During the last year, AI has undergone its most disruptive transformation in half a century, largely due to OpenAI's innovations and cloud democratization.ChatGPT, stable diffusion, and Dall-E have reshaped our AI understanding and applications across society.Despite ethical concerns, it's vital to examine this shift's impact on resourceconstrained entities like small and medium-sized enterprises and research groups.At its core, this transformation relies on Large Language Models (LLMs) built upon Transformer and GPT architectures.Yet, it's essential to recognize that training these models demands significant hardware budgets and access to curated datasets.Equally crucial is the labor-intensive data curation, supervision, and refinement often overlooked.Beyond training challenges, we must address inference cost, latency, concurrency, and scalability-affecting real-world viability and the quality decline in ChatGPT due to high request loads and its business model while maintaining profitability.In this keynote, we introduce LLM fundamentals and explore challenges.We delve into how this AI revolution impacts medium-sized enterprises, providing speech technology as a use case, focusing on technical and business aspects.We'll emphasize customercentric customization, domain specialization, and value addition in process automation.We will try to understand the profound implications of this AI renaissance for enterprises of all sizes.

In communications, channel models are useful approximations to the performance of a real channel, which most of the time is not available for repeated tests. In this work we present the problem of the real Near Vertical Incidence Skywave (NVIS) ionospheric scenario channel sounding, and evaluate the channel propagation characteristics in terms of frequency and time spread, with the final goal of designing a channel model. An NVIS channel model can be obtained from the evaluated channel parameters; however, on one hand, there is the problem of missing data due to bad channel performance in some frequencies, and, on the other hand, the measured parameters have strong dependencies between them that cannot be evinced directly. In this work, we conduct a first set of analyses of the measured parameters of the soundings to determine the dependencies in terms of quality of the channel propagation but refer mainly to the Doppler spread and the delay spread in the sensor. This classification approach allows us to face the second part of the research focusing on the design of the channel model for the ionospheric communication of remote sensors.

This paper describes the design, development and analysis of a MOOC entitled “Approaches to Machine Translation: rule-based, statistical and hybrid” providing lessons learnt on conclusions to be take into account in the future. The course was developed within a Canvas platform, used by recognized European universities. The course contains video-lectures, quizzes and laboratory assignments. Evaluation is done across on-line quizzes, programming assignments (PAs) evaluated by means of a specific code evaluation and peer-to-peer strategies. This MOOC allows to introduce people from various areas to the Machine Translation theory and practice. It also allows to internationally publisize different tools developed at the Universitat Polit`ecnica de Catalunya.

In this article we present a three-step methodology for dynamically improving a statistical machine translation (SMT) system by incorporating human feedback in the form of free edits on the system translations. We target at feedback provided by casual users, which is typically error-prone. Thus, we first propose a filtering step to automatically identify the better user-edited translations and discard the useless ones. A second step produces a pivot-based alignment between source and user-edited sentences, focusing on the errors made by the system. Finally, a third step produces a new translation model and combines it linearly with the one from the original system. We perform a thorough evaluation on a real-world dataset collected from the Reverso.net translation service and show that every step in our methodology contributes significantly to improve a general purpose SMT system. Interestingly, the quality improvement is not only due to the increase of lexical coverage, but to a better lexical selection, reordering, and morphology. Finally, we show the robustness of the methodology by applying it to a different scenario, in which the new examples come from an automatically Web-crawled parallel corpus. Using exactly the same architecture and models provides again a significant improvement of the translation quality of a general purpose baseline SMT system.

This paper describes the design, development and execution of a MOOC entitled “Approaches to Machine Translation: rule-based, statistical and hybrid”. The course is launched from the Canvas platform used by recognized European universities. The course contains video-lecture, quizzes and laboratory assignments. Evaluation is done using a virtual learning environment for computer programming and peer-to-peer strategies. This MOOC allows to introduce people from various areas to the Machine Translation theory and practice. It also allows to internationalize different tools developed at the Universitat Politècnica de Catalunya.

Post-editing feedback provided by users of on-line translation services offers an excellent opportunity for automatic improvement of statistical machine translation (SMT) systems. However, feedback provided by casual users is very noisy, and must be automatically filtered in order to identify the potentially useful cases. We present a study on automatic feedback filtering in a real weblog collected from Reverso.net. We extend and re-annotate a training corpus, define an extended set of simple features and approach the problem as a binary classification task, experimenting with linear and kernel-based classifiers and feature selection. Results on the feedback filtering task show a significant improvement over the majority class, but also a precision ceiling around 70-80%. This reflects the inherent difficulty of the problem and indicates that shallow features cannot fully capture the semantic nature of the problem. Despite the modest results on the filtering task, the classifiers are proven effective in an application-based evaluation. The incorporation of a filtered set of feedback instances selected from a larger corpus significantly improves the performance of a phrase-based SMT system, according to a set of standard evaluation metrics.

