Searcharxiv⌕ Search

arXiv subjects

Lizhi Cheng

Publications and source records attributed to Lizhi Cheng.

At least 19 recordsLinked to original sources
arXiv2022

A Scope Sensitive and Result Attentive Model for Multi-Intent Spoken Language Understanding

Multi-Intent Spoken Language Understanding (SLU), a novel and more complex scenario of SLU, is attracting increasing attention. Unlike traditional SLU, each intent in this scenario has its specific scope. Semantic information outside the scope even hinders the prediction, which tremendously increases the difficulty of intent detection. More seriously, guiding slot filling with these inaccurate intent labels suffers error propagation problems, resulting in unsatisfied overall performance. To solve these challenges, in this paper, we propose a novel Scope-Sensitive Result Attention Network (SSRAN) based on Transformer, which contains a Scope Recognizer (SR) and a Result Attention Network (RAN). Scope Recognizer assignments scope information to each token, reducing the distraction of out-of-scope tokens. Result Attention Network effectively utilizes the bidirectional interaction between results of slot filling and intent detection, mitigating the error propagation problem. Experiments on two public datasets indicate that our model significantly improves SLU performance (5.4\% and 2.1\% on Overall accuracy) over the state-of-the-art baseline.

cs.CL↗
arXiv2022

Capture Salient Historical Information: A Fast and Accurate Non-Autoregressive Model for Multi-turn Spoken Language Understanding

Spoken Language Understanding (SLU), a core component of the task-oriented dialogue system, expects a shorter inference facing the impatience of human users. Existing work increases inference speed by designing non-autoregressive models for single-turn SLU tasks but fails to apply to multi-turn SLU in confronting the dialogue history. The intuitive idea is to concatenate all historical utterances and utilize the non-autoregressive models directly. However, this approach seriously misses the salient historical information and suffers from the uncoordinated-slot problems. To overcome those shortcomings, we propose a novel model for multi-turn SLU named Salient History Attention with Layer-Refined Transformer (SHA-LRT), which composes of an SHA module, a Layer-Refined Mechanism (LRM), and a Slot Label Generation (SLG) task. SHA captures salient historical information for the current dialogue from both historical utterances and results via a well-designed history-attention mechanism. LRM predicts preliminary SLU results from Transformer's middle states and utilizes them to guide the final prediction, and SLG obtains the sequential dependency information for the non-autoregressive encoder. Experiments on public datasets indicate that our model significantly improves multi-turn SLU performance (17.5% on Overall) with accelerating (nearly 15 times) the inference process over the state-of-the-art baseline as well as effective on the single-turn SLU tasks.

cs.CL↗
arXiv2022

A Parallel Direct Eigensolver for Sequences of Hermitian Eigenvalue Problems with No Tridiagonalization

In this paper, a Parallel Direct Eigensolver for Sequences of Hermitian Eigenvalue Problems with no tridiagonalization is proposed, denoted by \texttt{PDESHEP}, and it combines direct methods with iterative methods. \texttt{PDESHEP} first reduces a Hermitian matrix to its banded form, then applies a spectrum slicing algorithm to the banded matrix, and finally computes the eigenvectors of the original matrix via backtransform. Therefore, compared with conventional direct eigensolvers, \texttt{PDESHEP} avoids tridiagonalization, which consists of many memory-bounded operations. In this work, the iterative method in \texttt{PDESHEP} is based on the contour integral method implemented in FEAST. The combination of direct methods with iterative methods for banded matrices requires some efficient data redistribution algorithms both from 2D to 1D and from 1D to 2D data structures. Hence, some two-step data redistribution algorithms are proposed, which can be $10\times$ faster than ScaLAPACK routine \texttt{PXGEMR2D}. For the symmetric self-consistent field (SCF) eigenvalue problems, \texttt{PDESHEP} can be on average $1.25\times$ faster than the state-of-the-art direct solver in ELPA when using $4096$ processes. Numerical results are obtained for dense Hermitian matrices from real applications and large real sparse matrices from the SuiteSparse collection.

math.NA↗
arXiv2021

An Effective Non-Autoregressive Model for Spoken Language Understanding

Spoken Language Understanding (SLU), a core component of the task-oriented dialogue system, expects a shorter inference latency due to the impatience of humans. Non-autoregressive SLU models clearly increase the inference speed but suffer uncoordinated-slot problems caused by the lack of sequential dependency information among each slot chunk. To gap this shortcoming, in this paper, we propose a novel non-autoregressive SLU model named Layered-Refine Transformer, which contains a Slot Label Generation (SLG) task and a Layered Refine Mechanism (LRM). SLG is defined as generating the next slot label with the token sequence and generated slot labels. With SLG, the non-autoregressive model can efficiently obtain dependency information during training and spend no extra time in inference. LRM predicts the preliminary SLU results from Transformer's middle states and utilizes them to guide the final prediction. Experiments on two public datasets indicate that our model significantly improves SLU performance (1.5\% on Overall accuracy) while substantially speed up (more than 10 times) the inference process over the state-of-the-art baseline.

