We present a theoretically grounded approach to train deep neural networks, including recurrent subject class-dependent label noise. propose two procedures for loss correction that are agnostic both application domain and network architecture. They simply amount at most matrix inversion multiplication, provided we know the probability of each class being corrupted into another. further show how one can estimate these probabilities, adapting recent technique noise estimation multi-class...

Dat Quoc Nguyen, Kairit Sirts, Lizhen Qu, Mark Johnson. Proceedings of the 2016 Conference North American Chapter Association for Computational Linguistics: Human Language Technologies. 2016.

Recent progress in information extraction has shown how to automatically build large ontologies from high-quality sources like Wikipedia. But knowledge evolves over time; facts have associated validity intervals. Therefore, should include time as a first-class dimension. In this paper, we introduce Timely YAGO, which extends our previously built base YAGO with temporal aspects. This prototype system extracts Wikipedia infoboxes, categories, and lists articles, integrates these into the base....

In this paper, we propose the first model to be able generate visually grounded questions with diverse types for a single image. Visual question generation is an emerging topic which aims ask in natural language based on visual input. To best of our knowledge, it lacks automatic methods meaningful various same circumvent problem, that automatically generates varying types. Our takes as input both images and captions generated by dense caption model, samples most probable types, sequel. The...

A major cause of security incidents such as cyber attacks is rooted in software vulnerabilities. These vulnerabilities should ideally be found and fixed before the code gets deployed. Machine learning-based approaches achieve state-of-the-art performance capturing methods are predominantly supervised. Their prediction models trained on a set ground truth data where training test assumed to drawn from same probability distribution. However, practice, often differs terms distribution because...

With increasing concerns about data privacy, there is an necessity of fine-tuning pre-trained language models (PLMs) for adapting to downstream tasks located in end-user devices or local clients without transmitting the central server. This urgent therefore calls research investigating federated learning (FL) PLMs. However, large PLMs bring curse prohibitive communication overhead and model adaptation costs FL system. To this end, we investigate parameter-efficient tuning (PETuning) develop...

There have been major advances on automatically constructing large knowledge bases by extracting relational facts from Web and text sources. However, the world is dynamic: periodic events like sports competitions need to be interpreted with their respective timepoints, such as coaching a team, holding political or business positions, even marriages do not hold forever should augmented timespans. This paper addresses problem of harvesting temporal extended time-awareness. We employ...

Knowledge bases are useful resources for many natural language processing tasks, however, they far from complete.In this paper, we define a novel entity representation as mixture of its neighborhood in the knowledge base and apply technique on TransE-a well-known embedding model completion.Experimental results show that information significantly helps to improve TransE, leading better performance than obtained by other state-of-the-art models three benchmark datasets triple classification,...

In social media, demographic inference is a critical task in order to gain better understanding of cohort and facilitate interacting with one’s audience. Most previous work has made independence assumptions over topological, textual label information on networks. this work, we employ recursive neural networks break down these obtain about characteristics Twitter. We show that our model performs than existing models including the state-of-the-art.

Application Programming Interfaces (APIs) have been widely discussed on social-technical platforms (e.g., Stack Overflow). Extracting API mentions from such informal software texts is the prerequisite for API-centric search and summarization of programming knowledge. Machine learning based extraction has demonstrated superior performance than rule-based methods in that lack consistent writing forms annotations. However, machine a significant overhead preparing training data effective...

Existing object detection methods are bounded in a fixed-set vocabulary by costly labeled data. When dealing with novel categories, the model has to be retrained more bounding box annotations. Natural language supervision is an attractive alternative for its annotation-free attributes and broader concepts. However, learning open-vocabulary from challenging since image-text pairs do not contain fine-grained object-language alignments. Previous solutions rely on either expensive grounding...

Large language models (LLMs) have made significant strides in various natural processing (NLP) tasks. Recent research shows that the moderately-sized LLMs often outperform their larger counterparts after task-specific fine-tuning. In this work, we delve into process of adapting to specialize document-level machine translation (DocMT) for a specific pair. Firstly, explore how prompt strategies affect downstream performance. Then, conduct extensive experiments with two fine-tuning methods,...

In named entity recognition, we often don't have a large in-domain training corpus or knowledge base with adequate coverage to train model directly.In this paper, propose method where, given data in related domain similar (but not identical) (NE) types and small amount of data, use transfer learning learn domain-specific NE model.That is, the novelty task setup is that assume just mismatch, but also label mismatch.

In this paper, we investigate the diversity aspect of paraphrase generation. Prior deep learning models employ either decoding methods or add random input noise for varying outputs. We propose a simple method Diverse Paraphrase Generation (D-PAGE), which extends neural machine translation (NMT) to support generation diverse paraphrases with implicit rewriting patterns. Our experimental results on two real-world benchmark datasets demonstrate that our model generates at least one order...

Natural Language Generation (NLG) supports the creation of personalized, contextualized, and targeted content. However, algorithms underpinning NLG have come under scrutiny for reinforcing gender, racial, other problematic biases. Recent research in seeks to remove these biases through principles fairness privacy. Drawing on gender queer theories from sociology Science Technology studies, we consider how can contribute towards advancement equity society. We propose a conceptual framework...

Randomised controlled trials (RCTs) are the cornerstone of evidence-based medicine. Unfortunately, not all RCTs based on real data. This serious breach research integrity compromises reliability systematic reviews and meta-analyses, leading to misinformed clinical guidelines posing a risk both individual public health. While methods detect problematic have been proposed, they time-consuming labour-intensive. The use artificial intelligence large language models (LLM) has potential accelerate...

Teacher-forcing training for audio captioning usually leads to exposure bias due and inference mismatch. Prior works propose the contrastive method deal with caption degeneration. However, ignores temporal information when measuring similarity across acoustic linguistic modalities, leading inferior performance. In this work, we develop temporal-similarity score by introducing unbiased sliced Wasserstein RBF (USW-RBF) kernel equipped rotary positional embedding account modalities. contrast...

Alignment tuning is crucial for ensuring large language models (LLMs) behave ethically and helpfully. Current alignment approaches require high-quality annotations significant training resources. This paper proposes a low-cost, tuning-free method using in-context learning (ICL) to enhance LLM alignment. Through an analysis of ICL demos, we identified style as key factor influencing capabilities explicitly restyled exemplars based on this stylistic framework. Additionally, combined the demos...

Due to the ubiquity of computer software, software vulnerability detection (SVD) has become crucial in industry and field security. Two significant issues SVD arise when using machine learning, namely: i) how learn automatic features that can help improve predictive performance ii) overcome scarcity labeled vulnerabilities projects require laborious labeling code by security experts. In this paper, we address these two concerns proposing a novel architecture which leverages deep domain...

Content Warning: this paper may contain content that is offensive or upsetting.

The rapid success of Large Language Models (LLMs) has unlocked vast potential for AI applications in privacy-sensitive domains. However, the traditional centralized training LLMs poses significant challenges due to privacy concerns regarding collecting sensitive data from diverse sources. This paper offers a promising and privacy-enhancing solution LLMs: collaboratively via Federated Learning (FL) across multiple clients, eliminating need raw transmission. To this end, we present F4LLM, new...