基于Prompt的两阶段澄清问题生成方法

在自然语言相关系统中,当用户输入存在歧义时,生成澄清问题询问用户有助于系统理解用户需求;基于Prompt的方法可以更好地挖掘预训练语言模型的潜在知识,但往往需要手动设计模板,限制其生成澄清问题的多样性。为解决这一问题,提出了TSCQG(two-stage clarification question generation)方法。首先,在动态Prompt模板生成阶段,利用歧义上下文和预训练语言模型生成动态的Prompt模板;然后在缺失信息生成阶段,将Prompt模板与外部知识相结合,充分利用预训练语言模型的生成能力生成相应的缺失信息。实验结果表明,在CLAQUA数据集的多轮对话情况中,BLEU值和ROUGE-L值分别达到了58.31和84.33,在ClariQ-FKw数据集上,BLEU值和ROUGE-L值分别达到了31.18和58.86。实验结果证明了TSCQG方法在澄清问题生成任务上的有效性。

GUARDIAN: A Multi-Tiered Defense Architecture for Thwarting Prompt Injection Attacks on LLMs

This paper introduces a novel multi-tiered defense architecture to protect language models from adversarial prompt attacks. We construct adversarial prompts using strategies like role emulation and manipulative assistance to simulate real threats. We introduce a comprehensive, multi-tiered defense framework named GUARDIAN (Guardrails for Upholding Ethics in Language Models) comprising a system prompt filter, pre-processing filter leveraging a toxic classifier and ethical prompt generator, and pre-display filter using the model itself for output screening. Extensive testing on Meta’s Llama-2 model demonstrates the capability to block 100% of attack prompts. The approach also auto-suggests safer prompt alternatives, thereby bolstering language model security. Quantitatively evaluated defense layers and an ethical substitution mechanism represent key innovations to counter sophisticated attacks. The integrated methodology not only fortifies smaller LLMs against emerging cyber threats but also guides the broader application of LLMs in a secure and ethical manner.

文本分类中Prompt Learning方法研究综述

文本分类是自然语言处理中的一项基础任务,在情感分析、新闻分类等领域具有重要应用。相较于传统的机器学习和深度学习模型,提示学习可以在数据不足的情况下通过构建提示来进行文本分类。近年来,GPT-3的出现推动了提示学习方法的发展,并且在文本分类领域取得了显著的进展。对以往的文本分类方法进行简要梳理,分析其存在的问题与不足;阐述了提示学习的发展进程,以及构建提示模板的方法,并对用于文本分类的提示学习方法研究及成果进行了介绍和总结。最后,对提示学习在文本分类领域的发展趋势和有待进一步研究的难点进行了总结和展望。

基于prompt tuning的中文文本多领域情感分析研究

不同领域的情感文本表达方式不一样,通常需要为各个领域训练相应的情感分析模型。针对无法用一个模型进行高效多领域情感分析的问题,提出了基于提示微调(prompt tuning)的多领域文本情感分析方法MSAPT。借助hard prompt,指示情感文本的所属领域和待选的情感标签,调动不同领域情感分析相关的知识,再为情感分析预训练一个统一的“通才模型”,在下游的各领域文本学习中,保持模型冻结,通过prompt tuning使模型学习到下游各领域情感文本的特征。MSAPT仅需保存一个模型和一些参数量远远小于模型的prompt,实现了多领域情感分析。在多个属于不同领域的情感文本数据集上进行实验,结果表明仅进行prompt tuning时,MSAPT效果优于模型微调(model tuning)的。最后,分别对适应特定领域的prompt tuning、hard prompt、soft prompt的长度和中间训练数据集的大小进行消融实验,从证明其对情感分析效果的影响。

