A5036423275- Natural Language Processing Techniques
- Topic Modeling
- Semantic Web and Ontologies
- Text Readability and Simplification
- Interpreting and Communication in Healthcare
- Advanced Text Analysis Techniques
- Heart Failure Treatment and Management
- Machine Learning in Healthcare
- Data Quality and Management
- Intelligent Tutoring Systems and Adaptive Learning
- Sentiment Analysis and Opinion Mining
- Hate Speech and Cyberbullying Detection
- Multimodal Machine Learning Applications
- Software Engineering Research
- Speech and dialogue systems
- Artificial Intelligence in Law
- Language, Metaphor, and Cognition
- Handwritten Text Recognition Techniques
- Health disparities and outcomes
- Global Cancer Incidence and Screening
- Health, Environment, Cognitive Aging
- Forensic and Genetic Research
- Deception detection and forensic psychology
University of Copenhagen
2022-2024
University of Chile
2017-2020
Daniel Hershcovich, Stella Frank, Heather Lent, Miryam de Lhoneux, Mostafa Abdou, Stephanie Brandl, Emanuele Bugliarello, Laura Cabello Piqueras, Ilias Chalkidis, Ruixiang Cui, Constanza Fierro, Katerina Margatina, Phillip Rust, Anders Søgaard. Proceedings of the 60th Annual Meeting Association for Computational Linguistics (Volume 1: Long Papers). 2022.
In this paper we present the dataset of 200,000+ political arguments produced in local phase 2016 Chilean constitutional process. We describe human processing data by government officials, and manual tagging performed members our research group. Afterwards focus on classification tasks that mimic processes, comparing linear methods with neural network architectures. The experiments show some are suitable for automatization. particular, best achieve a 90% top-5 accuracy multi-class arguments,...
Pretrained language models can be queried for factual knowledge, with potential applications in knowledge base acquisition and tasks that require inference. However, that, we need to know how reliable this is, recent work has shown monolingual English lack consistency when predicting they fill-in-the-blank differently paraphrases describing the same fact. In paper, extend analysis of a multilingual setting. We introduce resource, mParaRel, investigate (i) whether such as mBERT XLM-R are more...
Abstract Large-scale pretrained language models (LMs) are said to “lack the ability connect utterances world” (Bender and Koller, 2020), because they do not have “mental of (Mitchell Krakauer, 2023). If so, one would expect LM representations be unrelated induced by vision models. We present an empirical evaluation across four families LMs (BERT, GPT-2, OPT, LLaMA-2) three model architectures (ResNet, SegFormer, MAE). Our experiments show that partially converge towards isomorphic those...
The increasing demand for the deployment of LLMs in information-seeking scenarios has spurred efforts creating verifiable systems, which generate responses to queries along with supporting evidence. In this paper, we explore attribution capabilities plan-based models have been recently shown improve faithfulness, grounding, and controllability generated text. We conceptualize plans as a sequence questions serve blueprints content its organization. propose two that utilize different variants...
Various efforts in the Natural Language Processing (NLP) community have been made to accommodate linguistic diversity and serve speakers of many different languages. However, it is important acknowledge that content they produce require, vary not just by language, but also culture. Although language culture are tightly linked, there differences. Analogous cross-lingual multilingual NLP, cross-cultural multicultural NLP considers these differences order better users systems. We propose a...
Facts are subject to contingencies and can be true or false in different circumstances. One such contingency is time, wherein some facts mutate over a given period, e.g., the president of country winner championship. Trustworthy language models ideally identify mutable as process them accordingly. We create MuLan, benchmark for evaluating ability English anticipate time-contingency, covering both 1:1 1:N relations. hypothesize that encoded differently than immutable ones, hence being easier...
The purpose of instruction tuning is enabling zero-shot performance, but has also been shown to improve chain-of-thought reasoning and value alignment (Si et al., 2023). Here we consider the impact on $\textit{consistency}$, i.e., sensitivity language models small perturbations in input. We compare 10 instruction-tuned LLaMA original LLaMA-7b model show that almost across-the-board they become more consistent, both terms their representations predictions downstream tasks. explain these...
Knowledge claims are abundant in the literature on large language models (LLMs); but can we say that GPT-4 truly "knows" Earth is round? To address this question, review standard definitions of knowledge epistemology and formalize interpretations applicable to LLMs. In doing so, identify inconsistencies gaps how current NLP research conceptualizes with respect epistemological frameworks. Additionally, conduct a survey 100 professional philosophers computer scientists compare their...
Large Language Models (LLMs) store and retrieve vast amounts of factual knowledge acquired during pre-training. Prior research has localized identified mechanisms behind recall; however, it primarily focused on English monolingual models. The question how these processes generalize to other languages multilingual LLMs remains unexplored. In this paper, we address gap by conducting a comprehensive analysis two highly LLMs. We assess the extent which previously components recall in apply...
Linking information on family members in the Danish Civil Registration System (CRS) with national registers provides unique possibilities for research familial aggregation of diseases, health patterns, social factors and demography. However, CRS is limited number generations that it can identify. To allow more complete linkages, we introduce lite Multi-Generation Register (lite MGR) future full MGR currently being developed.
Cross-lingual summarization consists of generating a summary in one language given an input document different language, allowing for the dissemination relevant content across speakers other languages. The task is challenging mainly due to paucity cross-lingual datasets and compounded difficulty summarizing translating. This work presents $\mu$PLAN, approach that uses intermediate planning step as bridge. We formulate plan sequence entities capturing summary's order which it should be...
Deep learning techniques have been successfully applied to predict unplanned readmissions of patients in medical centers. The training data for these models is usually based on historical records that contain a significant amount free-text from admission reports, referrals, exam notes, etc. Most the proposed so far are tailored English text and assume electronic follow standards common developed countries. These two characteristics make them difficult apply developing countries do not...
Pretrained language models can be queried for factual knowledge, with potential applications in knowledge base acquisition and tasks that require inference. However, that, we need to know how reliable this is, recent work has shown monolingual English lack consistency when predicting they fill-in-the-blank differently paraphrases describing the same fact. In paper, extend analysis of a multilingual setting. We introduce resource, mParaRel, investigate (i) whether such as mBERT XLM-R are more...