Search

In the context of an epidemiological study involving multilingual social media, this paper reports on the ability of machine translation systems to preserve content relevant for a document classification task designed to determine whether the social media text is related to covid-19. The results indicate that machine translation does provide a feasible basis for scaling epidemiological social media surveillance to multiple languages. Moreover, a qualitative error analysis revealed that the majority of classification errors are not caused by MT errors.

We report our efforts in identifying a set of previous humane valuations in NLP that would be suitable for a coordinated study examining what makes human evaluations in NLP more/less reproducible. We present our results and findings, which include that just13% of papers had (i) sufficiently low barriers to reproduction, and (ii) enough obtainable information, to be considered for reproduction, and that all but one of the experiments we selected for reproduction was discovered to have flaws that made the meaningfulness of conducting are production questionable. As a result, we had to change our coordinated study design from a reproduce approach to a standardise-then-reproduce-twice approach. Our overall (negative)finding that the great majority of human evaluations in NLP is not repeatable and/or not reproducible and/or too flawed to justify reproduction, paints a dire picture, but presents an opportunity for a rethink about how to design and report human evaluations in NLP

The goal of this paper is to learn more about how idiomatic information is structurally encoded in embeddings, using a structural probing method. We repurpose an existing English verbal multi-word expression (MWE) dataset to suit the probing framework and perform a comparative probing study of static (GloVe) and contextual (BERT) embeddings. Our experiments indicate that both encode some idiomatic information to varying degrees, but yield conflicting evidence as to whether idiomaticity is encoded in the vector norm, leaving this an open question. We also identify some limitations of the used dataset and highlight important directions for future work in improving its suitability for a probing analysis., 9 pages, 5 tables, In proceedings of the 19th Workshop on Multiword Expressions @ EACL2023

Modelling taxonomic and thematic relatedness is important for building AI with comprehensive natural language understanding. The goal of this paper is to learn more about how taxonomic information is structurally encoded in embeddings. To do this, we design a new hypernym-hyponym probing task and perform a comparative probing study of taxonomic and thematic SGNS and GloVe embeddings. Our experiments indicate that both types of embeddings encode some taxonomic information, but the amount, as well as the geometric properties of the encodings, are independently related to both the encoder architecture, as well as the embedding training data. Specifically, we find that only taxonomic embeddings carry taxonomic information in their norm, which is determined by the underlying distribution in the data., 9 pages, 1 figure, 4 tables, In proceedings of the 12th International Global Wordnet Conference

Consistency is a key requirement of high-quality translation. It is especially important to adhere to pre-approved terminology and adapt to corrected translations in domain-specific projects. Machine translation (MT) has achieved significant progress in the area of domain adaptation. However, real-time adaptation remains challenging. Large-scale language models (LLMs) have recently shown interesting capabilities of in-context learning, where they learn to replicate certain input-output text generation patterns, without further fine-tuning. By feeding an LLM at inference time with a prompt that consists of a list of translation pairs, it can then simulate the domain and style characteristics. This work aims to investigate how we can utilize in-context learning to improve real-time adaptive MT. Our extensive experiments show promising results at translation time. For example, LLMs can adapt to a set of in-domain sentence pairs and/or terminology while translating a new sentence. We observe that the translation quality with few-shot in-context learning can surpass that of strong encoder-decoder MT systems, especially for high-resource languages. Moreover, we investigate whether we can combine MT from strong encoder-decoder models with fuzzy matches, which can further improve translation quality, especially for less supported languages. We conduct our experiments across five diverse language pairs, namely English-to-Arabic (EN-AR), English-to-Chinese (EN-ZH), English-to-French (EN-FR), English-to-Kinyarwanda (EN-RW), and English-to-Spanish (EN-ES)., Comment: EAMT 2023 - Research: technical

Preservation of domain knowledge from the source to target is crucial in any translation workflow. It is common in the translation industry to receive highly specialized projects, where there is hardly any parallel in-domain data. In such scenarios where there is insufficient in-domain data to fine-tune Machine Translation (MT) models, producing translations that are consistent with the relevant context is challenging. In this work, we propose a novel approach to domain adaptation leveraging state-of-the-art pretrained language models (LMs) for domain-specific data augmentation for MT, simulating the domain characteristics of either (a) a small bilingual dataset, or (b) the monolingual source text to be translated. Combining this idea with back-translation, we can generate huge amounts of synthetic bilingual in-domain data for both use cases. For our investigation, we use the state-of-the-art Transformer architecture. We employ mixed fine-tuning to train models that significantly improve translation of in-domain texts. More specifically, in both scenarios, our proposed methods achieve improvements of approximately 5-6 BLEU and 2-3 BLEU, respectively, on the Arabic-to-English and English-to-Arabic language pairs. Furthermore, the outcome of human evaluation corroborates the automatic evaluation results., AMTA 2022 - MT Research Track