cs.CL↗
arXiv2021

A Result based Portable Framework for Spoken Language Understanding

Spoken language understanding (SLU), which is a core component of the task-oriented dialogue system, has made substantial progress in the research of single-turn dialogue. However, the performance in multi-turn dialogue is still not satisfactory in the sense that the existing multi-turn SLU methods have low portability and compatibility for other single-turn SLU models. Further, existing multi-turn SLU methods do not exploit the historical predicted results when predicting the current utterance, which wastes helpful information. To gap those shortcomings, in this paper, we propose a novel Result-based Portable Framework for SLU (RPFSLU). RPFSLU allows most existing single-turn SLU models to obtain the contextual information from multi-turn dialogues and takes full advantage of predicted results in the dialogue history during the current prediction. Experimental results on the public dataset KVRET have shown that all SLU models in baselines acquire enhancement by RPFSLU on multi-turn SLU tasks.

cs.CL↗
arXiv2019

Bregman Proximal Gradient Algorithm with Extrapolation for a class of Nonconvex Nonsmooth Minimization Problems

In this paper, we consider an accelerated method for solving nonconvex and nonsmooth minimization problems. We propose a Bregman Proximal Gradient algorithm with extrapolation(BPGe). This algorithm extends and accelerates the Bregman Proximal Gradient algorithm (BPG), which circumvents the restrictive global Lipschitz gradient continuity assumption needed in Proximal Gradient algorithms (PG). The BPGe algorithm has higher generality than the recently introduced Proximal Gradient algorithm with extrapolation(PGe), and besides, due to the extrapolation step, BPGe converges faster than BPG algorithm. Analyzing the convergence, we prove that any limit point of the sequence generated by BPGe is a stationary point of the problem by choosing parameters properly. Besides, assuming Kurdyka-{Ĺ}ojasiewicz property, we prove the whole sequences generated by BPGe converges to a stationary point. Finally, to illustrate the potential of the new method BPGe, we apply it to two important practical problems that arise in many fundamental applications (and that not satisfy global Lipschitz gradient continuity assumption): Poisson linear inverse problems and quadratic inverse problems. In the tests the accelerated BPGe algorithm shows faster convergence results, giving an interesting new algorithm.

math.OC↗
arXiv2019

Training GANs with Centripetal Acceleration

Training generative adversarial networks (GANs) often suffers from cyclic behaviors of iterates. Based on a simple intuition that the direction of centripetal acceleration of an object moving in uniform circular motion is toward the center of the circle, we present the Simultaneous Centripetal Acceleration (SCA) method and the Alternating Centripetal Acceleration (ACA) method to alleviate the cyclic behaviors. Under suitable conditions, gradient descent methods with either SCA or ACA are shown to be linearly convergent for bilinear games. Numerical experiments are conducted by applying ACA to existing gradient-based algorithms in a GAN setup scenario, which demonstrate the superiority of ACA.

cs.LG↗
arXiv2018

A linear time algorithm for multiscale quantile simulation

Change-point problems have appeared in a great many applications for example cancer genetics, econometrics and climate change. Modern multiscale type segmentation methods are considered to be a statistically efficient approach for multiple change-point detection, which minimize the number of change-points under a multiscale side-constraint. The constraint threshold plays a critical role in balancing the data-fit and model complexity. However, the computation time of such a threshold is quadratic in terms of sample size $n$, making it impractical for large scale problems. In this paper we proposed an $\mathcal{O}(n)$ algorithm by utilizing the hidden quasiconvexity structure of the problem. It applies to all regression models in exponential family with arbitrary convex scale penalties. Simulations verify its computational efficiency and accuracy. An implementation is provided in R-package "linearQ" on CRAN.

stat.CO↗
arXiv2017

A convergence framework for inexact nonconvex and nonsmooth algorithms and its applications to several iterations

In this paper, we consider the convergence of an abstract inexact nonconvex and nonsmooth algorithm. We promise a pseudo sufficient descent condition and a pseudo relative error condition, which are both related to an auxiliary sequence, for the algorithm; and a continuity condition is assumed to hold. In fact, a lot of classical inexact nonconvex and nonsmooth algorithms allow these three conditions. Under a special kind of summable assumption on the auxiliary sequence, we prove the sequence generated by the general algorithm converges to a critical point of the objective function if being assumed Kurdyka- Lojasiewicz property. The core of the proofs lies in building a new Lyapunov function, whose successive difference provides a bound for the successive difference of the points generated by the algorithm. And then, we apply our findings to several classical nonconvex iterative algorithms and derive the corresponding convergence results

math.OC↗
arXiv2017

Iteratively Linearized Reweighted Alternating Direction Method of Multipliers for a Class of Nonconvex Problems

In this paper, we consider solving a class of nonconvex and nonsmooth problems frequently appearing in signal processing and machine learning research. The traditional alternating direction method of multipliers encounters troubles in both mathematics and computations in solving the nonconvex and nonsmooth subproblem. In view of this, we propose a reweighted alternating direction method of multipliers. In this algorithm, all subproblems are convex and easy to solve. We also provide several guarantees for the convergence and prove that the algorithm globally converges to a critical point of an auxiliary function with the help of the Kurdyka-{\L}ojasiewicz property. Several numerical results are presented to demonstrate the efficiency of the proposed algorithm.

