Automated Essay Scoring (AES) presents a key opportunity to improve student experience while reducing the administrative burden of academic staff. However existing methods for AES are reliant on large volumes of data and fail to consider the ordinal aspect of grading. As a result, when an institution introduces a new assessment, there may be no data available to train algorithms. In this paper, we demonstrate that metric learning architectures, specifically Prototypical Networks, offer robust performance on few-shot ordinal classification essay grading tasks. We introduce three novel weighted prototype calculation strategies designed to enhance class representation in ordinal few-shot text classification. These strategies improve how class knowledge is modeled from limited examples by refining the way prototypes are computed, incorporating weighted mechanisms for better differentiation. Results across four datasets show that our methods outperform existing baselines and the current state-of-the-art in ordinal few-shot text classification. Additionally, we compare our approach with three large language models (LLMs) using a prompt-based approach to few-shot learning and find that we achieve superior or comparable performance in all evaluated tasks.

WiFi Channel State Information (CSI) is widely-used in research for human sensing applications, yet its actual deployment in commercial real-time applications remains sparse with few examples. Existing demonstrations in research literature predominantly rely on specialised deployments of a single sensing apparatus, which cannot efficiently be used in large-scale deployments. Additionally packet loss is common which leads to an over-reliance on interpolation for missing points. Addressing these gaps, this paper presents a low-cost, and scalable solution for CSI-based human sensing, tailored for high performance and consistent operation in residential environments.Our approach leverages ESP32 hardware which is renowned for its high availability and low-cost compared to popular CSI collection solutions. We define a methodology for remotely collecting CSI data from multiple sensors concurrently over WiFi, by employing a single beacon for traffic generation while CSI data is gathered over a separate channel. We further optimize this process using DEFLATE compression on CSI payloads to minimize airtime contention during transmission. This proposed system has been evaluated through a series of experiments designed to assess its viability, scalability, and environmental adaptation capability. Notably, we demonstrate the system’s capability to support 30 sensors sampling CSI data at over 90Hz simultaneously, with additional projected capacity. This validation has been conducted across two distinct residential environments, affirming the adaptability and effectiveness of our approach for high-performance CSI sensing in real-world scenarios.

Data-to-Text Generation (D2T) problems can be considered as a stream of time-stamped events with a text summary being produced for each. The problem becomes more challenging when event summaries contain complex insights derived from multiple records either within an event, or across several events from the event stream. It is important to understand the different types of content present in the summary to help us better define the system requirements so that we can build better systems. In this paper, we propose a novel typology of content types, that we use to classify the contents of event summaries. Using the typology, a profile of a dataset is generated as the distribution of the aggregated content types which captures the specific characteristics of the dataset and gives a measure of the complexity present in the problem. Extensive experimentation on different D2T datasets is performed and these demonstrate that neural systems struggle in generating contents of complex types.

Abstract Theoretical frameworks in psychology map the relationships between emotions and sentiments. In this paper, we study the role of such mapping for computational emotion detection from text (e.g., social media) with an aim to understand the usefulness of an emotion‐rich corpus of documents (e.g., tweets) to learn polarity lexicons for sentiment analysis. We propose two different methods that leverage a corpus of emotion‐labelled tweets to learn word‐polarity lexicons. The proposed methods model the emotion corpus using a generative unigram mixture model, combined with the emotion‐sentiment mapping proposed in psychology for automated generation of word‐polarity lexicons that capture emotion‐rich vocabulary. We comparatively evaluate the quality of the proposed mixture model in learning emotion‐aware sentiment lexicons with those generated using supervised latent dirichlet allocation (sLDA) and word‐document‐frequency (WDF) statistics. Sentiment analysis experiments on benchmark Twitter data sets confirm the quality of our proposed lexicons. Further, a comparative analysis with sLDA, WDF‐based emotion‐aware lexicons, and standard sentiment lexicons that are agnostic to emotion knowledge suggests that the proposed lexicons lead to a significantly better performance in both sentiment classification and sentiment intensity prediction tasks.

Abstract There is currently a large amount of e‐Learning resources available to learners on the Web. However, learners often have difficulty finding and retrieving relevant materials to support their learning goals because they lack the domain knowledge to craft effective queries that convey what they wish to learn. In addition, the unfamiliar vocabulary often used by domain experts makes it difficult to map a learner's query to a relevant learning material. We address these challenges by introducing an innovative method that automatically builds background knowledge for a learning domain. In creating our method, we exploit a structured collection of teaching materials as a guide for identifying the important domain concepts. We enrich the identified concepts with discovered text from an encyclopedia, thereby increasing the richness of our acquired knowledge. We employ the developed background knowledge for influencing the representation and retrieval of learning resources to improve e‐Learning recommendation. The effectiveness of our method is evaluated using a collection of Machine Learning and Data Mining papers. Our method outperforms the benchmark, demonstrating the advantage of using background knowledge for improving the representation and recommendation of e‐Learning materials.

Text classification is one of the most important tasks in the field of Natural Language Processing. There are many approaches that focus on two main aspects: generating an effective representation; and selecting and refining algorithms to build the classification model. Traditional machine learning methods represent documents in vector space using features such as term frequencies, which have limitations in handling the order and semantics of words. Meanwhile, although achieving many successes, deep learning classifiers require substantial resources in terms of labelled data and computational complexity. In this work, a weighted ensemble of classifiers (WEC) is introduced to address the text classification problem. Instead of using majority vote as the combining method, we propose to associate each classifier's prediction with a different weight when combining classifiers. The optimal weights are obtained by minimising a loss function on the training data with the Particle Swarm Optimisation algorithm. We conducted experiments on 5 popular datasets and report classification performance of algorithms with classification accuracy and macro F1 score. WEC was run with several different combinations of traditional machine learning and deep learning classifiers to show its flexibility and robustness. Experimental results confirm the advantage of WEC, especially on smaller datasets.