Collaborative robots can improve ergonomics on factory floors while allowing a higher level of flexibility in production. The evolution of robotics and cyber-physical systems in size and functionality has enabled new applications which were never foreseen in traditional industrial robots. However, the current human-robot collaboration (HRC) technologies are limited in reliability and safety, which are vital in risk-critical scenarios. Certainly, confusion about European safety regulations has led to situations where collaborative robots operate behind security barriers, thus negating their advantages while reducing overall application productivity. Despite recent advances, developing a safe collaborative robotic system for performing complex industrial or daily tasks remains a challenge. Multiple influential factors in HRC make it difficult to define a clear classification to understand the depth of collaboration between humans and robots. In this article, we review the state of the art in reliable collaborative robotic work cells and propose a reference model to combine influential factors such as robot autonomy, collaboration, and safety modes to redefine HRC categorization.

Background The aetiology of ANCA-associated vasculitis (AAV) and triggers of relapse are poorly understood. Vitamin D (vitD) is an important immunomodulator, potentially responsible for the observed latitudinal differences between granulomatous and non-granulomatous AAV phenotypes. A narrow ultraviolet B spectrum induces vitD synthesis (vitD-UVB) via the skin. We hypothesised that prolonged periods of low ambient UVB (and by extension vitD deficiency) are associated with the granulomatous form of the disease and an increased risk of AAV relapse. Methods Patients with AAV recruited to the Irish Rare Kidney Disease (RKD) (n = 439) and UKIVAS (n = 1961) registries were studied. Exposure variables comprised latitude and measures of ambient vitD-UVB, including cumulative weighted UVB dose (CW-D-UVB), a well-validated vitD proxy. An n-of-1 study design was used to examine the relapse risk using only the RKD dataset. Multi-level models and logistic regression were used to examine the effect of predictors on AAV relapse risk, phenotype and serotype. Results Residential latitude was positively correlated (OR 1.41, 95% CI 1.14–1.74, p = 0.002) and average vitD-UVB negatively correlated (0.82, 0.70–0.99, p = 0.04) with relapse risk, with a stronger effect when restricting to winter measurements (0.71, 0.57–0.89, p = 0.002). However, these associations were not restricted to granulomatous phenotypes. We observed no clear relationship between latitude, vitD-UVB or CW-D-UVB and AAV phenotype or serotype. Conclusion Our findings suggest that low winter ambient UVB and prolonged vitD status contribute to AAV relapse risk across all phenotypes. However, the development of a granulomatous phenotype does not appear to be directly vitD-mediated. Further research is needed to determine whether sufficient vitD status would reduce relapse propensity in AAV.

Improving our understanding of how information is encoded in vector space can yield valuable interpretability insights. Alongside vector dimensions, we argue that it is possible for the vector norm to also carry linguistic information. We develop a method to test this: an extension of the probing framework which allows for relative intrinsic interpretations of probing results. It relies on introducing noise that ablates information encoded in embeddings, grounded in random baselines and confidence intervals. We apply the method to well-established probing tasks and find evidence that confirms the existence of separate information containers in English GloVe and BERT embeddings. Our correlation analysis aligns with the experimental findings that different encoders use the norm to encode different kinds of information: GloVe stores syntactic and sentence length information in the vector norm, while BERT uses it to encode contextual incongruity., Comment: 10 pages, 3 tables, Workshop on analyzing and interpreting neural networks for NLP

An open research question in deep reinforcement learning is how to focus the policy learning of key decisions within a sparse domain. This paper emphasizes combining the advantages of inputoutput hidden Markov models and reinforcement learning towards interpretable maintenance decisions. We propose a novel hierarchical-modeling methodology that, at a high level, detects and interprets the root cause of a failure as well as the health degradation of the turbofan engine, while, at a low level, it provides the optimal replacement policy. It outperforms the baseline performance of deep reinforcement learning methods applied directly to the raw data or when using a hidden Markov model without such a specialized hierarchy. It also provides comparable performance to prior work, however, with the additional benefit of interpretability.

