The Transformer model is a very recent, fast and powerful discovery in neural machine translation. We experiment with multi-task learning for enriching the source side of the Transformer with linguistic resources to provide it with additional information to learn linguistic and world knowledge better. We analyze two approaches: the basic shared model with multi-tasking through simple data manipulation, and multi-decoder models. We test joint models for machine translation (MT) and POS tagging, dependency parsing and named entity recognition as the secondary tasks. We evaluate them in comparison with the baseline and with dummy, linguistically unrelated tasks. We focus primarily on the standard- size data setting for German-to-Czech MT. Although our enriched models did not significantly outperform the baseline, we empirically document that (i) the MT models benefit from the secondary linguistic tasks; (ii) considering the amount of training data consumed, the multi-tasking models learn faster; (iii) in low-resource conditions, the multi-tasking significantly improves the model; (iv) the more fine-grained annotation of the source as the secondary task, the higher benefit to MT.

Promotion of cycling is important to reach the goals for climate mitigation of the Paris Agreement and Goals ofthe Agenda 2030. Sustainable transport, both rural and urban, could contribute to at least seven of the 17 Sustainable Development Goals (ITDP 2015). There is relatively little research on cycling in Africa, and there is also much less research on cycling at night. Some studies show the importance of road lighting for minimising the reduction in the numbers of cyclists after-dark and suggest 'only a minimal amount of lighting can promote cycling after-dark, making it an attractive mode of transport year-round' (Uttley at el. 2020). So far, these studies have little relation to the situation in developing countries, which is why a first study in Nairobi, Kenya, is carried out here as an example. ... [From: Introduction]

Urban quality of life is measured by how clean the environment is, how safe people feel, how close they are to green spaces, and in general by the quality of outdoor space. Good quality public spaces are spaces that reduce road accidents through managing appropriately different transport modes, especially walking and cycling [1]. Cycling is healthy, economical, and environmentally sound form of mobility that is fundamental to life. More than one billion of the people in African cities walk or cycle for more than 55 minutes every day - to reach work, home, school, and other essential services [2]. One-third of the population of the African continent uses active mobility as a daily means of transport. This reveals that there is a potential of using cycling as a daily mode of travel in Africa. However, the poor quality of infrastructure for cycling sends a message that cyclists are not welcome in the urban environment. Despite the widespread use of non-motorized modes, transport planning and the provision of infrastructure in most of the cities in Africa have become carcentered, undermining the importance of cycling and walking. While the majority in the global south are active mobility users, they are not being respected by the public policies and experience 93% of the world's traffic fatalities and injuries [3]. Road traffic accidents are a major shes are not different in Africa. The World Health Organization Global Status report on Road Safety 2018 showed that the African region had 26.6 deaths per 100,000 populations, which is the highest among all regions [5]. Sub-Saharan Africa still has the highest per capita rate of road fatalities of any region in the world. Unfortunately, in most cases, the victims of traffic casualties are primarily pedestrians and cyclists [6]. Much of that is linked to the neglect of the infrastructure needs for pedestrians' and cyclists' safety. [From: Introduction]

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

​​Optical character recognition (Optical Character Recognition (OCR)) has many applications, such as digitizing historical documents, automating processes, and helping visually impaired people read. However, extracting text from images into a digital format is not an easy problem to solve, and the outputs from the OCR frameworks often include errors. The complexity comes from the many variations in (digital) fonts, handwriting, lighting, etc. To tackle this problem, this thesis investigates two different methods for correcting the errors in OCR output. The used dataset consists of Swedish addresses. The methods are therefore applied to postal automation to investigate the usage of these methods for further automating postal work by automatically reading addresses on parcels using OCR. The main method, the lexical implementation, uses a dataset of Swedish addresses so that any valid address should be in this dataset (hence there is a known and limited vocabulary), and misspelled addresses are corrected to the address in the lexicon with the smallest Levenshtein distance. The second approach is to use the same dataset, but with artificial errors, or artificial noise, added. The addresses with this artificial noise are then used together with their correct spelling to train a machine learning model based on Neural machine translation (Neural Machine Translation (NMT)) to automatically correct errors in OCR read addresses. The results from this study could contribute by defining in what direction future work connected to OCR and postal addresses should go. The results were that the lexical implementation outperformed the NMT model. However, more experiments including real data would be required to draw definitive conclusions as to how the methods would work in real-life applications. / Optisk teckenigenkänning (Optical Character Recognition (OCR)) har många användningsområden, till exempel att digitalisera historiska dokument, automatisera processer och hjälpa synskadade att läsa. Att extrahera text från bilder till ett digitalt format är dock inte ett lätt problem att lösa, och utdata från OCR-ramverken innehåller ofta fel. Komplexiteten kommer från de många variationerna i (digitala) typsnitt, handstil, belysning, etc. För att lösa problemet undersöker den här avhandling två olika metoder för att rätta fel i OCR-utdata. Det använda datasetet består av svenska adresser. Metoderna tillämpas därför på postautomatisering för att undersöka användningen av dessa metoder för att ytterligare automatisera postarbetet genom att automatiskt läsa adresser på paket med OCR. Den första metoden, den lexikaliska metoden, använder en datauppsättning av svenska adresser så att alla giltiga adresser bör finnas i denna datauppsättning (därav finns det ett känt och begränsat ordförråd). Denna datauppsättning används sedan som en ordbok för att hitta adressen med det minsta Levenshtein-avståndet till någon felstavad adress. Det andra tillvägagångssättet använder samma datauppsättning, men med artificiella fel tillagda. Adresserna med dessa artificiella fel används sedan tillsammans med deras korrekta stavning för att träna en Neural Machine Translation (NMT)-modell för att automatiskt korrigera fel i OCR-lästa adresser. Resultaten från denna studie skulle kunna bidra genom att definiera i vilken riktning framtida arbete kopplat till OCR och postadresser ska gå. Resultaten var att den lexikaliska metoden presterade bättre än NMT-modellen. Fler experiment gjorde med verklig data skulle dock behövas för att dra definitiva slutsatser om hur metoderna skulle fungera i verkliga tillämpningar.