Research on translation quality annotation and estimation usually makes use of standard language, sometimes related to a specific language genre or domain. However, real-life machine translation (MT), performed for instance by on-line translation services, has to cope with some extra dif- ficulties related to the usage of open, non-standard and noisy language. In this paper we study the learning of quality estimation (QE) models able to rank translations from real-life input according to their goodness without the need of translation references. For that, we work with a corpus collected from the 24/7 Reverso.net MT service, translated by 5 different MT systems, and manually annotated with quality scores. We define several families of features and train QE predictors in the form of regressors or direct rankers. The predictors show a remarkable correlation with gold standard rankings and prove to be useful in a system combination scenario, obtaining better results than any individual translation system.

This paper describes the TALP-UPC participation in the WMT’13 Shared Task on Quality Estimation (QE). Our participation is reduced to task 1.2 on System Selection. We used a broad set of features (86 for German-to-English and 97 for English-to-Spanish) ranging from standard QE features to features based on pseudo-references and semantic similarity. We approached system selection by means of pairwise ranking decisions. For that, we learned Random Forest classifiers especially tailored for the problem. Evaluation at development time showed considerably good results in a cross-validation experiment, with Kendall’s values around 0.30. The results on the test set dropped significantly, raising different discussions to be taken into account.

This paper describes the TALP participation in the WMT13 evaluation campaign. Our participation is based on the combination of several statistical machine translation systems: based on standard phrasebased Moses systems. Variations include techniques such as morphology generation, training sentence filtering, and domain adaptation through unit derivation. The results show a coherent improvement on TER, METEOR, NIST, and BLEU scores when compared to our baseline system.

This paper presents a detailed study of a method for morphology generalization and generation to address out-of-domain translations in English-to-Spanish phrase-based MT. The paper studies whether the morphological richness of the target language causes poor quality translation when translating out-ofdomain. In detail, this approach first translates into Spanish simplified forms and then predicts the final inflected forms through a morphology generation step based on shallow and deep-projected linguistic information available from both the source and targetlanguage sentences. Obtained results highlight the importance of generalization, and therefore generation, for dealing with out-ofdomain data.

This paper describes the UPC participation in the WMT 12 evaluation campaign. All systems presented are based on standard phrase-based Moses systems. Variations adopted several improvement techniques such as morphology simplification and generation and domain adaptation. The morphology simplification overcomes the data sparsity problem when translating into morphologically-rich languages such as Spanish by translating first to a morphology-simplified language and secondly leave the morphology generation to an independent classification task. The domain adaptation approach improves the SMT system by adding new translation units learned from MT-output and reference alignment. Results depict an improvement on TER, METEOR, NIST and BLEU scores compared to our baseline system, obtaining on the official test set more benefits from the domain adaptation approach than from the morphological generalization method.

Misspelled words have a direct impact on the final quality obtained by Statistical Machine Translation (SMT) systems as the input becomes noisy and unpredictable. This paper presents some improvement strategies for translating real-life noisy input. The proposed strategies are based on a preprocessing step consisting in a character-based translator (MT) from noisy into cleaned text. The use of a character-level translator allows us to provide various spelling alternatives in a lattice format to the final bilingual translator. Therefore, the final MT is the one that decides the best path to be translated. The different hypotheses are obtained under the assumption of a noisy channel model for this task. This paper shows the experiments done with real-life noisy input and a standard phrase-based SMT system from English into Spanish.