math.NA↗
arXiv2016

Accurate Quotient-Difference algorithm: error analysis, improvements and applications

The compensated quotient-difference (Compqd) algorithm is proposed along with some applications. The main motivation is based on the fact that the standard quotient-difference (qd) algorithm can be numerically unstable. The Compqd algorithm is obtained by applying error-free transformations to improve the traditional qd algorithm. We study in detail the error analysis of the qd and Compqd algorithms and we introduce new condition numbers so that the relative forward rounding error bounds can be derived directly. Our numerical experiments illustrate that the Compqd algorithm is much more accurate than the qd algorithm, relegating the influence of the condition numbers up to second order in the rounding unit of the computer. Three applications of the new algorithm in the obtention of continued fractions and in pole and zero detection are shown.

math.NA↗
arXiv2016

A note on alternating minimization algorithms: Bregman frame

In this paper, we propose a Bregman frame for several classical alternating minimization algorithms. In the frame, these algorithms have uniform mathematical formulation. We also present convergence analysis for the frame algorithm. Under the Kurdyka-Lojasiewicz property, stronger convergence is obtained.

math.NA↗
arXiv2015

Variants of alternating minimization method with sublinear rates of convergence for convex optimization

The alternating minimization (AM) method is a fundamental method for minimizing convex functions whose variable consists of two blocks. How to efficiently solve each subproblems when applying the AM method is the most concerned task. In this paper, we investigate this task and design two new variants of the AM method by borrowing proximal linearized techniques. The first variant is very suitable for the case where half of the subproblems are hard to be solved and the other half can be directly computed. The second variant is designed for parallel computation. Both of them are featured by simplicity at each iteration step. Theoretically, with the help of the proximal operator we first write the new as well as the existing AM variants into uniform expressions, and then prove that they enjoy sublinear rates of convergence under very minimal assumptions.

math.OC↗
arXiv2014

Projected shrinkage algorithm for box-constrained L1-minimization

Box-constrained L1-minimization can perform remarkably better than classical L1-minimization when correction box constraints are available. And also many practical L1-minimization models indeed involve box constraints because they take certain values from some interval. In this paper, we propose an efficient iteration scheme, namely projected shrinkage (ProShrink) algorithm, to solve a class of box-constrained L1-minimization problems. A key contribution in our technique is that a complicated proximal point operator appeared in the deduction can be equivalently simplified into a projected shrinkage operator. Theoretically, we prove that ProShrink enjoys a convergence of both the primal and dual point sequences. On the numerical level, we demonstrate the benefit of adding box constraints via sparse recovery experiments.

math.OC↗
arXiv2014

Proximal linearized iteratively reweighted least squares for a class of nonconvex and nonsmooth problems

For solving a wide class of nonconvex and nonsmooth problems, we propose a proximal linearized iteratively reweighted least squares (PL-IRLS) algorithm. We first approximate the original problem by smoothing methods, and second write the approximated problem into an auxiliary problem by introducing new variables. PL-IRLS is then built on solving the auxiliary problem by utilizing the proximal linearization technique and the iteratively reweighted least squares (IRLS) method, and has remarkable computation advantages. We show that PL-IRLS can be extended to solve more general nonconvex and nonsmooth problems via adjusting generalized parameters, and also to solve nonconvex and nonsmooth problems with two or more blocks of variables. Theoretically, with the help of the Kurdyka- Lojasiewicz property, we prove that each bounded sequence generated by PL-IRLS globally converges to a critical point of the approximated problem. To the best of our knowledge, this is the first global convergence result of applying IRLS idea to solve nonconvex and nonsmooth problems. At last, we apply PL-IRLS to solve three representative nonconvex and nonsmooth problems in sparse signal recovery and low-rank matrix recovery and obtain new globally convergent algorithms.

math.OC↗
arXiv2013

A dual algorithm for a class of augmented convex models

Convex optimization models find interesting applications, especially in signal/image processing and compressive sensing. We study some augmented convex models, which are perturbed by strongly convex functions, and propose a dual gradient algorithm. The proposed algorithm includes the linearized Bregman algorithm and the singular value thresholding algorithm as special cases. Based on fundamental properties of proximal operators, we present a concise approach to establish the convergence of both primal and dual sequences, improving the results in the existing literature.

math.OC↗
arXiv2013

New bounds for circulant Johnson-Lindenstrauss embeddings

This paper analyzes circulant Johnson-Lindenstrauss (JL) embeddings which, as an important class of structured random JL embeddings, are formed by randomizing the column signs of a circulant matrix generated by a random vector. With the help of recent decoupling techniques and matrix-valued Bernstein inequalities, we obtain a new bound $k=O(ε^{-2}\log^{(1+δ)} (n))$ for Gaussian circulant JL embeddings. Moreover, by using the Laplace transform technique (also called Bernstein's trick), we extend the result to subgaussian case. The bounds in this paper offer a small improvement over the current best bounds for Gaussian circulant JL embeddings for certain parameter regimes and are derived using more direct methods.

cs.IT↗