Additional file 1: Fig. S1. Flowchart of participant selection for the UKIVAS cohort (1). Fig. S2. Flowchart of participant selection for the RKD cohort (2). Fig. S3. Participant locations at diagnosis for i). UKIVAS and ii). RKD cohort. Shading represents the absolute number of participants recruited to the study in that region, at the time of their diagnosis. Tendency to cluster around vasculitis referral centres is observed. Fig. S4. Distribution of vitD-UVB (kJ/m2) (a and b) and CW-D-UVB (kJ/m2) (c and d) by month and season: CW-D-UVB peaks in late summer and nadirs in winter. Fig. S5. The relationship between latitude (degrees) and a). annual daily average vitD-UVB (kJ/m2, Pearson correlation coefficient (corr) -0.98, p

Since our communication is multimodal in terms of our ability to express ourselves via different channels and our perception of the world, the automatic production and analysis of natural language content requires the integration of these multiple modalities in order to rival human performance. However, multimodal, or cross-modal, Natural Language Processing (NLP) has long been in the minority, perhaps because it is more complex. In the wake of recent advances in artificial intelligence, which are bringing multimodality to the fore, this special issue aims to highlight, through three articles on a variety of subjects, the questions that remain, particularly with regard to data requirements, understanding the links between modalities, and the need for convergence in terms of representation and modelling.

Applying deep learning models to MRI scans of acute stroke patients to extract features that are indicative of short-term outcome could assist a clinician’s treatment decisions. Deep learning models are usually accurate but are not easily interpretable. Here, we trained a convolutional neural network on ADC maps from hyperacute ischaemic stroke patients for prediction of short-term functional outcome and used an interpretability technique to highlight regions in the ADC maps that were most important in the prediction of a bad outcome. Although highly accurate, the model’s predictions were not based on aspects of the ADC maps related to stroke pathophysiology.

Recently, CLPB deficiency has been shown to cause a genetic syndrome with cataracts, neutropenia, and 3-methylglutaconic aciduria. Surprisingly, the neurological presentation ranges from completely unaffected to patients with virtual absence of development. Muscular hypo- and hypertonia, movement disorder and progressive brain atrophy are frequently reported. We present the foetal, peri- and neonatal features of 31 patients, of which five are previously unreported, using a newly developed clinical severity scoring system rating the clinical, metabolic, imaging and other findings weighted by the age of onset. Our data are illustrated by foetal and neonatal videos. The patients were classified as having a mild (n = 4), moderate (n = 13) or severe (n = 14) disease phenotype. The most striking feature of the severe subtype was the neonatal absence of voluntary movements in combination with ventilator dependency and hyperexcitability. The foetal and neonatal presentation mirrored the course of disease with respect to survival (current median age 17.5 years in the mild group, median age of death 35 days in the severe group), severity and age of onset of all findings evaluated. CLPB deficiency should be considered in neonates with absence of voluntary movements, respiratory insufficiency and swallowing problems, especially if associated with 3-methylglutaconic aciduria, neutropenia and cataracts. Being an important differential diagnosis of hyperekplexia (exaggerated startle responses), we advise performing urinary organic acid analysis, blood cell counts and ophthalmological examination in these patients. The neonatal presentation of CLPB deficiency predicts the course of disease in later life, which is extremely important for counselling.

Despite the widespread use of Knowledge Graph Embeddings (KGE), little is known about the security vulnerabilities that might disrupt their intended behaviour. We study data poisoning attacks against KGE models for link prediction. These attacks craft adversarial additions or deletions at training time to cause model failure at test time. To select adversarial deletions, we propose to use the model-agnostic instance attribution methods from Interpretable Machine Learning, which identify the training instances that are most influential to a neural model's predictions on test instances. We use these influential triples as adversarial deletions. We further propose a heuristic method to replace one of the two entities in each influential triple to generate adversarial additions. Our experiments show that the proposed strategies outperform the state-of-art data poisoning attacks on KGE models and improve the MRR degradation due to the attacks by up to 62% over the baselines., 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP 2021)