The tuning process of Unit Selection TTS (US-TTS) system is usually performed by an expert that typically conducts the task of weighting the cost function by hand. However, hand tuning is costly in terms of the required training time and inaccurate and ambiguous in terms of methodology. With the purpose of easing the task of properly tuning the weights of the cost function, this thesis make its contribution from a perceptual-based approach using of active interactive Genetic Algorithms (aiGAs). The thesis pursues four major guidelines: i) accuracy when tuning the weights, ii) robustness of the obtained weights, iii)real world applicability of the methodology to any cost function design, and iv)finding consensus of the different users when tuning the weights. The experimentation is carried out through a small and medium sized corpus (1.9h) applied to different configurations (type of features) of the US-TTS cost function. The thesis concludes that aiGAs are highly competitive in comparison to other weight tuning techniques from the state-of-the-art

We present a corpus consisting of 11,292 real-world English to Spanish automatic translations annotated with relative (ranking) and absolute (adequate/non-adequate) quality assessments. The translation requests, collected through the popular translation portal http://reverso.net, provide a most variated sample of real-world machine translation (MT) usage, from complete sentences to units of one or two words, from well-formed to hardly intelligible texts, from technical documents to colloquial and slang snippets. In this paper, we present 1) a preliminary annotation experiment that we carried out to select the most appropriate quality criterion to be used for these data, 2) a graph-based methodology inspired by Interactive Genetic Algorithms to reduce the annotation effort, and 3) the outcomes of the full-scale annotation experiment, which result in a valuable and original resource for the analysis and characterization of MT-output quality.

We present a detailed analysis of a graph- based annotation strategy that we employed to annotate a corpus of 11,292 real-world En- glish to Spanish automatic translations with relative (ranking) and absolute (adequate/non- adequate) quality assessments. The proposed approach, inspired by previous work in In- teractive Evolutionary Computation and Inter- active Genetic Algorithms, results in a sim- pler and faster annotation process. We em- pirically compare the method against a tra- ditional, explicit ranking approach, and show that the graph-based strategy: 1) is consider- ably faster, and 2) produces consistently more reliable annotations

Misspelled words have a direct impact on the final quality obtained by Statistical Machine Translation (SMT) systems as the input becomes noisy and unpredictable. This paper presents some improvement strategies for translating real-life noisy input. The proposed strategies are based on a preprocessing step consisting in a character-based translator (MT) from noisy into cleaned text. The use of a character-level translator allows us to provide various spelling alternatives in a lattice format to the final bilingual translator. Therefore, the final MT is the one that decides the best path to be translated. The different hypotheses are obtained under the assumption of a noisy channel model for this task. This paper shows the experiments done with real-life noisy input and a standard phrase-based SMT system from English into Spanish.

Unit-selection speech synthesis is one of the current corpus-based text-to-speech synthesis techniques. The quality of the generated speech depends on the accuracy of the unit selection process, which in turn relies on the cost function definition. This function should map the user perceptual preferences when selecting synthesis units, which is still an open research issue. This paper proposes a complete methodology for the tuning of the cost function weights by fusing the human judgments with the cost function, through efficient and reliable interactive weight tuning. To that effect, active interactive genetic algorithms (aiGAs) are used to guide the subjective weight adjustments. The application of aiGAs to this process allows mitigating user fatigue and frustration by improving user consistency. However, it is still unfeasible to subjectively adjust the weights of the whole corpus units (diphones and triphones in this work). This makes it mandatory to perform unit clustering before conducting the tuning process. The aiGA-based weight tuning proposal is evaluated in a small speech corpus as a proof-of-concept and results in more natural synthetic speech when compared to previous objective and subjective-based approaches.

This paper describes different pivot approaches to built SMT systems for language pairs with scarce parallel resources. The strategy is particularly interesting for Spain, a country with three official languages (Catalan, Basque, and Galician) besides Spanish, where it is difficult to find parallel corpora between two of the first three mentioned languages but it is relatively easy to collect it between Spanish and any of them. This characteristic, however, allow us to develop machine translation systems from major languages like English, to Catalan for instance, using Spanish as pivot. Such systems help these minority languages giving them global presence and promoting their use in content collaboration. We describe a English-Catalan baseline system built following the synthetic approach, we compare it with the transfer approach and comment about future enhancement that could be implemented for this language pair.

This paper presents the text-to-speech (TTS) synthesis sys tem of La Salle (Universitat Ramon Llull, URL) and its adaptatio n t the Albayzin Evaluation Campaign of FALA2010 conference. The URL-TTS system follows the classical scheme of unit selection TTS synthesis systems. However, it presents two dis tinguishable particularities: i) prosody prediction learned from labelled data by means of Case-Based-Reasoning (CBR) and perceptual weight tuning by means of active interactive Gen etic Algorithms (aiGA). The aiGA-based weights are compared to multilinear regression (MLR) weights both considering cla ssical averaged cost function and its root-mean squared varian t. The internal validation tests and the results of the evaluat ion campaing are described, and finally discussed.