采用自然语言的移动机器人任务编程

本文在机器人具备基本运动技能的基础上［1］，采用基于指令教导的学习方法．通 过自然语言教会机器人完成抽象化任务，并以程序体方式保存所学知识，也即通过自然语言 对话自动生成程序流．通过让机器人完成导航等任务，验证所提自然语言编程方法的可行性 ．     

A Model to Disambiguate Natural Language Parses on the Basis of User Language Proficiency: Design and Evaluation

This paper discusses the design and evaluation of an implemented user model in ICICLE, an instruction system for users writing in a second language. We show that in the task of disambiguating natural language parses, a blended model combining overlay techniques with user stereotyping representing typical linguistic acquisition sequences captures user individuality while supplementing incomplete information with stereotypic reasoning     

基于强化学习的对抗预训练语言建模方法

在大规模无监督语料上的BERT、XLNet等预训练语言模型,通常采用基于交叉熵损失函数的语言建模任务进行训练。模型的评价标准则采用困惑度或者模型在其他下游自然语言处理任务中的性能指标,存在损失函数和评测指标不匹配等问题。为解决这些问题,该文提出一种结合强化学习的对抗预训练语言模型RL-XLNet(Reinforcement Learning-XLNet)。RL-XLNet采用对抗训练方式训练一个生成器,基于上下文预测选定词,并训练一个判别器判断生成器预测的词是否正确。通过对抗网络生成器和判别器的相互促进作用,强化生成器对语义的理解,提高模型的学习能力。由于在文本生成过程中存在采样过程,导致最终的损失无法直接进行回传,故提出采用强化学习的方式对生成器进行训练。基于通用语言理解评估基准(GLUE Benchmark)和斯坦福问答任务(SQuAD 1.1)的实验,结果表明,与现有BERT、XLNet方法相比,RL-XLNet模型在多项任务中的性能上表现出较明显的优势: 在GLUE的六个任务中排名第1,一个任务排名第2,一个任务排名第3。在SQuAD 1.1任务中F₁值排名第1。考虑到运算资源有限,基于小语料集的模型性能也达到了领域先进水平。     

基于混合深度学习的多模态场景指令分类方法

为提高家庭服务机器人指令中目标对象预测的准确率,提出一种基于混合深度学习的多模态自然语言理处理(Natural Language Processing,NLP)指令分类方法.该方法从语言特征、视觉特征和关系特征多模态入手,采用两种深度学习方法分别以多模态特征进行编码.对于语言指令,采用多层双向长短期记忆(Bi-LSTM...     

基于信息增强BERT的关系分类

关系分类是自然语言处理领域中重要的语义处理任务,随着机器学习技术的发展,预训练模型BERT在多项自然语言处理任务中取得了大量研究成果,但在关系分类领域尚有待探索。该文针对关系分类的问题特点,提出一种基于实体与实体上下文信息增强BERT的关系分类方法(EC_BERT),该方法利用BERT获取句子特征表示向量,并结合两个目标实体以及实体上下文语句信息,送入简单神经网络进行关系分类。此外,该文还对BERT的改进模型RoBERTa、DistilBERT进行了实验,发现BERT对于关系分类能力更为突出。实验结果显示,该方法在SemEval-2010任务8数据集和KBP-37数据集上Macro-F₁值最高取得了89.69%和65.92%的结果,与以往方法相比,其在关系分类任务上表现出较好的性能。     

Effect of symmetry bias on linguistic evolution

As described in this paper, we investigated the effect of the symmetry bias on linguistic evolution. We specifically examined symmetry bias, which indicates the meaning in a state of environment. For this task, we constructed a meaning selection iterated learning model based on Simon Kirby’s iterated learning Model, and used it for simulation with three strategies: perfect matching symmetry bias, imperfect matching symmetry bias, and random strategy. Results of applying imperfect matching symmetry bias show that the language of the agent evolved into more compositional language. The agent acquired a more expressive, and a more similar language to the parent’s language than with the Random strategy agent. However, application of perfect matching symmetry bias showed that the language of the agent did not evolve. The agent acquired a less expressive and a more different language to the parent’s language than with Random strategy agent. Our experimentally obtained results demonstrate that the effect of imperfect matching symmetry bias accelerates linguistic evolution into compositional language, whereas perfect matching symmetry bias disturbs linguistic evolution.     

