Word embedding-based similarity measures are currently among the top-performing methods on unsupervised semantic textual similarity (STS) tasks.Recent work has increasingly adopted a statistical view on these embeddings, with some of the top approaches being essentially various correlations (which include the famous cosine similarity).Another excellent candidate for a similarity measure is mutual information (MI), which can capture arbitrary dependencies between the variables and has a simple and intuitive expression.Unfortunately, its use in the context of dense word embeddings has so far been avoided due to difficulties with estimating MI for continuous data.In this work we go through a vast literature on estimating MI in such cases and single out the most promising methods, yielding a simple and elegant similarity measure for word embeddings.We show that mutual information is a viable alternative to correlations, gives an excellent signal that correlates well with human judgements of similarity and rivals existing state-of-the-art unsupervised methods.

A promising application of AI to healthcare is the retrieval of information from electronic health records (EHRs), e.g. to aid clinicians in finding relevant information for a consultation or to recruit suitable patients for a study.This requires search capabilities far beyond simple string matching, including the retrieval of concepts (diagnoses, symptoms, medications, etc.) related to the one in question.The suitability of AI methods for such applications is tested by predicting the relatedness of concepts with known relatedness scores.However, all existing biomedical concept relatedness datasets are notoriously small and consist of hand-picked concept pairs.We opensource a novel concept relatedness benchmark overcoming these issues: it is six times larger than existing datasets and concept pairs are chosen based on co-occurrence in EHRs, ensuring their relevance for the application of interest.We present an in-depth analysis of our new dataset and compare it to existing ones, highlighting that it is not only larger but also complements existing datasets in terms of the types of concepts included.Initial experiments with state-of-the-art embedding methods show that our dataset is a challenging new benchmark for testing concept relatedness models.

In this work we approach the task of learning multilingual word representations in an offline manner by fitting a generative latent variable model to a multilingual dictionary.We model equivalent words in different languages as different views of the same word generated by a common latent variable representing their latent lexical meaning.We explore the task of alignment by querying the fitted model for multilingual embeddings achieving competitive results across a variety of tasks.The proposed model is robust to noise in the embedding space making it a suitable method for distributed representations learned from noisy corpora.

Vitalii Zhelezniak, April Shen, Daniel Busbridge, Aleksandar Savkov, Nils Hammerla. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.

We introduce a probabilistic framework for quantifying the semantic similarity between two groups of embeddings. We formulate the task of semantic similarity as a model comparison task in which we contrast a generative model which jointly models two sentences versus one that does not. We illustrate how this framework can be used for the Semantic Textual Similarity tasks using clear assumptions about how the embeddings of words are generated. We apply model comparison that utilises information criteria to address some of the shortcomings of Bayesian model comparison, whilst still penalising model complexity. We achieve competitive results by applying the proposed framework with an appropriate choice of likelihood on the STS datasets.

Expressiveness, which we define as the extent to which rich and complex intent can be conveyed through action, is a vital aspect of many human interactions. For instance, paint on canvas is said to be an expressive medium, because it affords the artist the ability to convey multifaceted emotional intent through intricate manipulations of a brush. To date, touch devices have failed to offer users a level of expressiveness in their interactions that rivals that experienced by the painter and those completing other skilled physical tasks. We investigate how data about hand movement -- provided by a motion sensor, similar to those found in many smart watches or fitness trackers -- can be used to expand the expressiveness of touch interactions. We begin by introducing a conceptual model that formalizes a design space of possible expressive touch interactions. We then describe and evaluate Expressy, an approach that uses a wrist-worn inertial measurement unit to detect and classify qualities of touch interaction that extend beyond those offered by today's typical sensing hardware. We conclude by describing a number of sample applications, which demonstrate the enhanced, expressive interaction capabilities made possible by Expressy.

Management of Parkinson's Disease (PD) could be improved significantly if reliable, objective information about fluctuations in disease severity can be obtained in ecologically valid surroundings such as the private home. Although automatic assessment in PD has been studied extensively, so far no approach has been devised that is useful for clinical practice. Analysis approaches common for the field lack the capability of exploiting data from realistic environments, which represents a major barrier towards practical assessment systems. The very unreliable and infrequent labelling of ambiguous, low resolution movement data collected in such environments represents a very challenging analysis setting, where advances would have significant societal impact in our ageing population. In this work we propose an assessment system that abides practical usability constraints and applies deep learning to differentiate disease state in data collected in naturalistic settings. Based on a large data-set collected from 34 people with PD we illustrate that deep learning outperforms other approaches in generalisation performance, despite the unreliable labelling characteristic for this problem setting, and how such systems could improve current clinical practice.

The ability to generalise towards either new users or unforeseen behaviours is a key requirement for activity recognition systems in ubiquitous computing. Differences in recognition performance for the two application cases can be significant, and user-dependent performance is typically assumed to be an upper bound on performance. We demonstrate that this assumption does not hold for the widely used cross-validation evaluation scheme that is typically employed both during system bootstrapping and for reporting results. We describe how the characteristics of segmented time-series data render random cross-validation a poor fit, as adjacent segments are not statistically independent. We develop an alternative approach -- meta-segmented cross validation -- that explicitly circumvents this issue and evaluate it on two data-sets. Results indicate a significant drop in performance across a variety of feature extraction and classification methods if this bias is removed, and that prolonged, repetitive activities are particularly affected.

