A Unified Framework For Automated, Accurate and Flexible Knowledge Graphs from Free Text
Research in Machine Learning AI has made great advances but recently, there has been much discussion centering around the issues with state-of-the-art models: lack of interpretability and transparency, inability to generalize and reason in unseen situations, and, frequently, the need to train these models on large labeled datasets that are often difficult and expensive to generate at the scale required. There is increasing interest in integrating symbolic knowledge representation and reasoning methods into Machine Learning solutions in order to tackle these issues and create adaptable systems that can be applied to a variety of domains and settings. Knowledge Graphs have been achieving increasing visibility in the research community as a form of structured representation of information. Integration of Knowledge Graphs into downstream tasks has already shown great potential in use-cases such as Question Answering and recommendation. A major bottleneck, however, is the still-unsolved challenge of automated creation and curation of Knowledge Graphs that are accurate, can be maintained with minimal manual intervention, and balance the tradeoff between adherence to design requirements and the flexibility necessary for integration of new knowledge and generalization. Our work focuses on addressing the Knowledge Graph creation and curation bottleneck, with specific focus on extracting knowledge from free text. We intend to tackle this by considering all aspects of end-to-end Knowledge Graph construction and application to downstream tasks: domain discovery, schema design, information extraction for KG population, Knowledge Graph completion, evaluation, as well as maintenance and downstream tasks. These aspects are usually considered in isolation in research work, whereas we propose to approach the bottleneck problem via a more unified framework that pushes the boundaries of existing methods individually and establishes end-to-end systems that benefit from mutual interaction and feedback. The first contribution addresses the bottleneck of Knowledge Graph construction and curation with particular attention to knowledge extraction and Knowledge Graph population from free text. Our approach consists of incorporating existing linguistic and domain-specific knowledge bases for downstream linguistic tasks, as well as enriching distributed semantic representations with syntactic information through recursive structures and neuro-symbolic reasoning. The goal of our efforts in this part of the project is to establish a two-way relationship between NLP methods and symbolic knowledge representation and reasoning.
Post hoc explanation for RNNs using state transition representations
AI and advanced machine learning techniques have had a significant impact on several facets of our life in recent years, taking over human positions in a variety of complex tasks. In domains as diverse as healthcare, banking, justice, and defence, their applications have had great success. Deep neural networks (DNNs), such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs), have gotten a lot of attention in recent years among these machine learning techniques. Since then, they have demonstrated unparalleled performance in tasks such as speech recognition, image recognition, natural language processing, and recommendation systems, outperforming humans in learning. However, it is observed that these best performing models are way too complex, abstract and opaque due to their complex deep architecture and non-linearity. Henceforth, they lack explainability and reliability. As they do not justify their decisions and predictions, it is difficult for humans to trust them. Unsurprisingly until recently, state-of-the-art CNNs, RRNs and other deep learning models, in general have been commonly regarded as “black boxes” or “black-box models”. Building trust in the deep learning model by validating its predictions and ensuring that it works as predicted and dependably on unseen or unfamiliar real-world data is unquestionably vital. In critical domains such as healthcare applications and autonomous vehicles, a single incorrect decision can have a catastrophic effect on society. Understanding, analyzing, visualizing, and explaining the rationale behind the model’s judgments and predictions is critical for ensuring the model’s reliability and understanding the model’s potential limitations and faultsA recurrent neural network (RNN) is a type of artificial neural network that uses sequential or time-series data. Recurrent neural networks, like feedforward and convolutional neural networks (CNNs), use training data to learn. They are distinguished by their “memory,” which allows them to use information from previous inputs to influence the current input and output. One of the challenging tasks with RNNs is to comprehend and evaluate their behavior. This is because it is difficult to understand what exactly they learn and also, they tend to work poorly on sequences requiring long-term memorization, despite having this capacity in principle. Many scholars have previously investigated a variety of strategies to address the aforementioned difficulties in recent years. The fundamental goal of this study is to investigate alternative methods for extracting interpretable state representations such as graphs, finite state machines, deterministic finite state machines from trained recurrent neural networks. The findings of this study will be useful across various domains in the industry and research community in order to provide better explanations to society for the deep learning applications they build and also comply with the GDPR rules.
Have you ever noticed that human behavior changes depending on surroundings and people nearby? Do you use different language styles with your friends, colleagues, or boss? As AI impacts more and more on multiple spheres of our life, we want it to be more tuned to our preferences and adapt as a human being. We want AI to respond in appropriate style and show us relevant information. We can argue that there are three levels of personalization: 1. Low level. This level of personalization is focused on low-level grammatical structures (choosing between simple or complicated words, sentence construction, etc.). 2. Medium level. Medium level focuses on semantics and content. It is at this level that AI decides how much information to show to a user and how many words he or she would need for this. 3. High level. At this level AI should be focused on what to say and how often to address a user and customized to user’s preferences. Which means totally or partially understanding of human personality. My Ph.D. will focus on optimizing conversational AI and conversational systems at a low level, and I will try to use some level 2 concepts. So, the main challenge I want to solve at these levels is to provide the user with relevant content. On the fly and in the context of dialogue systems. A good approach to tackling this problem is to use some basic pre-trained model, and fine tuning is so that we learn more and more about the user (similar to recommendation systems). Another approach that can be used is described in the article “PROTOTYPE-K-STYLE”: Create dialogs with style editing in RAM Su et al. We can also consider stylistic customization using GAN, but there are many problems due to the fact that the text is discrete in nature (therefore differentiation is not possible), however there are some heuristics that can solve this problem, so it is worth considering. I also want to look at the explainability of conversational AI, since this topic is closely related to human interaction. I believe that this work will make the interaction of users with dialog systems more convenient and natural, as well as help to find more relevant information easily and faster.