Ambient, non-intrusive approaches to smart home health monitoring, while limited in capability, are preferred by residents. More intrusive methods of sensing, such as video and wearables, can offer richer data but at the cost of lower resident uptake, in part due to privacy concerns. A radio frequency-based approach to sensing, Channel State Information (CSI), can make use of low cost off-the-shelf WiFi hardware. We have implemented an activity recognition system on the Raspberry Pi 4, one of the world's most popular embedded boards. We have implemented an classification system using the Pi to demonstrate its capability for activity recognition. This involves performing data collection, interpretation and windowing, before supplying the data to a classification model. In this paper, the capabilities of the Raspberry Pi 4 at performing activity recognition on CSI data are investigated. We have developed and publicly released a data interaction framework, capable of interpreting, processing and visualising data from a range of CSI-capable hardware. Furthermore, CSI data captured for these experiments during various activity performances have also been made publicly available. We then train a Deep Convolutional LSTM model to classify the activities. Our experiments, performed in a small apartment, achieve 92% average accuracy on 11 activity classes.

Long term health conditions, such as fall risk, are traditionally diagnosed through testing performed in hospital environments. Smart Homes offer the opportunity to perform continuous, long-term behavioural and vitals monitoring of residents, which may be employed to aid diagnosis and management of chronic conditions without placing additional strain on health services. A profile of the resident's behaviour can be produced from sensor data, and then compared over time. Activity Recognition is a primary challenge for profile generation, however many of the approaches adopted fail to take full advantage of the inherent temporal dependencies that exist in the activities taking place. Long Short Term Memory (LSTM) is a form of recurrent neural network that uses previously learned examples to inform classification decisions. In this paper we present a variety of approaches to human activity recognition using LSTMs which consider the temporal dependencies present in the sensor data in order to produce richer representations and improved classification accuracy. The LSTM approaches are compared to the performance of a selection of baseline classification algorithms on several real world datasets. In general, it was found that accuracy in LSTMs improved as additional temporal information was presented to the classifier.

Learning materials are increasingly available on the Web making them an excellent source of information for building e-Learning recommendation systems. However, learners often have difficulty finding the right materials to support their learning goals because they lack sufficient domain knowledge to craft effective queries that convey what they wish to learn. The unfamiliar vocabulary often used by domain experts creates a semantic gap between learners and experts, and also makes it difficult to map a learner's query to relevant learning materials. We build an e-Learning recommender system that uses background knowledge extracted from a collection of teaching materials and encyclopedia sources to support the refinement of learners' queries. Our approach allows us to bridge the gap between learners and teaching experts. We evaluate our method using a collection of realistic learner queries and a dataset of Machine Learning and Data Mining documents. Evaluation results show our method to outperform benchmark approaches and demonstrates its effectiveness in assisting learners to find the right materials.

Techniques for music recommendation are increasingly relying on hybrid representations to retrieve new and exciting music. A key component of these representations is musical content, with texture being the most widely used feature. Current techniques for representing texture however are inspired by speech, not music, therefore music representations are not capturing the correct nature of musical texture. In this paper we investigate two parts of the well-established mel-frequency cepstral coefficients (MFCC) representation: the resolution of mel-frequencies related to the resolution of musical notes; and how best to describe the shape of texture. Through contextualizing these parts, and their relationship to music, a novel music-inspired texture representation is developed. We evaluate this new texture representation by applying it to the task of music recommendation. We use the representation to build three recommendation models, based on current state-of-the-art methods. Our results show that by understanding two key parts of texture representation, it is possible to achieve a significant recommendation improvement. This contribution of a music-inspired texture representation will not only improve content-based representation, but will allow hybrid systems to take advantage of a stronger content component.

We have developed a novel hybrid representation for Music Information Retrieval. Our representation is built by incorporating audio content into the tag space in a tag-track matrix, and then learning hybrid concepts using latent semantic analysis. We apply this representation to the task of music recommendation, using similarity-based retrieval from a query music track. We also develop a new approach to evaluating music recommender systems, which is based upon the relationship of users liking tracks. We are interested in measuring the recommendation quality, and the rate at which cold-start tracks are recommended. Our hybrid representation is able to outperform a tag-only representation, in terms of both recommendation quality and the rate that cold-start tracks are included as recommendations.

Surrogate models of fitness have been presented as a way of reducing the number of fitness evaluations required by an evolutionary algorithm. This is of particular interest with expensive fitness functions where the cost of building the model is outweighed by the saving of using fewer function evaluations. In this paper we show how a Markov network model can be used as a surrogate fitness function in a genetic algorithm. We demonstrate this applied to a number of well-known benchmark functions and although the results are good in terms of function evaluations the model-building overhead requires a substantially more expensive fitness function to be worthwhile. We move on to describe a fitness function for feature selection in Case-Based Reasoning, which is considerably more expensive than the other benchmark functions we used. We show that for this problem using the surrogate offers a significant decrease in total run time compared to a GA using the true fitness function.