基于新词发现的跨领域中文分词方法

深度神经网络(DNN)是目前中文分词的主流方法,但将针对某一领域训练的网络模型用于其他领域时,会因存在跨领域的未登录词(OOV)和表达鸿沟而造成性能显著下降,而在实际中对所有未知领域的训练语料进行人工标注和训练模型并不可行。为了解决这个问题,该文构建了一个基于新词发现的跨领域中文分词系统,可以自动完成从目标领域语料中提取新词、标注语料和训练网络模型的工作。此外,针对现有新词发现算法提取出的词表垃圾词串多以及自动标注语料中存在噪声样本的问题,提出了一种基于向量增强互信息和加权邻接熵的无监督新词发现算法以及一种基于对抗式训练的中文分词模型。实验中将使用北大开源新闻语料训练的网络模型提取出的特征迁移到医疗、发明专利和小说领域,结果表明该文所提方法在未登录词率、准确率、召回率和分词F值方面均优于现有方法。     

An Efficient Geometric Integrator for Thermostatted Anti-/Ferromagnetic Models

(Anti)-/ferromagnetic Heisenberg spin models arise from discretization of Landau–Lifshitz models in micromagnetic modelling. In many applications it is essential to study the behavior of the system at a fixed temperature. A formulation for thermostatted spin dynamics was given by Bulgac and Kusnetsov, which incorporates a complicated nonlinear dissipation/driving term while preserving spin length. It is essential to properly model this term in simulation, and simplified schemes give poor numerical performance, e.g., requiring an excessively small timestep for stable integration. In this paper we present an efficient, structure-preserving method for thermostatted spin dynamics.     

Inhibition of domain wall formation and its effects on the bulk magnetic properties of certain spinels

It is shown that hysteresis and susceptibility studies at various temperatures could provide an idea whether a magnetic sample contains multidomain, single-domain or superparamagnetic particles. Our results on titanomagnetites with those of others on cobalt substituted magnesium ferrites suggest that domain wall formation is inhibited in some of them whereby only single-domain or superparamagnetic particles occur irrespective of the physical grain size of such materials. At high concentration of titanium in titanomagnetites, the magnetic behaviour is similar to a spin glass, which we interpret as a transition of optimum single-domains going over to superparamagnetic state.     

Compactification of Strictly Convex Real Projective Structures

In this paper we compactify the space of convex real projective structures, known as Hitchin’s component in a representation variety. We give geometric meanings to the boundary points.Mathematics Subject Classifications (2000). 51M10, 57S25.     

On the parametrization of autoregressive models by partial autocorrelations

One of the difficulties that arise in the statistical analysis of autoregressive schemes is the very complex nature of the domain of the regression parameters. In the present paper we study an alternative parametrization of autoregressive models of finite order, namely the parametrization by the partial autocorrelations. These are shown to vary freely from −1 to +1 and to be in a one-to-one, continuously differentiable correspondence with the regression parameters. Properties of the asymptotic normal distribution of the maximum likelihood estimates are discussed, and we present a new deduction of Quenouille's result on the asymptotic independence of some of the estimated partial autocorrelations.     

An algorithm of finite element grid generation for three-dimensional complex connected domain

An automated quasi three-dimensional finite element grid generation method is presented for particular three-dimensional complex connected domain, across which some are simply connected two-dimensional.regions and some are multiply connected two-dimensional regions. A subdivision algorithm based on the variational principle has been developed to ascertain the smoothness and orthogonality of the generated grid in any cross sections. Smooth transition between the simply and multiply connected regions is maintained. For illustration, the method is applied to generate a finite element three-dimensional grid for human above knee stump.     

Optical properties and g factors of GaAs quantum rods

The Hamiltonian of the zinc-blende quantum rods in the framework of eight-band effective-mass approximation in the presence of external homogeneous magnetic field is given. The electronic structure, optical properties and electron g factors of GaAs quantum rods are investigated. We found that the electron g factors are very sensitively dependent on the dimensions of the quantum rods. As some of the three dimensions increase, the electron g factors decrease. The more the dimensions increase, the more the electron g factors decrease. The dimensions perpendicular to the direction of the magnetic field affect the electron g factors more than the other dimension.