförs med och presterar bättre än ordbaserad statistisk maskinöversättning och GPT-3. Den bäst presterande modellen uppnådde endast 2.8 BLEU och 19 chrF, vilket belyser begränsningarna av maskinöversättning på lågresursspråk samt det kritiska behovet av mer data. / Recent neural models have led to huge improvements in machine translation, but performance is still suboptimal for languages without large parallel datasets, so called low resource languages. Gawarbati is a small, threatened low resource language with only 5000 parallel sentences. This thesis uses transfer learning and hyperparameters optimized for small datasets to explore possibilities and limitations for neural machine translation from Gawarbati to English. Transfer learning, where the parent model was trained on parallel data between Hindi and English, improved results by 1.8 BLEU and 1.3 chrF. Hyperparameters optimized for small datasets increased BLEU by 0.6 but decreased chrF by 1. Combining transfer learning and hyperparameters optimized for small datasets led to a decrease in performance by 0.5 BLEU and 2.2 chrF. The neural models outperform a word based statistical machine translation and GPT-3. The highest performing model only achieved 2.8 BLEU and 19 chrF, which illustrates the limitations of machine translation for low resource languages and the critical need for more data. / VR 2020-01500

The struggle of training adequate MT models using data-hungry NMT frameworks for low-resource language pairs has created a need to alleviate the scarcity of sufficiently large parallel corpora. Different transfer learning methods have been introduced as possible solutions to this problem, where a new model for a target task is initialized using parameters learned from some other high-resource task. Many of these methods are claimed to increase the translation quality of NMT systems in some low-resource environments, however, they are often proven to do so using different parent and child language pairs, a variation in data size, NMT frameworks, and training hyperparameters, which makes comparing them impossible. In this thesis project, three such transfer learning methods are put head-to-head in a controlled environment where the target task is to translate from the under-resourced Baltic languages Lithuanian and Latvian to English. In this controlled environment, the same parent language pairs, data sizes, data domains, transformer framework, and training parameters are used to ensure fair comparisons between the three transfer learning methods. The experiments involve training and testing models using all different combinations of transfer learning methods, parent language pairs, and either in-domain or out-domain data for an extensive study where different strengths and weaknesses are observed. The results display that Multi-Round Transfer Learning improves the overall translation quality the most but, at the same time, requires the longest training time by far. The Parameter freezing method provides a marginally lower overall improvement of translation quality but requires only half the training time, while Trivial Transfer learning improves quality the least. Both Polish and Russian work well as parents for the Baltic languages, while web-crawled data improves out-domain translations the most. The results suggest that all transfer learning methods are effective in a simulated low-resource environment, however, none of them can compete with simply having a larger target language pair data set, due to none of them overcoming the strong higher-resource baseline.

The availability of the stenographic manuscripts of Astrid Lindgren have sparked an interest in the creation of a language model for stenography. By its very nature stenography is low-resource and the unavailability of data requires a tool for using normal data. The tool presented in this thesis is to create stenographic data from manipulating orthographic data. Stenographic data is distinct from orthographic data through three different types manipulations that can be carried out. Firstly stenography is based on a phonetic version of language, secondly it used its own alphabet that is distinct from normal orthographic data, and thirdly it used several techniques to compress the data. The first type of manipulation is done by using a grapheme-to-phoneme converter. The second type is done by using an orthographic representation of a stenographic alphabet. The third type of manipulation is done by manipulating based on subword level, word level and phrase level. With these manipulations different datasets are created with different combinations of these manipulations. Results are measured for both perplexity on a GPT-2 language model and for compression rate on the different datasets. These results show a general decrease of perplexity scores and a slight compression rate across the board. We see that the lower perplexity scores are possibly due to the growth of ambiguity.