Research on actuated interfaces has shown that people respond in certain socialized ways to interfaces that exhibit autonomous behaviours. We wished to explore the elements of design that drive people to regard an autonomous, interactive system as a social agent. To explore perceptions of autonomous behaviour in interfaces we created Diri - an autonomous helium balloon, used to document activity in spaces. We implemented two different technological sophistications of Diri, to compare the outcomes of our design decisions. We present our design process, technical details and evaluation workshops, concluding with implications for designing for autonomous behaviour in interfaces.

We present a novel approach to graspable interfaces using Touchbugs, actuated physical objects for interacting with interactive surface computing applications. Touchbugs are active tangibles that are able to move across surfaces by employing vibrating motors and can communicate with camera based multi-touch surfaces using infrared LEDs. Touchbug's embedded inertial sensors and computational capabilities open a new interaction space by providing autonomous capabilities for tangibles that allow goal directed behavior.

Health and well-being of dogs, either domesticated pets or service animals, are major concerns that are taken seriously for ethical, emotional, and financial reasons. Welfare assessments in dogs rely on objective observations of both frequency and variability of individual behaviour traits, which is often difficult to obtain in a dog's everyday life. In this paper we have identified a set of activities, which are linked to behaviour traits that are relevant for a dog's wellbeing. We developed a collar-worn accelerometry platform that records dog behaviours in naturalistic environments. A statistical classification framework is used for recognising dog activities. In an experimental evaluation we analysed the naturalistic behaviour of 18 dogs and were able to recognise a total of 17 different activities with approximately 70% classification accuracy. The presented system is the first of its kind that allows for robust and detailed analysis of dog activities in naturalistic environments.

In recent years the sport of climbing has seen consistent increase in popularity. Climbing requires a complex skill set for successful and safe exercising. While elite climbers receive intensive expert coaching to refine this skill set, this progression approach is not viable for the amateur population. We have developed ClimbAX - a climbing performance analysis system that aims for replicating expert assessments and thus represents a first step towards an automatic coaching system for climbing enthusiasts. Through an accelerometer based wearable sensing platform, climber's movements are captured. An automatic analysis procedure detects climbing sessions and moves, which form the basis for subsequent performance assessment. The assessment parameters are derived from sports science literature and include: power, control, stability, speed. ClimbAX was evaluated in a large case study with 53 climbers under competition settings. We report a strong correlation between predicted scores and official competition results, which demonstrate the effectiveness of our automatic skill assessment system.

Severe behavior problems of children with developmental disabilities often require intervention by specialists. These specialists rely on direct observation of the behavior, usually in a controlled clinical environment. In this paper, we present a technique for using on-body accelerometers to assist in automated classification of problem behavior during such direct observation. Using simulated data of episodes of severe behavior acted out by trained specialists, we demonstrate how machine learning techniques can be used to segment relevant behavioral episodes from a continuous sensor stream and to classify them into distinct categories of severe behavior (aggression, disruption, and self-injury). We further validate our approach by demonstrating it produces no false positives when applied to a publicly accessible dataset of activities of daily living. Finally, we show promising classification results when our sensing and analysis system is applied to data from a real assessment session conducted with a child exhibiting problem behaviors.

First Person Shooter (FPS) is a popular genre in online gaming, unfortunately not everyone plays the game fairly, and this hinders the growth of the industry. The aiming robot (aimbot) is a common cheating mechanism employed in this genre, it differs from many other common online bots in that there is a human operating alongside the bot, and thus the in-game data exhibit both human and bot-like behaviour. The aimbot users can aim much better than the average player. However, there are also a large number of highly skilled players who can aim much better than the average player, some of these players have in the past been banned from servers due to false accusations from their peers. Therefore, it would be interesting to find out if and where the honest player's and the bot user's behaviour differ. In this paper we investigate the difference between the aiming abilities of aimbot users and honest human players. We introduce two novel features and have conducted an experiment using a modified open source FPS game. Our data shows that there is significant difference between behaviours of honest players and aimbot users. We propose a voting scheme to improve aimbot detection in FPS based on distribution matching, and have achieved approximately 93% in both True positive and True negative rates with one of our features.

We present GymSkill, a personal trainer for ubiquitous monitoring and assessment of physical activity using standard fitness equipment. The system records and analyzes exercises using the sensors of a personal smartphone attached to the gym equipment. Novel fine-grained activity recognition techniques based on pyramidal Principal Component Breakdown Analysis (PCBA) provide a quantitative analysis of the quality of human movements. In addition to overall quality judgments, GymSkill identifies interesting portions of the recorded sensor data and provides suggestions for improving the individual performance, thereby extending existing work. The system was evaluated in a case study where 6 participants performed a variety of exercises on balance boards. GymSkill successfully assessed the quality of the exercises, in agreement with the professional judgment provided by a physician. User feedback suggests that GymSkill has the potential to serve as an effective tool for motivating and supporting lay people to overcome sedentary, unhealthy lifestyles. GymSkill is available in the Android Market as `VMI Fit'.

We present the development of a socially acceptable cueing device for drooling in Parkinson's disease (PD). Sialorrhea, or drooling, is a significant problem associated with PD and has a strong negative emotional impact on those who experience it. Previous studies have shown the potential for managing drooling by using a cueing device. However, the devices used in these studies were deemed unacceptable by their users due to factors such as hearing impairment and social embarrassment. We conducted exploratory scoping work and high fidelity iterative prototyping with people with PD to get their input on the design of a cueing aid and this has given us an insight into challenges that confront users with PD and limit device usability and acceptability. The key finding from working with people with PD was the need for the device to be socially acceptable.