Background Stroke is a worldwide cause of disability; 40% of stroke survivors sustain cognitive impairments, most of them following inpatient rehabilitation at specialized clinical centers. Web-based cognitive rehabilitation tasks are extensively used in clinical settings. The impact of task execution depends on the ratio between the skills of the treated patient and the challenges imposed by the task itself. Thus, treatment personalization requires a trade-off between patients’ skills and task difficulties, which is still an open issue. In this study, we propose Elo ratings to support clinicians in tasks assignations and representing patients’ skills to optimize rehabilitation outcomes. Objective This study aims to stratify patients with ischemic stroke at an early stage of rehabilitation into three levels according to their Elo rating; to show the relationships between the Elo rating levels, task difficulty levels, and rehabilitation outcomes; and to determine if the Elo rating obtained at early stages of rehabilitation is a significant predictor of rehabilitation outcomes. Methods The PlayerRatings R library was used to obtain the Elo rating for each patient. Working memory was assessed using the DIGITS subtest of the Barcelona test, and the Rey Auditory Verbal Memory Test (RAVLT) was used to assess verbal memory. Three subtests of RAVLT were used: RAVLT learning (RAVLT075), free-recall memory (RAVLT015), and recognition (RAVLT015R). Memory predictors were identified using forward stepwise selection to add covariates to the models, which were evaluated by assessing discrimination using the area under the receiver operating characteristic curve (AUC) for logistic regressions and adjusted R2 for linear regressions. Results Three Elo levels (low, middle, and high) with the same number of patients (n=96) in each Elo group were obtained using the 50 initial task executions (from a total of 38,177) for N=288 adult patients consecutively admitted for inpatient rehabilitation in a clinical setting. The mid-Elo level showed the highest proportions of patients that improved in all four memory items: 56% (54/96) of them improved in DIGITS, 67% (64/96) in RAVLT075, 58% (56/96) in RAVLT015, and 53% (51/96) in RAVLT015R (P Conclusions Elo ratings can support clinicians in early rehabilitation stages in identifying cognitive profiles to be used for assigning task difficulty levels.

It is often the case with new technologies that it is very hard to predict their long-term impacts and as a result, although new technology may be beneficial in the short term, it can still cause problems in the longer term. This is what happened with oil by-products in different areas: the use of plastic as a disposable material did not take into account the hundreds of years necessary for its decomposition and its related long-term environmental damage. Data is said to be the new oil. The message to be conveyed is associated with its intrinsic value. But as in the case of real oil, we should take care to ensure that its use does not create harm in the future. We know from recent history that data can be used in harmful ways, but data also has enormous positive potential when applied to the service of communities. In this article, we highlight the opportunities, problems and best practice of using data.

Personalised response generation enables generating human-like responses by means of assigning the generator a social identity. However, pragmatics theory suggests that human beings adjust the way of speaking based on not only who they are but also whom they are talking to. In other words, when modelling personalised dialogues, it might be favourable if we also take the listener's social identity into consideration. To validate this idea, we use gender as a typical example of a social variable to investigate how the listener's identity influences the language used in Chinese dialogues on social media. Also, we build personalised generators. The experiment results demonstrate that the listener's identity indeed matters in the language use of responses and that the response generator can capture such differences in language use. More interestingly, by additionally modelling the listener's identity, the personalised response generator performs better in its own identity.

Personalised response generation enables generating human-like responses by means of assigning the generator a social identity. However, pragmatics theory suggests that human beings adjust the way of speaking based on not only who they are but also whom they are talking to. In other words, when modelling personalised dialogues, it might be favourable if we also take the listener's social identity into consideration. To validate this idea, we use gender as a typical example of a social variable to investigate how the listener's identity influences the language used in Chinese dialogues on social media. Also, we build personalised generators. The experiment results demonstrate that the listener's identity indeed matters in the language use of responses and that the response generator can capture such differences in language use. More interestingly, by additionally modelling the listener's identity, the personalised response generator performs better in its own identity.

Earlier research has shown that evaluation metrics based on textual similarity (e.g., BLEU, CIDEr, Meteor) do not correlate well with human evaluation scores for automatically generated text. We carried out an experiment with Chinese speakers, where we systematically manipulated image descriptions to contain different kinds of errors. Because our manipulated descriptions form minimal pairs with the reference descriptions, we are able to assess the impact of different kinds of errors on the perceived quality of the descriptions. Our results show that different kinds of errors elicit significantly different evaluation scores, even though all erroneous descriptions differ in only one character from the reference descriptions. Evaluation metrics based solely on textual similarity are unable to capture these differences, which (at least partially) explains their poor correlation with human judgments. Our work provides the foundations for future work, where we aim to understand why different errors are seen as more or less severe.

