I am a PhD student in Computational Linguistics under supervision of Prof. Massimo Poesio at the Cognitive Science Group (CogSci) at Queen Mary University of London (QMUL). As a member of the Disagreements and Language Interpretation (DALI) project I focus on underspecified expressions and ambiguous language use in dialogue settings and try to develop a better understanding of the special cases in which these theoretically problematic formulations don’t seem to disrupt an interaction and sometimes even make it more efficient.
You can download my CV here.
For a set of pilot experiment runs we are looking for pairs of native English speakers. The experiment takes about 30 minutes including instructions and has two participants re-ordering a story cut into a number of snippets. Nothing fancy and nothing complicated. If you are interested, please send a mail to firstname.lastname@example.org.
Visit our new website for the PhotoBook Task and Dataset at www.dmg-photobook.github.io.
Our paper “The PhotoBook Dataset: Building Common Ground through Visually-Grounded Dialogue” was accepted at ACL 2019. The paper is available on arxiv.org here.
I will be presenting the paper at ACL 2019 in Florence on Monday, the 29th of July in evening session 3E: Vision, Robotics, Multimodal, Grounding and Speech.
Our paper “Building Common Ground in Visual Dialogue: The PhotoBook Task and Dataset” was accepted at SemDial 2018. The paper is available in the conference proceedings here.
Our paper “Building Common Ground in Visual Dialogue: The PhotoBook Task and Dataset“ was accepted at the Workshop on Shortcomings in Vision and Language at ECCV 2018.
Bachelor Tehsis written under supervision of Dr. Roberto Valenti. Topic: Modeling Distributed Cybernetic Management for Resource Based Economies - A simulation approach to Stafford Beer’s 1971 CyberSyn Project.
Cum Laude, GPA: 8.3
Advanced courses in Mathematics, Chemistry and Politics & Economy
Final grade: 1.6
The past few years have seen an immense interest in developing and training computational agents for visually-grounded dialogue, the task of using natural language to communicate about visual input. While the resulting dialogue agents often already achieve reasonable performance on their respective task, none of the models can produce consistent and efficient outputs during a multi-turn conversation. We argue that this is primarily due to the fact that they cannot properly utilise the dialogue history. Human interlocutors on the other hand are shown to collaboratively establish a shared repository of mutual information during a conversation. This common ground then is used to optimise understanding and communication efficiency. We therefore propose that implementing a similar representation of dialogue context for computational dialogue agents is a pivotal next step in improving the quality of their dialogue output.
One of the main reasons why current research seems to eschew modelling common ground is that common ground is a conversation model concept that cannot be assessed directly in actual conversations. In order to address this problem, we propose to first investigate the generation of referring expressions: Being an indirect representation of a referent object, they too are not absolute, but conventions established with a specific conversation partner - based on the common ground established so far. By tracking the development of referring expressions during a conversation we therefore obtain a proxy of the underlying processes in the emerging common ground.
In order to develop an artificial dialogue agent that can utilise the conversation's common ground, we propose to implement a data-driven, modular agent architecture in an end-to-end training framework. With this setup, the dialogue agent is expected to learn the correct usage of referring expressions from recorded dialogue data directly and can be evaluated on downstream task performance. Opting for this approach however requires a large amount of dedicated dialogue training data that has never been collected before. To initiate this new track in dialogue modelling research, we therefore introduce a novel conversation task called the PhotoBook task that can be used to collect rich, human-human dialogue data for extended, goal-oriented conversations. We use the PhotoBook task to record more than 2,500 dialogues stemming from over 1,500 unique participants on crowd-sourcing platform Amazon Mechanical Turk (AMT). The resulting data contains a total of over 160k utterances, 130k actions and spans a vocabulary of close to 12k unique tokens. An extensive analysis of the data validates that the recorded conversations closely resemble the dialogue characteristics observed in natural human-human conversations. We therefore argue that this data provides a pivotal new repository to be used in further research which has the potential to significantly improve the dialogue output consistency, efficiency and naturalness of artificial dialogue agents.
My thesis is available as a download from the university's server here.
In 2018, Xu and Reitter introduced a novel, information-theoretic view of dialogue, in which they proposed modeling a conversation between two interlocutors as a two-way communication system. In such a system the information flow follows a number of general principles. One of those principles that is assumed to hold in dialogue as well is the Uniform Information Density hypothesis (UID). The UID hypothesis states that a communication system as a whole has the tendency to distribute data in such a way that the density of information remains constant.
In two-party dialogue both interlocutors are equal parts of the communication system. This means that they are jointly responsible for the level of information density at every moment of the conversation. In order to ensure the validity of the UID hypothesis, the two speakers must therefore have an agreement in an implicit sense about their contribution to the conversation. Xu and Reitter propose that speakers take on certain roles during a conversation: One leads the conversation by steering the ongoing topic, while the other follows along. These roles can switch during a conversation and rather than steering turn-taking behavior, they describe a higher-level segmentation of a conversation into topics.
