Neural architecture search has revolutionized automated machine learning and demonstrated transformative potential across natural language processing domains. This survey presents a comprehensive task-oriented analysis of NAS applications in NLP, systematically categorizing research into three core areas of text representation and classification, sequence modeling and generation, and information extraction. We trace the field’s evolution through three distinct developmental phases from foundational methodology transfer and task-specific architectural customization to deployment-oriented optimization. The analysis examines key methodological advances in search space construction, search strategy optimization, and efficient evaluation mechanisms, while highlighting architectural innovations including pre-trained model compression, multimodal fusion, and specialized designs for machine translation, speech recognition, and entity extraction. Building on this comprehensive synthesis of current progress, we identify critical future research directions in large language model optimization, zero-cost evaluation methods, and knowledge-aware architecture design for practical NLP systems.

This manuscript proposes an Improved Sparrow Search Algorithm with Multi-Strategies (MSISSA) to address the shortcomings of the Sparrow Search Algorithm (SSA), and applies it to optimize the deployment of 3D Wireless Sensor Nodes (3D-WSNs). The improvements of MSISSA encompass three main components: firstly, it utilizes an adaptive weight operator for updating position to enhance local search accuracy and expedite convergence; secondly, it introduces the spiral flight mechanism of the Moth Flame Optimization (MFO) algorithm, which balances the algorithm’s local and global search abilities; finally, it uses the Levy flight mechanism to jump out the local optima. The experimental results demonstrate that MSISSA outperforms SSA, its four variants, and two classical intelligent algorithms. The simulation results indicate that when the number of nodes is 30 and 50, the optimal coverage rate of MSISSA reaches 91.89% and 99.11%, respectively, surpassing those achieved by SSA and its four variants, as well as two classical intelligent algorithms. Therefore, it can be concluded that MSISSA is well-suited for optimizing the deployment of 3D-WSNs.

This manuscript presents ABSAS-CS-GSA, an improved iteration of the beetle antennae search (BAS) algorithm, tailored to address the shortcoming of the BAS with low convergence accuracy and easily falling into local optima. The improvements are structured into three principal components: first, dynamic step-size adjustment predicated on initial population to bolster solution accuracy and expedite convergence; second, position update mechanism integrating golden sine algorithm to diversify search patterns and expedite convergence; and third, population disturbance based on vertical and horizontal cross strategy to avoid falling into local optima. Experimental results on 12 benchmark functions demonstrate ABSAS-CS-GSA's superiority over the standard BAS and its four variants, as well as over two classic swarm intelligence algorithms. The simulation results of the algorithm applied to coverage optimization of wireless sensor networks with 30 nodes and 50 nodes reveal a marked improvement in both optimal and average coverage metrics relative to seven alternative algorithms. The refined algorithm exhibits excellent performance and is well-suited for tackling the coverage issue within wireless sensor networks.

This manuscript introduces an improved Cuckoo Search (CS) algorithm, known as BASCS, designed to address the inherent limitations of CS, including insufficient search space coverage, premature convergence, low search accuracy, and slow search speed. The proposed improvements encompass four main areas: the integration of tent chaotic mapping and random migration in population initialization to reduce the impact of random errors, the guidance of Levy flight by the directional determination strategy of the Beetle Antennae Search (BAS) algorithm during the global search phase to improve search accuracy and convergence speed, the adoption of the Sine Cosine Algorithm for local exploitation in later iterations to enhance local optimization and accuracy, and the adaptive adjustment of the step-size factor and elimination probability throughout the iterative process to convergence. The performance of BASCS is validated through ablation experiments on 10 benchmark functions, comparative experiments with the original CS and its four variants, and application to a robot path planning problem. The results demonstrate that BASCS achieves higher convergence accuracy and exhibits faster convergence speed and superior practical applicability compared to other algorithms.

Data-driven models perform poorly on part-of-speech tagging problems with the square Hmong language, a low-resource corpus. This paper designs a weight evaluation function to reduce the influence of unknown words. It proposes an improved harmony search algorithm utilizing the roulette and local evaluation strategies for handling the square Hmong part-of-speech tagging problem. The experiment shows that the average accuracy of the proposed model is 6%, 8% more than HMM and BiLSTM-CRF models, respectively. Meanwhile, the average F1 of the proposed model is also 6%, 3% more than HMM and BiLSTM-CRF models, respectively.

