Search

Large language models (LLMs) have demonstrated strong multilingual capabilities; yet, they are mostly English-centric due to the imbalanced training corpora. Existing works leverage this phenomenon to improve their multilingual performances on NLP tasks. In this work, we extend the evaluation from NLP tasks to real user queries. We find that even though translation into English can help improve the performance of multilingual NLP tasks for English-centric LLMs, it may not be optimal for all scenarios. For culture-related tasks that need deep language understanding, prompting in the native language proves to be more promising since it can capture the nuances related to culture and language. Therefore, we advocate for more efforts towards the development of strong multilingual LLMs instead of just English-centric LLMs., Comment: 18 pages

Hierarchical topic modeling aims to discover latent topics from a corpus and organize them into a hierarchy to understand documents with desirable semantic granularity. However, existing work struggles with producing topic hierarchies of low affinity, rationality, and diversity, which hampers document understanding. To overcome these challenges, we in this paper propose Transport Plan and Context-aware Hierarchical Topic Model (TraCo). Instead of early simple topic dependencies, we propose a transport plan dependency method. It constrains dependencies to ensure their sparsity and balance, and also regularizes topic hierarchy building with them. This improves affinity and diversity of hierarchies. We further propose a context-aware disentangled decoder. Rather than previously entangled decoding, it distributes different semantic granularity to topics at different levels by disentangled decoding. This facilitates the rationality of hierarchies. Experiments on benchmark datasets demonstrate that our method surpasses state-of-the-art baselines, effectively improving the affinity, rationality, and diversity of hierarchical topic modeling with better performance on downstream tasks., Comment: Accepted to AAAI2024 conference. Our code is available at https://github.com/bobxwu/TraCo

Existing solutions to zero-shot text classification either conduct prompting with pre-trained language models, which is sensitive to the choices of templates, or rely on large-scale annotated data of relevant tasks for meta-tuning. In this work, we propose a new paradigm based on self-supervised learning to solve zero-shot text classification tasks by tuning the language models with unlabeled data, called self-supervised tuning. By exploring the inherent structure of free texts, we propose a new learning objective called first sentence prediction to bridge the gap between unlabeled data and text classification tasks. After tuning the model to learn to predict the first sentence in a paragraph based on the rest, the model is able to conduct zero-shot inference on unseen tasks such as topic classification and sentiment analysis. Experimental results show that our model outperforms the state-of-the-art baselines on 7 out of 10 tasks. Moreover, the analysis reveals that our model is less sensitive to the prompt design. Our code and pre-trained models are publicly available at https://github.com/DAMO-NLP-SG/SSTuning ., Comment: Accepted to the Findings of ACL 2023

Cross-lingual topic models have been prevalent for cross-lingual text analysis by revealing aligned latent topics. However, most existing methods suffer from producing repetitive topics that hinder further analysis and performance decline caused by low-coverage dictionaries. In this paper, we propose the Cross-lingual Topic Modeling with Mutual Information (InfoCTM). Instead of the direct alignment in previous work, we propose a topic alignment with mutual information method. This works as a regularization to properly align topics and prevent degenerate topic representations of words, which mitigates the repetitive topic issue. To address the low-coverage dictionary issue, we further propose a cross-lingual vocabulary linking method that finds more linked cross-lingual words for topic alignment beyond the translations of a given dictionary. Extensive experiments on English, Chinese, and Japanese datasets demonstrate that our method outperforms state-of-the-art baselines, producing more coherent, diverse, and well-aligned topics and showing better transferability for cross-lingual classification tasks., Comment: Accepted to AAAI2023 conference. Code is available at https://github.com/BobXWu/InfoCTM

Despite the remarkable achievements of large language models (LLMs) in various tasks, there remains a linguistic bias that favors high-resource languages, such as English, often at the expense of low-resource and regional languages. To address this imbalance, we introduce SeaLLMs, an innovative series of language models that specifically focuses on Southeast Asian (SEA) languages. SeaLLMs are built upon the Llama-2 model and further advanced through continued pre-training with an extended vocabulary, specialized instruction and alignment tuning to better capture the intricacies of regional languages. This allows them to respect and reflect local cultural norms, customs, stylistic preferences, and legal considerations. Our comprehensive evaluation demonstrates that SeaLLM-13b models exhibit superior performance across a wide spectrum of linguistic tasks and assistant-style instruction-following capabilities relative to comparable open-source models. Moreover, they outperform ChatGPT-3.5 in non-Latin languages, such as Thai, Khmer, Lao, and Burmese, by large margins while remaining lightweight and cost-effective to operate., Comment: Technical report