基于prompt和知识增强的方面级情感分析

方面级情感分析是一种新兴的细粒度情感分析任务,旨在根据给定句子和方面词判断情感极性。目前广泛使用的预训练语言模型由于训练目标和方面级情感分析的目标有差异,分析结果不好。为了缓解预训练语言模型和情感分析目标的差异,prompt被引入到方面级情感分析中,采用伪标签加方面词和意见词的方式创建prompt连续模板,并使用prompt-encoder训练伪标签使其拥有语义信息;然后,使用主题图注意力机制融合关于方面词和意见词的外部知识,根据融合外部知识的隐藏向量预测由情感词典组成的候选标签词;最后,采用求和置信度分数的方式将候选标签词的概率映射到情感极性分布空间上。实验表明,该模型在SemEval 2014任务的笔记本电脑数据集和餐厅数据集上将正确率分别提高了1.53%和3.5%。

融合多Prompt模板的零样本关系抽取模型

Prompt范式被广泛应用于零样本的自然语言处理(NLP)任务中,但是现有基于Prompt范式的零样本关系抽取(RE)模型存在答案空间映射难构造与模板选择依赖人工的问题,无法取得较好的效果。针对这些问题,提出一种融合多Prompt模板的零样本RE模型。首先,将零样本RE任务定义为掩码语言模型(MLM)任务,舍弃答案空间映射的构造,将模板输出的词与关系描述文本在词向量空间中进行比较,以此判断关系类别;其次,引入待抽取关系类别的描述文本的词性作为特征,学习该特征与各个模板之间的权重;最后,利用该权重融合多个模板输出的结果,以此减少人工选取的Prompt模板引起的性能损失。在FewRel(Few-shot Relation extraction dataset)和TACRED(Text Analysis Conference Relation Extraction Dataset)这两个数据集上的实验结果显示,与目前最优的模型RelationPrompt相比,所提模型在不同数据资源设置下,F1值分别提升了1.48~19.84个百分点和15.27~15.75个百分点。可见,所提模型在零样本RE任务上取得了显著的效果提升。

基于Prompt学习的无监督关系抽取模型

无监督关系抽取旨在从无标签的自然语言文本中抽取实体之间的语义关系。目前,基于变分自编码器(VAE)架构的无监督关系抽取模型通过重构损失提供监督信号来训练模型,这为完成无监督关系抽取任务提供了新思路。针对此类模型无法有效地理解上下文信息、依赖数据集归纳偏置的问题,提出基于Prompt学习的无监督关系抽取(PURE)模型,其中包括关系抽取和链接预测两个模块。在关系抽取模块中设计了上下文感知的Prompt模板函数以融入上下文信息,并将无监督关系抽取任务转换为掩码预测任务,从而充分利用预训练阶段获得的知识完成关系抽取。在链接预测模块中则通过预测关系三元组中的缺失实体提供监督信号联合训练两个模块。在两个公开真实关系抽取数据集上进行了大量实验,得到的结果表明PURE模型能有效利用上下文信息并且不依赖数据集归纳偏置,相较于目前最优的基于VAE架构的模型UREVA(Variational Autoencoder-based Unsupervised Relation Extraction model)在NYT数据集上的B-cubed F1指标上提升了3.3个百分点。

基于Prompt和BERT的情感原因对抽取方法

情感原因对抽取是一项全新的自然语言处理任务,试图提取出文本中所有的情感子句以及相应的原因子句。现有的工作要么遵循多阶段方法,其中情感抽取、原因抽取及配对都是单独进行的;要么使用复杂的架构来解决其局限性,并且可解释性较低。针对这些问题,本文提出一种基于Prompt和BERT相结合的ECPE-Prompt模型。该模型以一种可解释性较强的两阶段方法将情感抽取和原因抽取链接起来,最后完成情感原因对抽取任务。实验结果表明,该模型的F1分数指标对比基线模型提高6%。

基于prompt的文本可读性评估

为了提高预训练模型在文本可读性分类中的表现,提出基于预训练模型BERT(bidirectional encoder representations from transformers)和GPT2(generative pre-training 2)的prompt文本可读性分类方法,通过手工模板和映射器将文本可读性分类转换为与模型预训练任务一致的文本填空形式,以此提高预训练模型的分类性能.实验表明,prompt方法在实验数据集上与传统机器学习和有监督的神经网络相比具有更好的分类性能.