For wireless sensor network (WSN) coverage problems, since the sensing range of sensor nodes is a circular area with symmetry, taking symmetry into account when deploying nodes will help simplify problem solving. In addition, in view of two specific problems of high node deployment costs and insufficient effective coverage in WSNs, this paper proposes a WSN coverage optimization method based on the improved grey wolf optimizer with multi-strategies (IGWO-MS). As far as IGWO-MS is concerned, first of all, it uses Sobol sequences to initialize the population so that the initial values of the population are evenly distributed in the search space, ensuring high ergodicity and diversity. Secondly, it introduces a search space strategy to increase the search range of the population, avoid premature convergence, and improve search accuracy. And then, it combines reverse learning and mirror mapping to expand the population richness. Finally, it adds Levy flight to increase the disturbance and improve the probability of the algorithm jumping out of the local optimum. To verify the performance of IGWO-MS in WSN coverage optimization, this paper rasterizes the coverage area of the WSN into multiple grids of the same size and symmetry with each other, thereby transforming the node coverage problem into a single-objective optimization problem. In the simulation experiment, not only was IGWO-MS selected, but four other algorithms were also selected for comparison, namely particle swarm optimization (PSO), grey wolf optimizer (GWO), grey wolf optimization based on drunk walk (DGWO), and grey wolf optimization led by two-headed wolves (GWO-THW). The experimental results demonstrate that when the number of nodes for WSN coverage optimization is 20 and 30, the optimal coverage rate and average coverage rate using IGWO-MS are both improved compared to the other four comparison algorithms. To make this clear, in the case of 20 nodes, the optimal coverage rate of IGWO-MS is increased by 13.19%, 1.68%, 4.92%, and 3.62%, respectively, compared with PSO, GWO, DGWO, and GWO-THW; while IGWO-MS performs even better in terms of average coverage rate, which is 16.45%, 3.13%, 11.25%, and 6.19% higher than that of PSO, GWO, DGWO, and GWO-THW, respectively. Similarly, in the case of 30 nodes, compared with PSO, GWO, DGWO, and GWO-THW, the optimal coverage rate of the IGWO-MS is increased by 15.23%, 1.36%, 5.55%, and 3.66%; the average coverage rate is increased by 16.78%, 1.56%, 10.91%, and 8.55%. Therefore, it can be concluded that IGWO-MS has certain advantages in solving WSN coverage problems, which is reflected in that not only can it effectively improve the coverage quality of network nodes, but it also has good stability.

An improved cuckoo search (CS) algorithm focusing on optimizing the updating of the bird’s nest position is proposed to overcome the shortcomings of CS algorithm, such as low search accuracy, easy premature convergence, and weak local search ability in the late stage. The proposed algorithm employs three main strategies. First, in the early and middle stage, aiming to expand the search domain, increase the population diversity, and prevent the premature convergence of the algorithm, the global optimization strategy combining opposition-based learning and Levy flight is used to update the nest position. Second, to avoid the algorithm from falling into a local optimum, a dynamic inertia weight is employed to reduce the effect of the current nest position when the nest position is updated during the random migration. Third, in the late stage, to improve the local search ability and search accuracy, the nest position is updated using the local exploitation strategy of the Aquila optimization (AO) algorithm to replace the Levy flight mechanism. The results of comparative experiments with CS algorithm and its four variants on function optimization show that the proposed algorithm has a better global search ability, better local development ability, faster convergence speed and better search accuracy than other selected algorithms.

The current machine learning algorithms classify human activities with inaccurate accuracy, poor generalization ability of the model, and poor classification effect. Proposing to use Random Forest classifier to classify the samples. The classifier has the advantage of good generalization ability, high accuracy, and the ability to handle a large number of sample data. After comparing with BP neural networks, Naive Bayesian networks, and decision trees, the random forest classifier achieved an accuracy of 98% on the test set, much higher than the 55%, 87%, and 88% of the other algorithms. In addition, we also tested the generalization ability of the model using the K-Folder cross-validation method, which yielded an average accuracy of 95.5%, also much higher than the 49.9%, 85%, and 90% of the other classifiers. The experimental results show that the random forest classifier has significantly improved in accuracy and generalization ability compared with BP neural network by 78% and 91%, respectively. Therefore, the superiority of the random forest classifier in human activity classification is proved.

