IBM (Ireland)

Neural machine translation (NMT) is often described as ‘data hungry’ as it typically requires large amounts of parallel data in order to build a good-quality machine translation (MT) system. However, most of the world's language-pairs are low-resource or extremely low-resource. This situation becomes even worse if a specialised domain is taken into consideration for translation. In this paper, we present a novel data augmentation method which makes use of bilingual word embeddings (BWEs) learned from monolingual corpora and bidirectional encoder representations from transformer (BERT) language models (LMs). We augment a parallel training corpus by introducing new words (i.e. out-of-vocabulary (OOV) items) and increasing the presence of rare words on both sides of the original parallel training corpus. Our experiments on the simulated low-resource German–English and French–English translation tasks show that the proposed data augmentation strategy can significantly improve state-of-the-art NMT systems and outperform the state-of-the-art data augmentation approach for low-resource NMT.

Posting information on social media platforms is a popular activity through which personal and confidential information can leak into the public domain. Consequently, social media can contain information that provides an indication that an organization has been compromised or suffered a data breach. This paper describes a technique for inferring if an organization has been compromised from information posted on social media. The proposed strategy forms the basis of an alarm system which generates an alert for possible unreported cybercrime incidents. The proposed strategy used two social media cybercrime related datasets that were collected from the Irish and New York regions from financial organizations’ Twitter accounts. The Tweets are labelled as either containing cybercrime indicators or not, and then the cybercrime Tweets were labelled further into crime categories. A deep dense pyramidal Neural Network model is used to classify the Tweets. This approach achieves an AUC of \(~0.85 \pm 0.03\) which outperforms the baseline of deep convolutional neural networks.

In this paper we propose a new approach to complement reinforcement learning (RL) with model-based control (in particular, Model Predictive Control - MPC). We introduce an algorithm, the MPC augmented RL (MPRL) that combines RL and MPC in a novel way so that they can augment each other’s strengths. We demonstrate the effectiveness of the MPRL by letting it play against the Atari game Pong. For this task, the results highlight how MPRL is able to outperform both RL and MPC when these are used individually.

Informative missingness is unavoidable in the digital processing of continuous time series, where the value for one or more observations at different time points are missing. Such missing observations are one of the major limitations of time series processing using deep learning. Practical applications, e.g., sensor data, healthcare, weather, generates data that is in truth continuous in time, and informative missingness is a common phenomenon in these datasets. These datasets often consist of multiple variables, and often there are missing values for one or many of these variables. This characteristic makes time series prediction more challenging, and the impact of missing input observations on the accuracy of the final output can be significant. A recent novel deep learning model called GRU-D is one early attempt to address informative missingness in time series data. On the other hand, a new family of neural networks called Neural ODEs (Ordinary Differential Equations) are natural and efficient for processing time series data which is continuous in time. In this paper, a deep learning model is proposed that leverages the effective imputation of GRU-D, and the temporal continuity of Neural ODEs. A time series classification task performed on the PhysioNet dataset demonstrates the performance of this architecture.

IBM Research in partnership with IBM Services developed an innovative set of tools that automatically and cognitively reads a request for proposal, extracts client requirements, and maps them to IBM offerings. It also performs accurate costing and recosting, pricing and repricing, and market benchmarking of the bid, and it predicts the status over time of the various deals being pursued to effectively manage IBM’s sales pipeline and align salesforce resources.

Traditional multimodal journey planners are deterministic. However, uncertainty in a transportation network can often lead to missed connections, causing delays and hurting the reliability level of the service. This article provides an overview of what is possibly the first multimodal journey advising system that is capable of reasoning under uncertainty and that provides more reliable journey plans.

The development of a city gradually forms different functional regions, such as residential districts and shopping areas. Discovering these functional regions in cities can enable new types of valuable applications that can benefit different end users: Urban planners can better identify the proximity of existing functional regions and hence, can contribute a better future planning for the cities. Tourists can differentiate scenic areas from other business and residential areas which will help in reducing effort for trip planning. Moreover, local people can better understand each part of their cities by finding areas with particular functionality. With the rise of Location-Based Social Networks (LBSNs) which attract lots of new users everyday with the potential of bridging the gap between the physical world and digital online social network services, we show in this paper that identifying functional regions taking into account temporal variations of geographic user activity has become possible and is more sensible when identifying functional regions. In this work, we propose a novel approach to modeling functional areas taking into account temporal variation by means of place categories. Our proposed approach compares between three clustering algorithms (Hierarchical, K-means, and Spectral) on areas and users of Manhattan borough in New York City using a dataset from one of the most vibrant LBSN, Foursquare. We demonstrate the impact of different temporal variations splits on the quality of the clustering algorithms comparing it to the default approach with no temporal variation. We believe that this research can not only yield a deeper understanding of a complex city but also can offer finer personalized recommendations based on regions' functionality that changes over space and time.

During the past few years, the analysis of data generated from Location-Based Social Networks (LBSNs) have aided in the identification of urban patterns, understanding activity behaviours in urban areas, as well as producing novel recommender systems that facilitate users' choices. However, the recent advancement in machine learning techniques promises new deeper insights with the possibility of finding new spatio-temporal patterns in cities. In this paper, we show that one of the recent advancements in machine learning, Deep Belief Networks (DBNs), can discover a new type of pattern, which we refer to in the paper as the Socio-demographic Regional Pattern. This pattern illustrates the ability of predicting the district of a city given a set of weekly activities captured from LBSNs. Specifically, we have found instances of this embedded pattern for the boroughs in New York City by training a DBN model that can classify with nearly 70% accuracy the location of weekly region-footprints. We further validated the existence and complexity of this type of pattern by applying a probabilistic topic model, namely Latent Dirichlet Allocation (LDA). We believe that this research can yield to a deeper understanding about social commonalities and the geographical evolution of different regions and areas, between cities across the globe.