No description available.

Most organizations use an increasing number of domain- or organization-specific words and phrases. A translation process, whether human or automated, must also be able to accurately and efficiently use these specific multilingual terminology collections. However, comparatively little has been done to explore the use of vetted terminology as an input to machine translation (MT) for improved results. In fact, no single established process currently exists to integrate terminology into MT as a general practice, and especially no established process for neural machine translation (NMT) exists to ensure that the translation of individual terms is consistent with an approved terminology collection. The use of tokenization as a method of injecting terminology and of evaluating terminology injection is the focus of this thesis. I use the attention mechanism prevalent in state-of-the-art NMT systems to produce the desired results. Attention vectors play an important part of this method to correctly identify semantic entities and to align the tokens that represent them. My methods presented in this thesis use these attention vectors to align the source tokens in the sentence to be translated with the target tokens in the final translation output. Then, supplied terminology is injected, where these alignments correctly identify semantic entities. My methods demonstrate significant improvement to the state-of-the-art results for NMT using terminology injection.

<p>We have 3G on the doorstep but nothing seems to attract ordinary people to this technology. To attract the mass market the telecom industry must show something beyond high bit rates. They must show how ordinary people can take advantage of this new technology. This is done by showing the possibilities of the new technology and by demonstrating applications that it will handle. The telecom industry must convince the telecom operators to invest in this technology and the only thing that matters to them is how much revenue they can make by adopting the upcoming technology. </p><p>To convince the operators industry must show how the operators can charge for the new types of applications that will be introduced soon. This is the main reason why this Master's Thesis has been conducted. The purpose of this thesis is to provide a demonstration to Ericsson's 3G lab in Katrineholm in the form of an IP application with a billing solution. This thesis describes the migration from 1G to 3G and examines existing and future billing strategies as well. </p><p>The IP application is an application that uses progressive streaming in order to stream multimedia content to a PDA connected to a 3G phone. This application is platform independent because it is placed on leading Web servers, Apache and IIS. </p><p>The billing application consists of a number of steps. The first step is logging, which is performed by the Web server on which the streaming application is placed. The second step, processing and billing, is performed in the BGw, which is Ericsson's mediation tool, and the SQL server.The third step is displaying the bill, which is done by using ASP to create an active HTML page.</p>

We have 3G on the doorstep but nothing seems to attract ordinary people to this technology. To attract the mass market the telecom industry must show something beyond high bit rates. They must show how ordinary people can take advantage of this new technology. This is done by showing the possibilities of the new technology and by demonstrating applications that it will handle. The telecom industry must convince the telecom operators to invest in this technology and the only thing that matters to them is how much revenue they can make by adopting the upcoming technology. To convince the operators industry must show how the operators can charge for the new types of applications that will be introduced soon. This is the main reason why this Master's Thesis has been conducted. The purpose of this thesis is to provide a demonstration to Ericsson's 3G lab in Katrineholm in the form of an IP application with a billing solution. This thesis describes the migration from 1G to 3G and examines existing and future billing strategies as well. The IP application is an application that uses progressive streaming in order to stream multimedia content to a PDA connected to a 3G phone. This application is platform independent because it is placed on leading Web servers, Apache and IIS. The billing application consists of a number of steps. The first step is logging, which is performed by the Web server on which the streaming application is placed. The second step, processing and billing, is performed in the BGw, which is Ericsson's mediation tool, and the SQL server.The third step is displaying the bill, which is done by using ASP to create an active HTML page.