Cross-lingual transfer learning has been the driving force of low-resource natural language processing in recent years, relying on massively multilingual language models with hopes of solving the data scarcity issue for languages with a limited digital presence. However, this "one-size-fits-all" approach is not equally applicable to all low-resource languages, suggesting limitations of such models in cross-lingual transfer. Besides, known similarities and phylogenetic relationships between source and target languages are often overlooked. In this work, the emphasis is placed on Faroese, a low-resource North Germanic language with several closely related resource-rich sibling languages. The cross-lingual transfer potential from these strong Scandinavian source candidates, as well as from additional genetically related, geographically proximate, and syntactically similar source languages is studied in single-source and multi-source experiments, in terms of Faroese syntactic parsing and part-of-speech tagging. In addition, the effect of task-specific fine-tuning on monolingual, linguistically informed smaller multilingual, and massively multilingual pre-trained language models is explored. The results suggest Icelandic as a strong source candidate, however, only when fine-tuning a monolingual model. With multilingual models, task-specific fine-tuning in Norwegian and Swedish seems even more beneficial. Although they do not surpass fully Scandinavian fine-tuning, models trained on genetically related and syntactically similar languages produce good results. Additionally, the findings indicate that multilingual models outperform models pre-trained on a single language, and that even better results can be achieved using a smaller, linguistically informed model, compared to a massively multilingual one.

Tato práce se zabývá vyhledáváním výrazů v řeči pomocí mluvených příkladů (QbE STD). Výrazy jsou zadávány v mluvené podobě a jsou vyhledány v množině řečových nahrávek, výstupem vyhledávání je seznam detekcí s jejich skóre a časováním. V práci popisujeme, analyzujeme a srovnáváme tři různé přístupy ke QbE STD v jazykově závislých a jazykově nezávislých podmínkách, s jedním a pěti příklady na dotaz. Pro naše experimenty jsme použili česká, maďarská, anglická a arabská (levantská) data, a pro každý z těchto jazyků jsme natrénovali 3-stavový fonémový rozpoznávač. To nám dalo 16 možných kombinací jazyka pro vyhodnocení a jazyka na kterém byl natrénovaný rozpoznávač. Čtyři kombinace byly tedy závislé na jazyce (language-dependent) a 12 bylo jazykově nezávislých (language-independent). Všechny QbE systémy byly vyhodnoceny na stejných datech a stejných fonémových posteriorních příznacích, pomocí metrik: nesdružené Figure-of-Merit (non pooled FOM) a námi navrhnuté nesdružené Figure-of-Merit se simulací normalizace přes promluvy (utterrance-normalized non-pooled Figure-of-Merit). Ty nám poskytly relevantní údaje pro porovnání těchto QbE přístupů a pro získání lepšího vhledu do jejich chování. QbE přístupy použité v této práci jsou: sekvenční statistické modelování (GMM/HMM), srovnávání vzorů v příznacích (DTW) a srovnávání grafů hypotéz (WFST). Abychom porovnali výsledky QbE přístupů s běžnými STD systémy vyhledávajícími textové výrazy, vyhodnotili jsme jazykově závislé konfigurace také s akustickým detektorem klíčových slov (AKWS) a systémem pro vyhledávání fonémových řetězců v grafech hypotéz (WFSTlat). Jádrem této práce je vývoj, analýza a zlepšení systému WFST QbE STD, který po zlepšení dosahuje podobných výsledků jako DTW systém v jazykově závislých podmínkách.

Dealing with low-resource languages is a challenging task, because of the absence of sufficient data to train machine-learning models to make predictions on these languages. One way to deal with this problem is to use data from higher-resource languages, which enables the transfer of learning from these languages to the low-resource target ones. The present study focuses on dependency parsing and part-of-speech tagging of low-resource languages belonging to the spoken genre, i.e., languages whose treebank data is transcribed speech. These are the following: Beja, Chukchi, Komi-Zyrian, Frisian-Dutch, and Cantonese. Our approach involves investigating different types of transfer languages, employing MACHAMP, a state-of-the-art parser and tagger that uses contextualized word embeddings, mBERT, and XLM-R in particular. The main idea is to explore how the genre, the language similarity, none of the two, or the combination of those affect the model performance in the aforementioned downstream tasks for our selected target treebanks. Our findings suggest that in order to capture speech-specific dependency relations, we need to incorporate at least a few genre-matching source data, while language similarity-matching source data are a better candidate when the task at hand is part-of-speech tagging. We also explore the impact of multi-task learning in one of our proposed methods, but we observe minor differences in the model performance.

No description available.