Modeling robotic operations controlled by natural language

There are multiple ways to control a robotic system. Most of them require the users to have prior knowledge about robots or get trained before using them. Natural language based control attracts increasing attention due to its versatility and less requirements for users. Since natural language instructions from users cannot be understood by the robots directly, the linguistic input has to be processed into a formal representation which captures the task specification and removes the ambiguity inherent in natural language. For most of existing natural language controlled robotic system, they assume the given language instructions are already in correct orders. However, it is very likely for untrained users to give commands in a mixed order based on their direct observation and intuitive thinking. Simply following the order of the commands can lead to failures of tasks. To provide a remedy for the problem, we propose a novel framework named dependency relation matrix (DRM) to model and organize the semantic information extracted from language input, in order to figure out an executable sequence of subtasks for later execution. In addition, the proposed approach projects abstract language input and detailed sensory information into the same space, and uses the difference between the goal specification and temporal status of the task under implementation to monitor the progress of task execution. In this paper, we describe the DRM framework in detail, and illustrate the utility of this approach with experiment results.     

MTL-BERT:一种结合BERT的中文文本多任务学习模型

单任务学习常常受限于单目标函数的不足,多任务学习能有效利用任务相关性的先验性,故而受到了学界的关注.在中文自然语言处理领域,关于多任务学习的研究极为匮乏,该领域需同时考虑到中文文本特征提取和多任务的建模.本论文提出了一种多任务学习模型MTL-BERT.首先将BERT作为特征提取器以提升模型的泛化性.其次分类和回归是机器学习中的两个主要问题,针对多标签分类和回归的混合任务,提出了一种任务权重自适应框架.该框架下,任务之间的权重由联合模型参数共同训练.最后从模型最大似然角度,理论验证了该多任务学习算法的有效性.在真实中文数据集上的实验表明,MTL-BERT具有较好的计算效果.     

Interpreting and Extracting Open Knowledge for Human-Robot Interaction

A more natural way for non-expert users to express their tasks in an open-ended set is to use natural language. In this case, a human-centered intelligent agent/robot is required to be able to understand and generate plans for these naturally expressed tasks. For this purpose, it is a good way to enhance intelligent robot's abilities by utilizing open knowledge extracted from the web, instead of hand-coded knowledge. A key challenge of utilizing open knowledge lies in the semantic interpretation of the open knowledge organized in multiple modes, which can be unstructured or semi-structured, before one can use it. Previous approaches used a limited lexicon to employ combinatory categorial grammar (CCG) as the underlying formalism for semantic parsing over sentences. Here, we propose a more effective learning method to interpret semi-structured user instructions. Moreover, we present a new heuristic method to recover missing semantic information from the context of an instruction. Experiments showed that the proposed approach renders significant performance improvement compared to the baseline methods and the recovering method is promising.      

Effectiveness of automatic and expert generated narrative and guided instructions for task-oriented web browsing

This study investigated how instructions may be created to facilitate web browsing tasks. Two types of instructions were considered. Narrative instructions are text-based while guided instructions use graphic annotations. One way to create both types of instructions is to utilize the specialty of human experts. A method is also provided for automatic generation of both types of instructions based on the learning of user experience in web browsing. An experiment was conducted to test the effect of types of instruction, sources of instruction, and task complexity setting on performance variables in web browsing. The results of the experiment indicated that (1) by using web annotations, guided instructions resulted in better performance and satisfactions than narrative instructions in web browsing, (2) based on learning of web browsing activities, automatically generated guided instructions are comparable to expert-created guided instructions in terms of their effect on the performance of web browsing.     