针对鲸鱼优化算法存在的求解精度不高、收敛速度较慢和易陷入局部最优等缺点,设计了 一种基于莱维飞行和布朗运动的鲸鱼优化算法.先利用莱维飞行方法对鲸鱼种群进行初始化,以增加初始种群的多样性;再根据布朗运动原理对鲸鱼种群的位置更新进行随机扰动,以避免算法提前陷入局部最优.将改进的鲸鱼优化算法与鲸鱼优化算法、粒子群优化算法、遗传算法和蚁群优化算法在7个不同的基准测试函数上进行对比测试,结果表明,改进的鲸鱼优化算法在求解精度、收敛速度方面均优于其他4种算法.对初始化阶段采用莱维飞行策略的改进鲸鱼优化算法与采用随机搜索策略的鲸鱼优化算法的初始解探索范围进行仿真对比实验,结果表明,改进鲸鱼优化算法一定程度上可以避免陷入局部最优.

This paper discusses a hybrid grey wolf optimizer utilizing a clone selection algorithm (pGWO-CSA) to overcome the disadvantages of a standard grey wolf optimizer (GWO), such as slow convergence speed, low accuracy in the single-peak function, and easily falling into local optimum in the multi-peak function and complex problems. The modifications of the proposed pGWO-CSA could be classified into the following three aspects. Firstly, a nonlinear function is used instead of a linear function for adjusting the iterative attenuation of the convergence factor to balance exploitation and exploration automatically. Then, an optimal α wolf is designed which will not be affected by the wolves β and δ with poor fitness in the position updating strategy; the second-best β wolf is designed, which will be affected by the low fitness value of the δ wolf. Finally, the cloning and super-mutation of the clonal selection algorithm (CSA) are introduced into GWO to enhance the ability to jump out of the local optimum. In the experimental part, 15 benchmark functions are selected to perform the function optimization tasks to reveal the performance of pGWO-CSA further. Due to the statistical analysis of the obtained experimental data, the pGWO-CSA is superior to these classical swarm intelligence algorithms, GWO, and related variants. Furthermore, in order to verify the applicability of the algorithm, it was applied to the robot path-planning problem and obtained excellent results.

针对基本和声搜索(Harmony search,HS)算法收敛速度较慢、易陷入局部最优和计算精度不高的缺点,结合正余弦优化算子、Levy飞行机制和参数动态调整策略,提出一种改进的和声搜索算法.该算法在即兴创作阶段,首先引入正余弦优化算子和微调带宽相结合的方式对和声向量进行微调操作,充分利用最优个体和当前个体的位置信息,提高算法的计算精度和收敛速度;再采用Levy飞行机制对微调带宽进行更新,避免算法陷入局部最优,提高全局搜索能力;在算法迭代过程中,对和声记忆库存储概率、基音微调概率和搜索域进行自适应动态调整,以进一步提高算法收敛性能.在10个基准函数上进行性能对比试验的结果表明,本文提出的算法具有较强的全局搜索能力,较快的收敛速度和较高的计算精度.

A self-adaptive whale optimization algorithm integrating four improvement strategies is proposed to overcome the shortcomings of the basic whale optimization algorithm that are easy to fall into local optima, slow convergence speed and low calculation accuracy. Firstly, the chaotic logistics mapping method is used to initialize the population randomly to increase the diversity of the population and the uniformity of individual distribution. Secondly, the convergence factor is non-linearized and dynamically adjusted to balance the global search and local search performance of the algorithm. Thirdly, the dynamic adjustment of the search step length is realized by setting different adaptive inertia weights according to stages. Finally, crossover and mutation operations are performed on the random dimensions of the population individuals to avoid the algorithm from falling into the local optimum. The results of comparative experiments with three related algorithms on six benchmark functions show that the proposed algorithm is superior to the others in terms of global search capability, convergence speed and calculation accuracy.

为了提高词性标注模型训练语料的质量,设计了一种利用FP Growth算法从训练语料库中自动获取词性标注规则的方法,并将该方法与基于Apriori算法的词性标注规则获取方法进行了对比实验.实验结果显示,对于0.1万、0.2万和l万词级的小规模语料库,2种方法获取的词性标注规则条数均相同,但基于FP-Growth算法的时间耗费分别仅为基于Apriori算法的0.013866％,0.010399％,0.003132％;对于10万、100万词级的训练语料库,基于Apriori算法无法获取任何规则,但基于FP-Growth算法依然可以在合理时间内获取有效的规则.这说明,基于FP-Growth算法的词性标注规则获取方法是可行且高效的,满足在优化训练语料库时能从不同规模的语料库中自动获取词性标注规则的实际需求.

