Zhiguo Wang, Patrick Ng, Xiaofei Ma, Ramesh Nallapati, Bing Xiang. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and 9th International Joint (EMNLP-IJCNLP). 2019.

The performance of deep neural models can deteriorate substantially when there is a domain shift between training and test data. For example, the pre-trained BERT model be easily fine-tuned with just one additional output layer to create state-of-the-art for wide range tasks. However, suffers considerably at zero-shot applied different domain. In this paper, we present novel two-step adaptation framework based on curriculum learning domain-discriminative data selection. conducted in mostly...

We present a systematic investigation of layer-wise BERT activations for general-purpose text representations to understand what linguistic information they capture and how transferable are across different tasks. Sentence-level embeddings evaluated against two state-of-the-art models on downstream probing tasks from SentEval, while passage-level four question-answering (QA) datasets under learning-to-rank problem setting. Embeddings the pre-trained model perform poorly in semantic...

Testing plays a pivotal role in ensuring software quality, yet conventional Search Based Software (SBST) methods often struggle with complex units, achieving suboptimal test coverage. Recent work using large language models (LLMs) for generation have focused on improving quality through optimizing the context and correcting errors model outputs, but use fixed prompting strategies that prompt to generate tests without additional guidance. As result LLM-generated testsuites still suffer from...

Introduction Glioma segmentation is vital for diagnostic decision-making, monitoring disease progression, and surgical planning. However, this task hindered by substantial heterogeneity within gliomas imbalanced region distributions, posing challenges to existing methods. Methods To address these challenges, we propose the DeepGlioSeg network, a U-shaped architecture with skip connections continuous contextual feature integration. The model includes two primary components. First, CTPC...

Generative models for Information Retrieval, where ranking of documents is viewed as the task generating a query from document's language model, were very successful in various IR tasks past. However, with advent modern deep neural networks, attention has shifted to discriminative functions that model semantic similarity and queries instead. Recently, generative such GPT2 BART have been shown be excellent text generators, but their effectiveness rankers not demonstrated yet. In this work, we...

Danilo Neves Ribeiro, Shen Wang, Xiaofei Ma, Rui Dong, Xiaokai Wei, Henghui Zhu, Xinchi Chen, Peng Xu, Zhiheng Huang, Andrew Arnold, Dan Roth. Findings of the Association for Computational Linguistics: NAACL 2022.

Abstract Glioblastoma multiforme (GBM) is the most common and deadly primary malignant brain tumor. As GBM tumor aggressive shows high biological heterogeneity, overall survival (OS) time extremely low even with treatment. If OS can be predicted before surgery, developing personalized treatment plans for patients will beneficial. Magnetic resonance imaging (MRI) a commonly used diagnostic tool tumors high‐resolution sound effects. However, in clinical practice, doctors mainly rely on...

Testing plays a pivotal role in ensuring software quality, yet conventional Search Based Software (SBST) methods often struggle with complex units, achieving suboptimal test coverage. Recent work using large language models (LLMs) for generation have focused on improving quality through optimizing the context and correcting errors model outputs, but use fixed prompting strategies that prompt to generate tests without additional guidance. As result LLM-generated suites still suffer from low...

Despite profound successes, contrastive representation learning relies on carefully designed data augmentations using domain-specific knowledge. This challenge is magnified in natural language processing, where no general rules exist for augmentation due to the discrete nature of language. We tackle this by presenting a Virtual Supported Contrastive Learning sentence representations (VaSCL). Originating from interpretation that essentially constructs neighborhoods each training instance, we,...

Summary Electromagnetic signal emitted by satellite communication (satcom) transmitters are used to identify specific individual uplink satcom terminals sharing the common transponder in real environment, which is known as emitter identification (SEI) that allows for early indications and warning (I&W) of targets carrying furnishment furthermore time electromagnetic situation awareness military operations. In this paper, authors first propose using probabilistic neural networks (PNN)...

In this paper, we propose a \textit{weak supervision} framework for neural ranking tasks based on the data programming paradigm \citep{Ratner2016}, which enables us to leverage multiple weak supervision signals from different sources. Empirically, consider two sources of signals, unsupervised functions and semantic feature similarities. We train BERT-based passage-ranking model (which achieves new state-of-the-art performances benchmark datasets with full supervision) in our framework....

BERT model has been successfully applied to open-domain QA tasks. However, previous work trains by viewing passages corresponding the same question as independent training instances, which may cause incomparable scores for answers from different passages. To tackle this issue, we propose a multi-passage globally normalize answer across all of question, and change enables our find better utilizing more In addition, that splitting articles into with length 100 words sliding window improves...

The performance of state-of-the-art neural rankers can deteriorate substantially when exposed to noisy inputs or applied a new domain.In this paper, we present novel method for fine-tuning that significantly improve their robustness out-of-domain data and query perturbations.Specifically, contrastive loss compares points in the representation space is combined with standard ranking during fine-tuning.We use relevance labels denote similar/dissimilar pairs, which allows model learn underlying...

Zhihan Zhou, Dejiao Zhang, Wei Xiao, Nicholas Dingwall, Xiaofei Ma, Andrew Arnold, Bing Xiang. Proceedings of the 2022 Conference North American Chapter Association for Computational Linguistics: Human Language Technologies. 2022.

We introduce STREET, a unified multi-task and multi-domain natural language reasoning explanation benchmark. Unlike most existing question-answering (QA) datasets, we expect models to not only answer questions, but also produce step-by-step structured explanations describing how premises in the question are used intermediate conclusions that can prove correctness of certain answer. perform extensive evaluation with popular such as few-shot prompting GPT-3 fine-tuned T5. find these still lag...

Prateek Yadav, Qing Sun, Hantian Ding, Xiaopeng Li, Dejiao Zhang, Ming Tan, Parminder Bhatia, Xiaofei Ma, Ramesh Nallapati, Murali Krishna Ramanathan, Mohit Bansal, Bing Xiang. Proceedings of the 61st Annual Meeting Association for Computational Linguistics (Volume 2: Short Papers). 2023.

Missing information is a common issue of dialogue summarization where some in the reference summaries not covered generated summaries. To address this issue, we propose to utilize natural language inference (NLI) models improve coverage while avoiding introducing factual inconsistencies. Specifically, use NLI compute fine-grained training signals encourage model generate content that have been covered, as well distinguish between factually consistent and inconsistent sentences. Experiments...

Brain connectivity analysis plays an essential role in the research of working memory that involves complex coordination various brain regions. In this research, we present a comprehensive view trans-states variation based on continuous scalp EEG, extending beyond traditional stimuli-lock averaging or restriction to short time scales hundreds milliseconds after stimulus onset. The EEG was collected under three conditions: quiet, memory, and control. only difference between control conditions...

Timely detection of dynamical complexity changes in natural and man-made systems has deep scientific practical meanings. We introduce a measure for time series: the base-scale entropy. The definition directly applies to arbitrary real-word data. illustrate our method on speech signal theoretical chaotic system. results show that simple easily calculated entropy can be effectively used detect qualitative quantitative changes.

Recent progress in pretrained Transformer-based language models has shown great success learning contextual representation of text. However, due to the quadratic self-attention complexity, most Transformers can only handle relatively short It is still a challenge when it comes modeling very long documents. In this work, we propose use graph attention network on top available model learn document embeddings. This allows us leverage high-level semantic structure document. addition, based our...