Referring expression generation (REG) algorithms offer computational models of the production of referring expressions. In earlier work, a corpus of referring expressions (REs) in Mandarin was introduced. In the present paper, we annotate this corpus, evaluate classic REG algorithms on it, and compare the results with earlier results on the evaluation of REG for English referring expressions. Next, we offer an in-depth analysis of the corpus, focusing on issues that arise from the grammar of Mandarin. We discuss shortcomings of previous REG evaluations that came to light during our investigation and we highlight some surprising results. Perhaps most strikingly, we found a much higher proportion of under-specified expressions than previous studies had suggested, not just in Mandarin but in English as well.

Referring expression generation (REG) algorithms offer computational models of the production of referring expressions. In earlier work, a corpus of referring expressions (REs) in Mandarin was introduced. In the present paper, we annotate this corpus, evaluate classic REG algorithms on it, and compare the results with earlier results on the evaluation of REG for English referring expressions. Next, we offer an in-depth analysis of the corpus, focusing on issues that arise from the grammar of Mandarin. We discuss shortcomings of previous REG evaluations that came to light during our investigation and we highlight some surprising results. Perhaps most strikingly, we found a much higher proportion of under-specified expressions than previous studies had suggested, not just in Mandarin but in English as well.

Earlier research has shown that evaluation metrics based on textual similarity (e.g., BLEU, CIDEr, Meteor) do not correlate well with human evaluation scores for automatically generated text. We carried out an experiment with Chinese speakers, where we systematically manipulated image descriptions to contain different kinds of errors. Because our manipulated descriptions form minimal pairs with the reference descriptions, we are able to assess the impact of different kinds of errors on the perceived quality of the descriptions. Our results show that different kinds of errors elicit significantly different evaluation scores, even though all erroneous descriptions differ in only one character from the reference descriptions. Evaluation metrics based solely on textual similarity are unable to capture these differences, which (at least partially) explains their poor correlation with human judgments. Our work provides the foundations for future work, where we aim to understand why different errors are seen as more or less severe.

This is a resource description paper that describes the creation and properties of a set of pseudo-corpora generated artificially from a random walk over the English WordNet taxonomy. Our WordNet taxonomic random walk implementation allows the exploration of different random walk hyperparameters and the generation of a variety of different pseudo-corpora. We find that different combinations of the walk’s hyperparameters result in varying statistical properties of the generated pseudo-corpora. We have published a total of 81 pseudo-corpora that we have used in our previous research, but have not exhausted all possible combinations of hyperparameters, which is why we have also published a codebase that allows the generation of additional WordNet taxonomic pseudo-corpora as needed. Ultimately, such pseudo-corpora can be used to train taxonomic word embeddings, as a way of transferring taxonomic knowledge into a word embedding space.

Supplemental material, sj-pdf-1-bds-10.1177_2053951720915939 for Expectations of artificial intelligence and the performativity of ethics: Implications for communication governance by Aphra Kerr, Marguerite Barry and John D Kelleher in Big Data & Society

Corporate smart city initiatives are just one example of the contemporary culture of surveillance. They rely on extensive information gathering systems and Big Data analysis topredict citizen behaviour and optimise city services. In this paper we argue that many smart city and social media technologies result in a paradox whereby digital inclusion for the purposes of service provision also results in marginalisation and disempowerment of citizens. Drawing upon insights garnered from a digital inclusion workshop conducted in the Galapagos islands, we propose that critically and creatively unpacking the computational techniques embedded in data services is needed as a first step if we are to reimagine neganthropic, sustainable and empowering data services for inhabitants in diverse localities. We propose a therapeutic inspired by the concept of ‘common ground’ from communication theory. Common ground presupposes a symmetry of purpose, shared values and accessible participation processes. When common ground is deployed in the smart city context it prompts us to reimagine data services as an ongoing dialogue between peers, to rethink citizen participation in terms of capabilities and empowerment, and to focus on clear lines ofaccountability and equality of citizen outcomes.

We present an approach to design the grid searches for hyper-parameter optimization for recurrent neural architectures. The basis for this approach is the use of mutual information to analyze long distance dependencies (LDDs) within a dataset. We also report a set of experiments that demonstrate how using this approach, we obtain state-of-the-art results for DilatedRNNs across a range of benchmark datasets.