基于PROMPT的本体映射实例分析

本体语义互操作可以有效促进相关本体基于语义的互联互通。目前,异构本体的语义互操作主要是通过本体映射来实现。本文以医学领域中的公共卫生领域疫情监测本体与医学领域顶层本体（BioCaster和UMLS Semantic Network）为研究案例,通过基于字符串匹配与锚遍历相结合的PROMPT系统,探讨两个本体的语义相似性,旨在从中发现两个医学本体间的相似性特点,为医药卫生领域本体语义互操作解决方案的优化提供一定的实践基础,从而促进语义网环境下医药卫生信息资源乃至其他领域中信息资源一体化。

图书馆网站易用性研究--基于A-Prompt软件的分析

论文分析了现阶段网站设计的研究方向,阐述了网站易用性研究的重要性及必要性;并以南京图书馆主页为研究对象,利用A-Prompt软件分析了该网页在网站易用性上存在的问题,同时提供了修改建议与方法。

使用PROMPT命令与DOS ANSI.SYS设置屏幕

在 DOS 中,有一个常常忽略的 PROMPT。一般 DOS 手册对 PROMPT 命令的介绍大多限于用它来设置(改变)系统的提示符,命令格式为PROMPT[string][$X]其中的 string 是任意的字符串,通过它来指定将要包含于(或作为)系统提示符的简单信息。例如运行命令'PROMPT*

Study on element detection and its correction in iron ore concentrate based on a prompt gamma-neutron activation analysis system

A prompt gamma-neutron activation analysis(PGNAA) system was developed to detect the iron content of iron ore concentrate. Because of the self-absorption effect of gamma-rays and neutrons, and the interference of chlorine in the neutron field, the linear relationship between the iron analytical coefficient and total iron content was poor, increasing the error in the quantitative analysis. To solve this problem, gamma-ray self-absorption compensation and a neutron field correction algorithm were proposed, and the experimental results have been corrected using this algorithm. The results show that the linear relationship between the iron analytical coefficient and total iron content was considerably improved after the correction. The linear correlation coefficients reached 0.99 or more.

Measurement of the prompt neutron spectrum from thermalneutron-induced fission in U-235 using the recoil proton method

A measurement of the ^235U prompt fission neutron spectrum (PFNS) by the recoil proton method was performed at the Institute of Nuclear Physics and Chemistry, China. Details of the method, which include the calculation and validation of the response matrix, are presented. The PFNS for ^235U in the energy range 1–12 MeV, induced by thermal neutrons, was obtained. The measured spectrum in the low-energy region was in good agreement with previous work and the ENDF/B-VII library, except for minor differences. In the high-energy region, however, the relative height of the measured spectrum was greater, and an analysis of the experiment indicated uncertainties of 13% at 10 MeV and 24% at 12 MeV. Experimental results showed that the recoil proton method could be used to measure prompt fission neutron spectra. Some directions for future work are included.

Analysis of influencing factors on the method for determining boron concentration and dose through dual prompt gamma detection

A method based on the cross-sectional relationship between^(10)B(n,α)^(7)Li and^(1)H(n,γ)^(2)H was proposed to detect and reconstruct the three-dimensional boron concentration/dose distribution in vivo during boron neutron capture therapy(BNCT).Factors such as the neutron energy,fluence rate,and degree of non-uniform distribution of the boron concentration in a voxel may affect the results of this method.A theoretical analysis of the accuracy of the method using a Monte Carlo simulation shows that the determining error is generally less than 1%under different tumor locations and neutron source configurations.When the voxel size is larger than 0.4 cm,the determining error might be higher for a non-uniformly distributed boron concentration in the voxel because of the changes in the neutron energy and fluence rate.In conclusion,the proposed method enables an accurate threedimensional boron determination in vivo during BNCT.