融合角色、结构和语义的口语对话预训练语言模型

口语语言理解是任务式对话系统的重要组件,预训练语言模型在口语语言理解中取得了重要突破,然而这些预训练语言模型大多是基于大规模书面文本语料。考虑到口语与书面语在结构、使用条件和表达方式上的明显差异,构建了大规模、双角色、多轮次、口语对话语料,并提出融合角色、结构和语义的四个自监督预训练任务：全词掩码、角色预测、话语内部反转预测和轮次间互换预测,通过多任务联合训练面向口语的预训练语言模型SPD-BERT（SPoken Dialog-BERT）。在金融领域智能客服场景的三个人工标注数据集——意图识别、实体识别和拼音纠错上进行详细的实验测试,实验结果验证了该语言模型的有效性。     

Sense representations for Portuguese: experiments with sense embeddings and deep neural language models

Sense representations have gone beyond word representations like Word2Vec, GloVe and FastText and achieved innovative performance on a wide range of natural language processing tasks. Although very useful in many applications, the traditional approaches for generating word embeddings have a strict drawback: they produce a single vector representation for a given word ignoring the fact that ambiguous words can assume different meanings. In this paper, we explore unsupervised sense representations which, different from traditional word embeddings, are able to induce different senses of a word by analyzing its contextual semantics in a text. The unsupervised sense representations investigated in this paper are: sense embeddings and deep neural language models. We present the first experiments carried out for generating sense embeddings for Portuguese. Our experiments show that the sense embedding model (Sense2vec) outperformed traditional word embeddings in syntactic and semantic analogies task, proving that the language resource generated here can improve the performance of NLP tasks in Portuguese. We also evaluated the performance of pre-trained deep neural language models (ELMo and BERT) in two transfer learning approaches: feature based and fine-tuning, in the semantic textual similarity task. Our experiments indicate that the fine tuned Multilingual and Portuguese BERT language models were able to achieve better accuracy than the ELMo model and baselines.

     

基于深度学习的机器阅读理解研究综述

机器阅读理解任务在近年来备受关注,它赋予计算机从文本数据中获取知识和回答问题的能力。如何让机器理解自然语言是人工智能领域长期存在的挑战之一,近年来大规模高质量数据集的发布和深度学习技术的运用,使得机器阅读理解取得了快速发展。基于神经网络的端到端的模型结构,基于预训练语言模型以及推理技术的应用,其性能在大规模评测数据集上有很大提升,但距离真正的理解语言还有较大差距。本文对机器阅读理解任务的研究现状与发展趋势进行了综述,主要包括任务划分、机器阅读理解模型与相关技术的分析,特别是基于知识推理的机器阅读理解技术,总结并讨论了该领域的发展趋势。     

结合五笔字形与上下文相关字向量的命名实体识别

命名实体识别（NER）作为自然语言处理的重要部分,在信息抽取和知识图谱等任务中得到广泛应用。然而目前中文预训练语言模型通常仅对上下文中的字符进行建模,忽略了中文字符的字形结构。提出2种结合五笔字形的上下文相关字向量表示方法,以增强字向量的语义表达能力。第一种方法分别对字符和字形抽取特征并联合建模得到字向量表示,第二种方法将五笔字形作为辅助信息拼接到字向量中,训练一个基于字符和五笔字形的混合语言模型。实验结果表明,所提两种方法可以有效提升中文NER系统的性能,且结合五笔字形的上下文相关字向量表示方法的系统性能优于基于单一字符的语言模型。     

预训练语言模型实体匹配的可解释性

实体匹配可以判断两个数据集中的记录是否指向同一现实世界实体,对于大数据集成、社交网络分析、网络语义数据管理等任务不可或缺.作为在自然语言处理、计算机视觉中取得大量成功的深度学习技术,预训练语言模型在实体识别任务上也取得了优于传统方法的效果,引起了大量研究人员的关注.然而,基于预训练语言模型的实体匹配技术效果不稳定、匹配结果不可解释,给这一技术在大数据集成中的应用带来了很大的不确定性.同时,现有的实体匹配模型解释方法主要面向机器学习方法进行模型无关的解释,在预训练语言模型上的适用性存在缺陷.因此,以Ditto、JointBERT等BERT类实体匹配模型为例,提出3种面向预训练语言模型实体匹配技术的模型解释方法来解决这个问题:(1)针对序列化操作中关系数据属性序的敏感性,对于错分样本,利用数据集元特征和属性相似度实现属性序反事实生成;(2)作为传统属性重要性衡量的补充,通过预训练语言模型注意力机制权重来衡量并可视化模型处理数据时的关联性;(3)基于序列化后的句子向量,使用k近邻搜索技术召回与错分样本相似的可解释性优良的样本,增强低置信度的预训练语言模型预测结果.在真实公开数据集上的实验结果...     