This paper connects a series of papers dealing with taxonomic word embeddings. It begins by noting that there are different types of semantic relatedness and that different lexical representations encode different forms of relatedness. A particularly important distinction within semantic relatedness is that of thematic versus taxonomic relatedness. Next, we present a number of experiments that analyse taxonomic embeddings that have been trained on a synthetic corpus that has been generated via a random walk over a taxonomy. These experiments demonstrate how the properties of the synthetic corpus, such as the percentage of rare words, are affected by the shape of the knowledge graph the corpus is generated from. Finally, we explore the interactions between the relative sizes of natural and synthetic corpora on the performance of embeddings when taxonomic and thematic embeddings are combined., Comment: 7 pages 0 figures

Brain vessel status is a promising biomarker for better prevention and treatment in cerebrovascular disease. However, classic rule-based vessel segmentation algorithms need to be hand-crafted and are insufficiently validated. A specialized deep learning method—the U-net—is a promising alternative. Using labeled data from 66 patients with cerebrovascular disease, the U-net framework was optimized and evaluated with three metrics: Dice coefficient, 95% Hausdorff distance (95HD) and average Hausdorff distance (AVD). The model performance was compared with the traditional segmentation method of graph-cuts. Training and reconstruction was performed using 2D patches. A full and a reduced architecture with less parameters were trained. We performed both quantitative and qualitative analyses. The U-net models yielded high performance for both the full and the reduced architecture: A Dice value of ~0.88, a 95HD of ~47 voxels and an AVD of ~0.4 voxels. The visual analysis revealed excellent performance in large vessels and sufficient performance in small vessels. Pathologies like cortical laminar necrosis and a rete mirabile led to limited segmentation performance in few patients. The U-net outperfomed the traditional graph-cuts method (Dice ~0.76, 95HD ~59, AVD ~1.97). Our work highly encourages the development of clinically applicable segmentation tools based on deep learning. Future works should focus on improved segmentation of small vessels and methodologies to deal with specific pathologies

From theoretical linguistic and cognitive perspectives, situated dialog systems are interesting as they provide ideal test-beds for investigating the interaction between language and perception. At the same time there are a growing number of practical applications, for example robotic systems and driver-less cars, where spoken interfaces, capable of situated dialog, promise many advantages. To date, however much of the work on situated dialog has focused resolving anaphoric or exophoric references. This paper, by contrast, opens up the question of how perceptual memory and linguistic references interact, and the challenges that this poses to computational models of perceptually grounded dialog., Comment: 18 Pages, 4 Figures

Creating word embeddings that reflect semantic relationships encoded in lexical knowledge resources is an open challenge. One approach is to use a random walk over a knowledge graph to generate a pseudo-corpus and use this corpus to train embeddings. However, the effect of the shape of the knowledge graph on the generated pseudo-corpora, and on the resulting word embeddings, has not been studied. To explore this, we use English WordNet, constrained to the taxonomic (tree-like) portion of the graph, as a case study. We investigate the properties of the generated pseudo-corpora, and their impact on the resulting embeddings. We find that the distributions in the psuedo-corpora exhibit properties found in natural corpora, such as Zipf’s and Heaps’ law, and also ob- serve that the proportion of rare words in a pseudo-corpus affects the performance of its embeddings on word similarity.

Dialogue State Tracking is arguably one of the most challenging tasks among dialogue processing problems due to the uncertainties of language and complexity of dialogue contexts. We argue that this problem is made more challenging by variable dependencies in the dialogue states that must be accounted for in processing. In this paper we give details on our motivation for this argument through statistical tests on a number of dialogue datasets. We also propose a machine learning-based approach called energy-based learning that tackles variable dependencies while performing prediction on the dialogue state tracking tasks.

Image Captioning is a task that requires models to acquire a multi-modal understanding of the world and to express this understanding in natural language text. While the state-of-the-art for this task has rapidly improved in terms of n-gram metrics, these models tend to output the same generic captions for similar images. In this work, we address this limitation and train a model that generates more diverse and specific captions through an unsupervised training approach that incorporates a learning signal from an Image Retrieval model. We summarize previous results and improve the state-of-the-art on caption diversity and novelty. We make our source code publicly available online., Accepted for presentation at The 27th International Conference on Artificial Neural Networks (ICANN 2018)

Incrementality is a fundamental feature of language in real world use. To this point, however, the vast majority of work in automated dialogue processing has focused on language as turn based. In this paper we explore the challenge of incremental dialogue state tracking through the development and analysis of a multi-task approach to incremental dialogue state tracking. We present the design of our incremental dialogue state tracker in detail and provide evaluation against the well known Dialogue State Tracking Challenge 2 (DSTC2) dataset. In addition to a standard evaluation of the tracker, we also provide an analysis of the Incrementality phenomenon in our model’s performance by analyzing how early our models can produce correct predictions and how stable those predictions are. We find that the Multi-Task Learning-based model achieves state-of-the-art results for incremental processing.