Analysis of prompt fission neutron spectrum and multiplicity for ^(237)Np(n, f) in the frame of multi-modal Los Alamos model

The improved version of Los Alamos model with the multi-modal fission approach is used to analyse the prompt fission neutron spectrum and multiplicity for the neutron-induced fission of 237Np. The spectra of neutrons emitted from fragments for the three most dominant fission modes （standard Ⅰ, standard Ⅱ and superlong） are calculated separately and the total spectrum is synthesized. The multi-modal parameters contained in the spectrum model are determined on the basis of experimental data of fission fragment mass distributions. The calculated total prompt fission neutron spectrum and multiplicity are better agreement with the experimental data than those obtained from the conventional treatment of the Los Alamos model.

Calculation of Prompt Fission Neutron from ^(233)U(n, f) Reaction byMulti-Modal Los Alamos Model

An attempt is made to improve the evaluation of the prompt fission neutron emis- sion from 233U（n, f） reaction for incident neutron energies below 6 MeV. The multi-modal fission approach is applied to the improved version of Los Alamos model and the point by point model. The prompt fission neutron spectra and the prompt fission neutron as a function of fragment mass （usually named ＂sawtooth＂ data） v（A） are calculated independently for the three most dominant fission modes （standard I, standard II and superlong）, and the total spectra and v（A） are syn- thesized. The multi-modal parameters are determined on the basis of experimental data of fission fragment mass distributions. The present calculation results can describe the experimental data very well, and the proposed treatment is thus a useful tool for prompt fission neutron emission prediction.

DEVELOPMENT AND APPLICATIONS OF AN ON-LINE THERMAL NEUTRON PROMPT-GAMMA ELEMENT ANALYSIS SYSTEM

This paper introduces the principles, instrumentation, implementation, and industrial applications of an on- line thermal neutron prompt- gamma element analysis system (using a 252Cf neutron source, Am- Be neutron source, or neutron generator). The energy resolution of the system at the H prompt- gamma full- energy photopeak (2.22325 MeV) is 3.6 keV. The concentration measurement error of A12O3, Fe2O3, CaO and SiO2 is ±0.3%,±0.1%.±0.4% and ±0.4%, respectively.The system has been tested on- site at both the Shandong and the Zhengzhou Aluminum Works. Our preliminary on- site measurements confirm that the stability, reliability, measurement range, and accuracy of the system can meet the requirements of the aluminum production process. Facilitation of this measurement at aluminum plants is expected to reduce plant costs by over 3 million dollars annually through reduced energy consumption, more rapid qualification of pulps being mixed during the production process, and in reduced labor costs.

3M Adper Prompt自酸蚀粘结剂治疗牙本质过敏的近期疗效观察

目的了解3MAdperPrompt治疗牙本质过敏的疗效。方法于2004年11月-2005年3月间采用3MAdperPrompt自酸蚀粘结剂与75%氟化钠甘油糊剂给左右对称的同名牙进行脱敏对比观察。将确诊为牙本质过敏症的42例,共51对牙,随机分为两组。治疗组用3MAdperPrompt粘结剂,对照组用75%氟化钠甘油,在统一的疗效标准下进行即刻、一周、1个月、3个月四时段的疗效对比。结果第一,二时段的疗效二者无显著性差异。而第三,四时段的疗效3MAdperPrompt明显优于氟化钠甘油糊剂。结论3MAdperPrompt是一种比较理想的脱敏剂。

用@…PROMPT设计内容是灵活多变的菜单

对foxbase+中传统光带菜单做改进,利用数据库文件和数组的特性,实现在不改变源程序的前提下,可任意修改显示菜单内容,适用于FOXBASE+程序设计中不改变软件功能仅改变菜单中名称的软件.