Using ML to Signal Gender and the use of Gendered Language
The focus of this project is to use machine learning and natural language processing to develop automatic techniques for identifying gender issues and bias in text content. There are a variety of application areas where such techniques can be useful. In 2018, Amazon scrapped the use of their AI internal recruitment model which showed significant bias against women. The model had been trained on the applications and CVs of successful applicants, most of whom were male, hence, it ‘learned’ that successful candidates were typically male. In the recent Labour leadership election in Britain, an analysis of the language used in news articles about the candidates showed discrepancies related to their gender in how they were described. The single male candidate was more likely to be discussed in terms of professional employment, politics and law and order and the two female candidates were much more likely to be discussed in terms of their families, in particular their fathers. Earlier projects in this area have used techniques pioneered by Google to help identify gender issues in news articles and to detect racist sentiment. The approach is based on the idea that gender attribution relies on language use, not on language itself; therefore, there are many other factors which should also be considered when determining who is being referred to in a text. Therefore, the inclusion of women’s representations in text can be argued to be not only important for simulating real-life occurrences, but also valuable as it allows us to understand how perception and social roles influence language use. Gender stereotype hypotheses about textual content tend to situate language use within a wider discourse about gender differences and the ways that they are constructed. Thus, the goal is that providing recommended linguistic modifications and positive reinforcement to authors about written text will influence and change behaviour. Signaling text content that suggests gendered language or is gender-biased can encourage and influence writing behaviour that is gender neutral. Hence, this project explores methods in supervised machine learning and natural language processing related to gender bias in text and gendered-language identification and prediction. The model will harness stylometric features, gender-specific language patterns, discourses of gender difference and principles of cognitive perception about an author’s identity and use NLP techniques to identify phrases, language, constructs or patterns in writing that signal use of textual content with gender bias or gender stereotyping.
Multi-Modal Generative Models of Stylistic Tuning — Making Personality more Personal.
The world is progressively globalising. As a result, many people want to learn new languages to exploit travel, business, and cultural enrichment opportunities. Unfortunately, not everyone has easy access to language learning classes, and they can be costly for people who do. The amount of virtual content and services people are using is increasing year to year. The rising volume of virtual services and content paired with the finite amount of spare time that all users have makes the selection of services and content within services more critical. Many services can be provided or supported by automatic text generation. There are support chatbots, virtual assistants, data to text generators, joke generators, and language tutors. Many of these advancements in text generation had become more feasible once deep neural networks were popularised and applied to the area. Personalisation applied to these services can make them more relevant to individual users. One fruitful application domain of these areas is language learning. An intelligent virtual tutor can service people regardless of location a lot cheaper than a human tutor. On the other hand, some people may not relate to a text generator like a human tutor. However, an intelligent virtual tutor that can adapt to user personality in a continuous fashion can become more relatable to the user. Another approach that may promote relatability is to combine a virtual reality avatar that can perform simple gestures in line with the generated text. The personality of a virtual tutor can be considered a form of style. This style can be customised to suit a user better. Generative adversarial networks or GANs are a neural mechanism often used to tune the outputs of media. Text can be passed through a network and tuned to provide a custom output. A research challenge would be to make these outputs fluid, dynamic, and automatic. We could combine GANs with other generative models to simulate personality for the language tutor. User studies would likely be used to evaluate this research. This project links to the ML-Labs challenge area of machine learning in language, a fundamental area of machine learning.
Project Title Web Accessibility Using Machine Learning Tools
The World Wide Web has become a ubiquitous platform for all kinds of everyday activities such as retail, entertainment, communication work, and education. The usability and accessibility of the web impact users of these services to varying degrees causing irritation to some and outright exclusion to others. People with disabilities and users of Assistive ICT such as screen readers, Switch devices and alternative I/O are particularly impacted by inaccessible websites. Access to Banking, Government Services and other important content and services is severely restricted. The response is in implementing Accessibility Standards and Guidelines such as the Web Content Accessibility Guidelines 2.1 and European Standards which embrace these such as the harmonised European standard. These guidelines are organised around the principles of Perceivable, Operable Understandable and Robust content. These are referred to as the P.O.U.R principles. They specify requirements such as Alternative Text for images and captioning for videos. They also insist on operational requirements such as good navigation and the ability to turn off animations and blinking effects. Recent work has looked at the role of how AI and ML can help meet these guidelines. This has produced tools for creating alternative text for images and auto-captioning for videos. ML-driven processes such as text simplification and word prediction are useful utilities. Technologies such as Speech recognition can provide alternative input mechanisms. These techniques have been aggregated into intelligent overlays with varying degrees of success. There are, however, problems with current versions of these technologies. They don’t work or only partially work in many instances causing major frustrations. There are ML approaches in other applications that the Web uses that could play a role here. Of particular interest is how ML can optimise navigation through websites. Work done in games for example could inform this work here. The question that this project addresses is what useful role can ML and AI play in making websites more accessible. Where the ML tools can best be deployed in development life-cycles. How they can best be supported through strategies like co-design and active inclusion. Of particular insight to this work is how intelligent web analytics and data insights from other web disciplines such as SEO and behaviour analysis can feed into the improvement of ML and AI web accessibility technologies.