In order to build efficient deep recurrent neural architectures, it is essential to analyze the complexityof long distance dependencies (LDDs) of the dataset being modeled. In this paper, we presentdetailed analysis of the dependency decay curve exhibited by various datasets. The datasets sampledfrom a similar process (e.g. natural language, sequential MNIST, Strictlyk-Piecewise languages,etc) display variations in the properties of the dependency decay curve. Our analysis reveal thefactors resulting in these variations; such as (i) number of unique symbols in a dataset, (ii) size ofthe dataset, (iii) number of interacting symbols within a given LDD, and (iv) the distance betweenthe interacting symbols. We test these factors by generating synthesized datasets of the Strictlyk-Piecewise languages. Another advantage of these synthesized datasets is that they enable targetedtesting of deep recurrent neural architectures in terms of their ability to model LDDs with differentcharacteristics. We also demonstrate that analysing dependency decay curves can inform the selectionof optimal hyper-parameters for SOTA deep recurrent neural architectures. This analysis can directlycontribute to the development of more accurate and efficient sequential models.

Natural language processing (NLP) can be done using either top-down (theory driven) and bottom-up (data driven) approaches, which we call mechanistic and phenomenological respectively. The approaches are frequently considered to stand in opposition to each other. Examining some recent approaches in deep learning we argue that deep neural networks incorporate both perspectives and, furthermore, that leveraging this aspect of deep learning may help in solving complex problems within language technology, such as modelling language and perception in the domain of spatial cognition., 18 pages, 1 figure, Appears in CLASP Papers in Computational Linguistics Vol. 1: Proceedings of the Conference on Logic and Machine Learning in Natural Language (LaML 2017)

An urgent limitation in current Image Captioning models is their tendency to produce generic captions that avoid the interesting detail which makes each image unique. To address this limitation, we propose an approach that enforces a stronger alignment between image regions and specific segments of text. The model architecture is composed of a visual region proposer, a region-order planner and a region-guided caption generator. The region-guided caption generator incorporates a novel information gate which allows visual and textual input of different frequencies and dimensionalities in a Recurrent Neural Network.

We examine the non-Markovian nature of human mobility by exposing the inability of Markov models to capture criticality in human mobility. In particular, the assumed Markovian nature of mobility was used to establish an upper bound on the predictability of human mobility, based on the temporal entropy. Since its inception, this bound has been widely used for validating the performance of mobility prediction models. We show that the variants of recurrent neural network architectures can achieve significantly higher prediction accuracy surpassing this upper bound. The central objective of our work is to show that human-mobility dynamics exhibit criticality characteristics which contributes to this discrepancy. In order to explain this anomaly, we shed light on the underlying assumption that human mobility characteristics follow an exponential decay that has resulted in this bias. By evaluating the predictability on real-world datasets, we show that human mobility exhibits scale-invariant long-distance dependencies, bearing resemblance to power-law decay, contrasting with the initial Markovian assumption. We experimentally validate that this assumption inflates the estimated mobility entropy, consequently lowering the upper bound on predictability. We demonstrate that the existing approach of entropy computation tends to overlook the presence of long-distance dependencies and structural correlations in human mobility. We justify why recurrent-neural network architectures that are designed to handle long-distance dependencies surpass the previously computed upper bound on mobility predictability.

This article revisits the question of ‘la bêtise’ or stupidity in the era of Artificial Intelligence driven by Big Data, it extends on the questions posed by Gille Deleuze and more recently by Bernard Stiegler. However, the framework for revisiting the question of la bêtise will be through the lens of contemporary computer science, in particular the development of data science as a mode of analysis, sometimes, misinterpreted as a mode of intelligence. In particular, this article will argue that with the advent of forms of hype (sometimes referred to as the hype cycle) in relation to big data and modalities of data analytics there is a form of computational stupidity or functional stupidity at work. The exaggerated promises of big data to solve everything are overblown expectations which will lead ultimately to a form of disillusionment with data science. This can be seen in a number of domains, for example smart city technologies, the internet of things, and machine translation. In addition to the negative effects of exaggerated claims of Big Data is the possibility that societal norms will facilitate Big Data technological change by incorporating the bêtise of Big Data, thus leading to a change in our relationship to technology, examples of this would be privacy standards and ownership of data. This paper will conclude by setting out the analysis some of the limitations of Artificial intelligence and Big Data in order to allow a re-examination of the claims made.

