The task of named entity recognition (NER) is normally divided into nested NER and flat depending on whether entities are or not.Models usually separately developed for the two tasks, since sequence labeling models, most widely used backbone NER, only able to assign a single label particular token, which unsuitable where token may be assigned several labels. In this paper, we propose unified framework that capable handling both tasks. Instead treating as problem, formulate it machine reading...

Many NLP tasks such as tagging and machine reading comprehension are faced with the severe data imbalance issue: negative examples significantly outnumber positive examples, huge number of easy-negative overwhelms training. The most commonly used cross entropy (CE) criteria is actually an accuracy-oriented objective, thus creates a discrepancy between training test: at time, each instance contributes equally to objective function, while test time F1 score concerns more about examples. In...

In this paper, we propose a new paradigm for the task of entity-relation extraction. We cast as multi-turn question answering problem, i.e., extraction entities and elations is transformed to identifying answer spans from context. This QA formalization comes with several key advantages: firstly, query encodes important information entity/relation class want identify; secondly, provides natural way jointly modeling entity relation; thirdly, it allows us exploit well developed machine reading...

Zijun Sun, Xiaoya Li, Xiaofei Yuxian Meng, Xiang Ao, Qing He, Fei Wu, Jiwei Li. Proceedings of the 59th Annual Meeting Association for Computational Linguistics and 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

Single domain generalization is a challenging case of model generalization, where the models are trained on single and tested other unseen domains. A promising solution to learn cross-domain invariant representations by expanding coverage training domain. These methods have limited performance gains in practical applications due lack appropriate safety effectiveness constraints. In this paper, we propose novel learning framework called progressive expansion network (PDEN) for generalization....

Despite the fact that large-scale Language Models (LLM) have achieved SOTA performances on a variety of NLP tasks, its performance NER is still significantly below supervised baselines. This due to gap between two tasks and LLMs: former sequence labeling task in nature while latter text-generation model. In this paper, we propose GPT-NER resolve issue. bridges by transforming generation can be easily adapted LLMs e.g., finding location entities input text "Columbus city" transformed generate...

Despite the remarkable success of large-scale Language Models (LLMs) such as GPT-3, their performances still significantly underperform fine-tuned models in task text classification.This is due to (1) lack reasoning ability addressing complex linguistic phenomena (e.g., intensification, contrast, irony etc); (2) limited number tokens allowed in-context learning. In this paper, we introduce Clue And Reasoning Prompting (CARP). CARP adopts a progressive strategy tailored involved...

This paper surveys research works in the quickly advancing field of instruction tuning (IT), a crucial technique to enhance capabilities and controllability large language models (LLMs). Instruction refers process further training LLMs on dataset consisting \textsc{(instruction, output)} pairs supervised fashion, which bridges gap between next-word prediction objective users' having adhere human instructions. In this work, we make systematic review literature, including general methodology...

Segmenting a chunk of text into words is usually the first step processing Chinese text, but its necessity has rarely been explored. In this paper, we ask fundamental question whether word segmentation (CWS) necessary for deep learning-based Natural Language Processing. We benchmark neural word-based models which rely on against char-based do not involve in four end-to-end NLP tasks: language modeling, machine translation, sentence matching/paraphrase and classification. Through direct...

It is intuitive that NLP tasks for logographic languages like Chinese should benefit from the use of glyph information in those languages. However, due to lack rich pictographic evidence glyphs and weak generalization ability standard computer vision models on character data, an effective way utilize remains be found. In this paper, we address gap by presenting Glyce, glyph-vectors representations. We make three major innovations: (1) historical scripts (e.g., bronzeware script, seal...

Leilei Gan, Jiwei Li, Tianwei Zhang, Xiaoya Yuxian Meng, Fei Wu, Yi Yang, Shangwei Guo, Chun Fan. Proceedings of the 2022 Conference North American Chapter Association for Computational Linguistics: Human Language Technologies. 2022.

Though nearest neighbor Machine Translation (kNN-MT) (CITATION) has proved to introduce significant performance boosts over standard neural MT systems, it is prohibitively slow since uses the entire reference corpus as datastore for search. This means each step beam in search corpus. kNN-MT thus two-orders slower than vanilla models, making hard be applied real-world applications, especially online services. In this work, we propose Fast address issue. constructs a significantly smaller...

The frustratingly fragile nature of neural network models make current natural language generation (NLG) systems prone to backdoor attacks and generate malicious sequences that could be sexist or offensive. Unfortunately, little effort has been invested how can affect NLG defend against these attacks. In this work, by giving a formal definition attack defense, we investigate problem on two important tasks, machine translation dialog generation. Tailored the inherent (e.g., producing sequence...

With the continuous development of computer network technology, security problems in are emerging one after another, and it is becoming more difficult to ignore. For current administrators, how successfully prevent malicious hackers from invading, so that systems computers at Safe normal operation an urgent task. This paper proposes a intrusion detection method based on deep learning. uses confidence neural extract features monitoring data, BP as top level classifier classify types. The was...

The task of named entity recognition (NER) is normally divided into nested NER and flat depending on whether entities are or not. Models usually separately developed for the two tasks, since sequence labeling models, most widely used backbone NER, only able to assign a single label particular token, which unsuitable where token may be assigned several labels. In this paper, we propose unified framework that capable handling both tasks. Instead treating as problem, formulate it machine...

Existing face identity (FaceID) customization methods perform well but are limited to generating identical faces as the input, while in real-world applications, users often desire images of same person with variations, such different expressions (e.g., smiling, angry) or angles side profile). This limitation arises from lack datasets controlled input-output facial restricting models' ability learn effective modifications. To address this issue, we propose CrossFaceID, first large-scale,...

Cross-domain authentication requires that there is no trust gap between different domains can cause cross-domain devices to exceed the security control scope of original domain and further expose systems threats. In addition, as relying on traditional means built by centralized institutions cannot meet data needs in a big environment. Therefore, it necessary design secure dynamic scheme. this paper, we propose scheme (DCAGS-IoT) Internet Things environment using group signature technology...

A standard paradigm for sentiment analysis is to rely on a singular LLM and makes the decision in single round under framework of in-context learning. This suffers key disadvantage that single-turn output generated by might not deliver perfect decision, just as humans sometimes need multiple attempts get things right. especially true task where deep reasoning required address complex linguistic phenomenon (e.g., clause composition, irony, etc) input. To this issue, paper introduces multi-LLM...

In this paper, we propose a new paradigm for the task of entity-relation extraction. We cast as multi-turn question answering problem, i.e., extraction entities and relations is transformed to identifying answer spans from context. This QA formalization comes with several key advantages: firstly, query encodes important information entity/relation class want identify; secondly, provides natural way jointly modeling entity relation; thirdly, it allows us exploit well developed machine reading...

At the time of writing, ongoing pandemic coronavirus disease (COVID-19) has caused severe impacts on society, economy and people's daily lives. People constantly express their opinions various aspects social media, making user-generated content an important source for understanding public emotions concerns. In this paper, we perform a comprehensive analysis affective trajectories American people Chinese based Twitter Weibo posts between January 20th, 2020 May 11th 2020. Specifically, by...

Despite the success of ChatGPT, its performances on most NLP tasks are still well below supervised baselines. In this work, we looked into causes, and discovered that subpar performance was caused by following factors: (1) token limit in prompt does not allow for full utilization datasets; (2) mismatch between generation nature ChatGPT tasks; (3) intrinsic pitfalls LLMs models, e.g., hallucination, overly focus certain keywords, etc. propose a collection general modules to address these...