Learning Local Transductions Is Hard

Local deterministic string-to-string transductions arise in natural language processing (NLP) tasks such as letter-to-sound translation or pronunciation modeling. This class of transductions is a simple generalization of morphisms of free monoids; learning local transductions is essentially the same as inference of certain monoid morphisms. However, learning even a highly restricted class of morphisms, the so-called fine morphisms, leads to intractable problems: deciding whether a hypothesized fine morphism is consistent with observations is an NP-complete problem; and maximizing classification accuracy of the even smaller class of alphabetic substitution morphisms is APX-hard. These theoretical results provide some justification for using the kinds of heuristics that are commonly used for this learning task.     

知识感知的预训练语言模型综述

随着自然语言处理（NLP）领域中预训练技术的快速发展,将外部知识引入到预训练语言模型的知识驱动方法在NLP任务中表现优异,知识表示学习和预训练技术为知识融合的预训练方法提供了理论依据。概述目前经典预训练方法的相关研究成果,分析在新兴预训练技术支持下具有代表性的知识感知的预训练语言模型,分别介绍引入不同外部知识的预训练语言模型,并结合相关实验数据评估知识感知的预训练语言模型在NLP各个下游任务中的性能表现。在此基础上,分析当前预训练语言模型发展过程中所面临的问题和挑战,并对领域发展前景进行展望。     

视觉语言预训练综述

近年来深度学习在计算机视觉(CV)和自然语言处理(NLP)等单模态领域都取得了十分优异的性能.随着技术的发展,多模态学习的重要性和必要性已经慢慢展现.视觉语言学习作为多模态学习的重要部分,得到国内外研究人员的广泛关注.得益于Transformer框架的发展,越来越多的预训练模型被运用到视觉语言多模态学习上,相关任务在性能上得到了质的飞跃.系统地梳理了当前视觉语言预训练模型相关的工作,首先介绍了预训练模型的相关知识,其次从两种不同的角度分析比较预训练模型结构,讨论了常用的视觉语言预训练技术,详细介绍了5类下游预训练任务,最后介绍了常用的图像和视频预训练任务的数据集,并比较和分析了常用预训练模型在不同任务下不同数据集上的性能.     

Balanced Cooperative Modeling

Machine learning techniques are often used for supporting a knowledge engineer in constructing a model of part of the world. Different learning algorithms contribute to different tasks within the modeling process. Integrating several learning algorithms into one system allows it to support several modeling tasks within the same framework. In this article, we focus on the distribution of work between several learning algorithms on the one hand and the user on the other hand. The approach followed by the MOBAL system is that ofbalanced cooperation, i.e., each modeling task can be done by the user or by a learning tool of the system. The MOBAL system is described in detail. We discuss the principle of multi-functionality of one representation for the balanced use by learning algorithms and users.     

融合预训练语言模型的成语完形填空算法

根据上下文语境选择恰当的成语,是自然语言处理领域的重要任务之一.现有的研究将成语完型填空任务看成是文本匹配问题,虽然预训练语言模型能够在文本匹配研究上取得较高的准确率,但也存在明显的缺陷:一方面,预训练语言模型作为特征提取器时,会丢失句子间相互信息;另一方面,预训练语言模型作为文本匹配器时,计算开销大,训练时间和推理时间较长.另外,上下文与候选成语之间的匹配是不对称的,会影响预训练语言模型发挥匹配器的效果.为了解决上述两个问题,利用参数共享的思想,提出了TALBERT-blank.TALBERT-blank是将成语选择从基于上下文的不对称匹配过程转换为填空与候选答案之间的对称匹配过程,将预训练语言模型同时作为特征提取器和文本匹配器,并对句向量作潜在语义匹配.这样可以减少参数量和内存的消耗,在保持准确度的情况下,提高了训练和推理速度,达到了轻量高效的效果.在CHID数据集上的实验结果表明:作为匹配器,TALBERT-blank相较于ALBERT,在保证准确率的情况下,更大限度地精简了模型的结构,计算时间进一步缩短54.35%.