Natural language processing (NLP) can be done using either top-down (theory driven) and bottom-up (data driven) approaches, which we call mechanistic and phenomenological respectively. The approaches are frequently considered to stand in opposition to each other. Examining some recent approaches in deep learning we argue that deep neural networks incorporate both perspectives and, furthermore, that leveraging this aspect of deep learning may help in solving complex problems within language technology, such as modelling language and perception in the domain of spatial cognition.

This paper examines to what degree current deep learning architectures for image caption generation capture spatial lan- guage. On the basis of the evaluation of examples of generated captions from the literature we argue that systems capture what objects are in the image data but not where these objects are located: the cap- tions generated by these systems are the output of a language model conditioned on the output of an object detector that cannot capture fine-grained location information. Although language models provide useful knowledge for image captions, we argue that deep learning image captioning architectures should also model geometric rela- tions between objects.

Dialogue State Tracking is an important task in dialogue management as it provides a mechanism to monitor dialogue contributions. In this paper we introduce an Incremental Joint Model as a new approach to the task. Our tracker is capable of incrementally tracking Dialogue States. We base our model and analysis on the datasets provided in the Second Dialogue State Tracking Challenge (DSTC2). Our early stage evaluations are based on comparisons of our tracker with both the baseline model provide by the DSTC2 and also LecTrack: a state-of-the-art incremental LSTM-based tracker. The main finding of our experiments is that moving from an utterance based to incremental word based tracker results in better performance for our RNN based joint task models.

The aim of this paper is to present techniques and results for identifying the key of Irish traditional music melodies, or tunes. Several corpora are used, consisting of both symbolic and audio representations. Monophonic and heterophonic recordings are present in the audio datasets. Some particularities of Irish traditional music are discussed, notably its modal nature. New key-profiles are defined, that are better suited to Irish music.

In this paper we argue that since the beginning of the natural language processing or computational linguistics there has been a strong connection between logic and machine learning. First of all, there is something logical about language or linguistic about logic. Secondly, we argue that rather than distinguishing between logic and machine learning, a more useful distinction is between top-down approaches and data-driven approaches. Examining some recent approaches in deep learning we argue that they incorporate both properties and this is the reason for their very successful adoption to solve several problems within language technology.

This paper presents an interactive visualisation tool that encodes the scope chain, and information related to the scope chain, within source code. The main goal of the tool is to support programmers when dealing with issues related to scope and to provide answers to questions such as to which scope does a specific variable or function belong to and can I access a specific variable from the scope I am currently located in. The design guidelines followed during the implementation of the tool, as well as the design rationale behind the main features of the tool are described. Finally, the results of a pilot user experience evaluation study are presented where an interesting observation was that the tool seemed to support programmers in verifying and correcting their assumptions when asked questions about specific scoping issues within a source code document.

This paper presents a short introduction to neural networks and how they are used for machine translation and concludes with some discussion on the current research challenges being addressed by neural machine translation (NMT) research. The primary goal of this paper is to give a no-tears introduction to NMT to readers that do not have a computer science or mathematical background. The secondary goal is to provide the reader with a deep enough understanding of NMT that they can appreciate the strengths of weaknesses of the technology. The paper starts with a brief introduction to standard feed-forward neural networks (what they are, how they work, and how they are trained), this is followed by an introduction to word-embeddings (vector representations of words) and then we introduce recurrent neural networks. Once these fundamentals have been introduced we then focus in on the components of a standard neural-machine translation architecture, namely: encoder networks, decoder language models, and the encoder-decoder architecture.

An emerging trend in music information retrieval (MIR) is the use of supervised machine learning to train automatic music transcription models. A prerequisite of adopting a machine learning methodology is the availability of annotated corpora. However, different genres of music have different characteristics and modelling these characteristics is an important part of creating state of the art MIR systems. Consequently, although some music corpora are available the use of these corpora is tied to the specific music genre, instrument type and recording context the corpus covers. This paper introduces the first corpus of annotations of audio recordings of Irish traditional dance music that covers multiple instrument types and both solo studio and live session recordings. We first discuss the considerations that motivated our design choices in developing the corpus. We then benchmark a number of automatic music transcription algorithms against the corpus. The underlying dataset for this research is available here at Github or here in Arrow .