Prediction without justification has limited applicability.As a remedy, we learn to extract pieces of input text as justifications -rationales -that are tailored be short and coherent, yet sufficient for making the same prediction.Our approach combines two modular components, generator encoder, which trained operate well together.The specifies distribution over fragments candidate rationales these passed through encoder prediction.Rationales never given during training.Instead, model is...

This paper focuses on style transfer the basis of non-parallel text. is an instance a broad family problems including machine translation, decipherment, and sentiment modification. The key challenge to separate content from other aspects such as style. We assume shared latent distribution across different text corpora, propose method that leverages refined alignment representations perform transfer. transferred sentences one should match example population. demonstrate effectiveness this...

Common recurrent neural architectures scale poorly due to the intrinsic difficulty in parallelizing their state computations. In this work, we propose Simple Recurrent Unit (SRU), a light unit that balances model capacity and scalability. SRU is designed provide expressive recurrence, enable highly parallelized implementation, comes with careful initialization facilitate training of deep models. We demonstrate effectiveness on multiple NLP tasks. achieves 5—9x speed-up over cuDNN-optimized...

Target detection in UAV images is of great significance fields such as traffic safety, emergency rescue, and environmental monitoring. However, captured by UAVs usually have multi-scale features, complex backgrounds, uneven illumination, low target resolution, which makes very challenging. To tackle these challenges, this paper introduces SPDC-YOLO, a novel model built upon YOLOv8. In the backbone, eliminates last C2f module final downsampling module, thus avoiding loss small features. neck,...

Accurate scoring of syntactic structures such as head-modifier arcs in dependency parsing typically requires rich, highdimensional feature representations.A small subset features is often selected manually.This problematic when lack clear linguistic meaning embeddings or the information blended across features.In this paper, we use tensors to map high-dimensional vectors into low dimensional representations.We explicitly maintain parameters a low-rank tensor obtain representations words...

The success of deep learning often derives from well-chosen operational building blocks.In this work, we revise the temporal convolution operation in CNNs to better adapt it text processing.Instead concatenating word representations, appeal tensor algebra and use low-rank n-gram tensors directly exploit interactions between words already at stage.Moreover, extend non-consecutive recognize patterns with intervening words.Through a combination lowrank tensors, pattern weighting, can...

Tao Lei, Hrishikesh Joshi, Regina Barzilay, Tommi Jaakkola, Kateryna Tymoshenko, Alessandro Moschitti, Lluís Màrquez. Proceedings of the 2016 Conference North American Chapter Association for Computational Linguistics: Human Language Technologies. 2016.

Large language models have recently achieved state of the art performance across a wide variety natural tasks. Meanwhile, size these and their latency significantly increased, which makes usage costly, raises an interesting question: do need to be large? We study this question through lens model compression. present generic, structured pruning approach by parameterizing each weight matrix using its low-rank factorization, adaptively removing rank-1 components during training. On modeling...

Deep learning algorithms have recently provided new ideas for various change detection (CD) tasks, which yielded promising results. However, accurately identifying urban land cover and use (LCLU) changes remains challenging in the very high-resolution (HR) remote sensing images due to difficulties effectively modeling features from ground objects with different times spatial locations. In this letter, a multiscale swin transformer (MST)-based deeply supervised network (MSTDSNet) is proposed...

Vehicle detection in aerial images is an important and challenging task. Traditionally, many target models based on sliding-window fashion were developed achieved acceptable performance, but these are time-consuming the phase. Recently, with great success of convolutional neural networks (CNNs) computer vision, state-of-the-art detectors have been designed deep CNNs. However, CNN-based inefficient when applied image data due to fact that existing struggle small-size object precise...

We address the problem of detecting duplicate questions in forums, which is an important step towards automating process answering new questions. As finding and annotating such potential duplicates manually very tedious costly, automatic methods based on machine learning are a viable alternative. However, many forums do not have annotated data, i.e., labeled by experts as duplicates, thus promising solution to use domain adaptation from another forum that has annotations. Here we focus...

Dependency parsing with high-order features results in a provably hard decoding problem.A lot of work has gone into developing powerful optimization methods for solving these combinatorial problems.In contrast, we explore, analyze, and demonstrate that substantially simpler randomized greedy inference algorithm already suffices near optimal parsing: a) analytically quantify the number local optima method to overcome context first-order parsing; b) show that, as algorithm, surpasses dual...

Most object detection methods based on remote sensing images are generally dependent a large amount of high-quality labeled training data. However, due to the slow acquisition cycle and difficulty in labeling, many types data samples scarce. This makes few-shot an urgent necessary research problem. In this paper, we introduce method text semantic fusion relation graph reasoning (TSF-RGR), which learns various relationships from common sense knowledge end-to-end manner, thereby empowering...

Infrared (IR) small target detection in sky scenes is crucial for aerospace, border security, and atmospheric monitoring. Most current works are typically designed generalized IR scenes, which may not be optimal the specific scenario of backgrounds, particularly detecting dim targets at long ranges. In these scenarios, presence heavy clouds usually causes significant false alarms due to factors such as strong edges, streaks, large undulations, isolated floating clouds. To address challenges,...

Tao Lei, Yuan Zhang, Lluís Màrquez, Alessandro Moschitti, Regina Barzilay. Proceedings of the 2015 Conference North American Chapter Association for Computational Linguistics: Human Language Technologies. 2015.

Semantic image segmentation, as one of the most popular tasks in computer vision, has been widely used autonomous driving, robotics and other fields. Currently, deep convolutional neural networks (DCNNs) are driving major advances semantic segmentation due to their powerful feature representation. However, DCNNs extract high-level representations by strided convolution, which makes it impossible segment foreground objects precisely, especially when locating object boundaries. This paper...