In this research we investigate whether we can detect the boundaries of these conversation segments, formally referred to as topic shifts, based on the speaker’s contribution to the conversation alone. To this end we extract a number of simple syntactical features that have been shown to correlate well with the amount of information transmitted and build a simple prediction model based on these features.
While we had to conclude that this simple approach does not yield a model expressive enough to correctly predict topic shifts produced by more involved methods, we beleive that it nonetheless produces coherent and intuitively sound topic segments even for noisy dialogue transcripts. As a next step, we will investigate different methods to validate this claim.You can view our preliminary results here.
The continuously increasing amount of online news articles requires new ways of filtering relevant information into a human-digestible form. Recently, research has focused on providing such selections by generating timelines for known entities through extending and extracting information from Knowledge Graphs. Contrasting this approach, we propose a new method to generate an entity timeline based directly on a non-curated, unstructured set of news items so as to allow this approach to be extended to long-tail entities.
In this research, Wikipedia pages of entities are seen as a gold-label timeline consisting of information cited from news-worthy articles, while other news articles about those entities that are not cited are treated as negative examples. To learn what makes an article news-worthy, we take a supervised approach based on a set of 28 handcrafted features.
One of our main contributions is a novel, larger dataset for this task, covering 379 unique entities and containing 13146 news articles with an equal distribution of positive and negative examples per entity. Using this dataset we obtain a basic classification accuracy of 68.9% for deciding whether an unseen news article contains relevant information about a given entity. As a baseline method of evaluation, the top article predictions per entity are then summarized and concatenated to generate a dummy Wikipedia entries which we compare to the original ones. As no standardized, gold-label evaluation methods were developed yet, we also propose an A/B testing method for a more qualitative performance estimate.
You can read the project paper here.
In the early 1970's, AI once before was THE big thing that would revolutionize the world as they knew it. Many great researchers were optimistic that artificial systems with general intelligence were within grasp - and big plans were made to apply such systems to solve real-world problems. Among the most notorious ones: 1971’s CyberSyn project of British economist Stafford Beer - which came to an abrupt and violent end just two years later.
The context: 1970 Chile elected its first socialist president, Salvador Allende, which in turn appointed Fernando Flores, a young scientist devoted to the study of operations research and scholar of Beer’s work on the subject of management cybernetics to be the General Technical Manager of the new-found state development agency. In that function, Flores invited Beer to design and implement a cybernetic system to automatize the administration of the entire Chilean economy. An ambitious project that after taking first steps was cut short by a military coup in 1973.
In this research we aim to investigate whether the simple cybernetic approach proposed by Stafford and his colleagues could have been sufficient to manage something as complex as the Chilean state economy. We do so by modeling a simplified economic setting governed by CyberSyn’s management principles and analyze the model’s performance under a range of different parameter settings. The results of these initial experiments suggest that the model indeed exhibits emerging self-sustainability and lead to the conclusion that CyberSyn’s approach might have been principally feasible.
You can find my Bachelor's Thesis here.
Many critics of AI argue that intentionality in computers - or any other artifact for that matter - can never be more than derivative. With the words of John Haugeland, their “tokens only have meaning because we give it to them” and consequently, “they only mean what we say that they do”. Contemporary philosopher Daniel Dennett however claims that “there is no principled (theoretically motivated) way to distinguish ‘original’ intentionality from ‘derived’ intentionality.” On the basis of this idea, he developed a three-stage model to explain the assignment of intentionality and refute the objection of derived intentionality in artifacts.
In this essay, we analyze Dennett’s model and answer the question How does Dennett’s elaborated model of the intentional stance answer Haugeland’s objection that intentionality in artifacts cannot be original?
You can read the essay here.
Courses: Computersystemen, Computational Logic, Brein & Cognitie and Natuurlijke Taalmodellen en Interfaces
Main tasks: Organizing and supporting regular youth and childrens' meetings, maitaining public relations, improving communications with supporters and helping out the center's founders and members in a wide range of tasks.
Internationaler Jugendfreiwilligendienst (IJFD) with CFI Freiwilligendienste
Courses: Brein & Cognitie and Natuurlijke Taalmodellen en Interfaces
Main tasks: Teaching classes in Informatics and Politics (seniors), assisting the Kindergarten supervisors and organizing activities for children in the affiliated orphanage.
Internationaler Jugendfreiwilligendienst (IJFD) with Co-Workers International
Main task: Developing whitepapers concerning the topic 'IT-Offshoring in India'.
Robotics Workshop for KUKA youBot standard platform.
In a team developed, programmed and presented an interactive learning software.
Every moment not filled with studies is definitively filled with music - either digital, analog, live or self-made. I play the guitar far longer now than my skill level might indicate - though somebody once might have said something about being me my own worst critic... whatever. My little home studio is growing slowly but constantly (mostly limited by my 20 sqm freight container appartment) and you might find some of my own music here once I decide it's ready for the great big world.