An adaptive harmony search algorithm utilizing differential evolution and opposition-based learning (AHS-DE-OBL) is proposed to overcome the drawbacks of the harmony search (HS) algorithm, such as its low fine-tuning ability, slow convergence speed, and easily falling into a local optimum. In AHS-DE-OBL, three main innovative strategies are adopted. First, inspired by the differential evolution algorithm, the differential harmonies in the population are used to randomly perturb individuals to improve the fine-tuning ability. Then, the search domain is adaptively adjusted to accelerate the algorithm convergence. Finally, an opposition-based learning strategy is introduced to prevent the algorithm from falling into a local optimum. The experimental results show that the proposed algorithm has a better global search ability and faster convergence speed than other selected improved harmony search algorithms and selected metaheuristic approaches.

针对传统和声算法收敛速度慢和搜索精度低等固有缺点,提出一种改进的自适应全局最优和声搜索算法.在即兴创作方案中,带宽由当前和声里的最优和声变量和最差和声变量之差表示,使得带宽具有针对具体情况的自适应能力,并且每次保存最优和声中一个随机和声变量.在产生的随机数大于和声记忆库存储考虑概率时,利用种群内差分随机生成一个和声变量.为了提高和声搜索算法的搜索能力,在即兴创作结束后产生一个新的和声的同时,再从当前种群中的最小和声到最大和声之间随机产生一个和声,然后将两个新产生和声中误差小的和声进入更新和声记忆库阶段.将所提出的算法与3个改进和声搜索算法在13个测试函数上进行对比.试验结果表明,提出的改进算法具有更好的全局搜索能力和收敛速度.

Recently, it is difficult to simulate, analyze and control a real knowledge-based system using the correspondence Petri net (PN) when there exist many current states. To overcome the state explosion problem of PN, an efficient decomposition algorithm is presented to divide a large-scale PN into a series of corresponding sub-PNs by keeping the consistency of dynamic properties. In this novel decomposition approach, an index function is defined to judge the subnet needs to be decomposed or not. Furthermore, an exhaustive analysis on the consistency of related dynamic properties is also discussed between the original PN and the corresponding sub-PNs. Finally, a case study is carried out to illustrate the feasibility and validity of the proposed approach.

为了进一步提高中文语料库中语料的词性标注效率,在分析最大熵模型（MEM）和隐马尔科夫模型（HMM）所涉及理论、算法及其在中文词性标注技术中的应用的基础上,进行了基于MEM和HMM的中文词性标注实验.实验结果显示,基于MEM和HMM的中文词性标注算法都获得了一致性很好且覆盖率较高的标注效果,中文词性标注的准确率、召回率和F1这3个指标均达到92%以上;MEM的标注效果总体上比HMM的稍佳.

To increase the convergence speed for tackling the complex-valued system of linear equation (CVSLE) problem online, a novel fully complex-valued gradient neural network (FCGNN) model is proposed. Firstly, to tackle the CVSLE problem in the complex number field instead of the real number field, the negative gradient descent information and a new nonlinear activation function are used to design a FCGNN model. Secondly, the proof for the global convergence property of the FCGNN model is deduced from the strict derivation. Finally, the simulation results show that this FCGNN model has a higher convergence rate than the original gradient neural network (GNN) model and the improved Zhang neural network (IZNN) model.

In order to reduce the storage capacity of fonts and improve the display quality of characters, this study proposes an approach to generate square Hmong characters using intelligent derivation mechanism. In the proposed approach, the process of generating square Hmong characters is transformed into the process of solving the bounding box of components using binary tree. Thus, various square Hmong characters can be generated by repeatedly performing topological transformation of limited components. A corresponding intelligent derivation model is developed to get square Hmong characters based on the glyph data of objective components obtained by intelligent derivation. Meanwhile, deterministic finite automata is also used to represent the square of square Hmong characters. Testing results of the editing software developed based on this model demonstrate that the square Hmong characters generated by the proposed intelligent derivation mechanism can meet the needs of practical applications.