Computational Fluid Dynamics: A Practical Approach, Third Edition, is an introduction to CFD fundamentals and commercial CFD software to solve engineering problems. The book is designed for a wide variety of engineering students new to CFD, and for practicing engineers learning CFD for the first time. Combining an appropriate level of mathematical background, worked examples, computer screen shots, and step-by-step processes, this book walks the reader through modeling and computing, as well as interpreting CFD results. This new edition has been updated throughout, with new content and improved figures, examples and problems.Includes a new chapter on practical guidelines for mesh generation Provides full coverage of high-pressure fluid dynamics and the meshless approach to provide a broader overview of the application areas where CFD can be used Includes online resources with a new bonus chapter featuring detailed case studies and the latest developments in CFD

Turbulence is one of the most important but unresolved problems in modern fluid dynamics. Mathematically, one counterexample can overthrow a theory and we really do not need a second example. However, classical and current turbulence theories are filled with not only one but many self-contradictions. Therefore, we really do not have the right turbulence theory. The transition community believes that turbulence is generated by “vortex breakdown,” but the turbulence community believes there are coherent structures in fully developed turbulence after vortex breakdown. It is equivalent to say we study the structure of a house after the house collapses and breaks down. This is really ridiculous. There are many others. According to Liu, (1) flow transition is not a process of vortex breakdown but turbulence vortex structure buildup; (2) the nature of turbulence generation is that fluids cannot tolerate high shear and shear must transfer to rotation and form a very fast rotation core; (3) “shear layer instability” is the “mother of turbulence”; (4) turbulence small scales are generated by multiple-level shear layers that are generated by multiple level sweeps, ejections, negative and positive spikes; (5) large vortex provides energy to smaller vortices through fast rotation, which causes strong sweeps. According to Liu, the nature of the flow transition is mainly caused by vorticity rollup from the wall. Flow transition is vorticity redistribution and increment from near wall to whole boundary layer. In addition, flow transition is a process of non-rotational vorticity transferring to rotational vorticity.

Dynamic tension variation caused by conductor galloping is a major impact on design and stable operation of overhead transmission lines. A formula to calculate tension variation caused by the conductor galloping was presented in this paper by using the energy balance method. And two important parameters and were proposed by the dimensionless analysis to study the factors influencing the dynamic tension. By comparison with the popular expression obtained by the length variation method, the formula deduced in this paper has more extensive applications and it was simplified to obtain the expression derived by the length variation method as the parameters and have the smaller values. Using the actual galloping conditions, the tension variation maximum against the different parameters was obtained and it can go up to 3-4 times as large as the initial tension for some extreme cases. Additional a finite element galloping model was used to verify the theoretical expression. The results by the theoretical calculation were agreement well with the numerical simulated values except the extremely large amplitude galloping conditions.

This final technical report serves as a summary of our recent work on LES and experiments on supersonic MVG. An implicitly implemented large eddy simulation (ILES) by using the fifth order WENO scheme is applied to study the flow around the microramp vortex generator (MVG) at Mach 2.5. A number of new discoveries on the flow around supersonic MVG has been made including spiral points, surface separation topology, source of the momentum deficit, inflection surface, Kelvin-Helmholtz instability, vortex ring generation, ring-shock interaction, 3-D recompression shock structure, influence of MVG decline angles. Most of the new discoveries, which were made in 2009, were confirmed by experiments conducted by the UTA experimental team in 2010. A new 5-pair-vortex-tube model near the MVG is given based on the ILES observation. The vortex ring-shock interaction is found as the new mechanism of the reduction of the separation zone induced by the shock-boundary layer interaction., The original document contains color images.

The critical problem of CFD is perhaps an accurate approximation of derivatives for a given discrete data set. Based on our previous work on the weighted compact scheme (WCS), a modified weighted compact scheme (MWCS) has been developed. Similar to WENO, three high order candidates, left, right, and central, are constructed by Hermite polynomials., The original document contains color images.

During the past two years (12/00-3/03), we have developed a very efficient and reliable direct numerical simulation (DNS) and large eddy simulation (LES) software (DNSUTA) for compressible flow in a curvilinear coordinates which is a continuation of our previous effort supported by AFOSR. The software includes a high-quality grid generation code which can generate smooth and near-orthogonal, body-fitted grids and a parallel efficient flow solver which applies sixth-order compact scheme, eighth-order filter, LU-SGS flow solver, non-reflecting boundary condition, high-order weighted compact scheme for shock capturing, structure function sub-grid model, and MPI parallel computation. The code has been successfully used for prediction of flow transition around flat plate and airfoils with various attack angles. The code has also been used for flow separation control for low Reynolds airfoils with steady and pulsed blowing jets. All of these simulations are real time dependent, 3-D, with general geometry and have demonstrated qualitative or quantitative agreement with experiment or theoretical work. We believe the topics described above are very challenging and our accomplishment is very significant for both basic science and Air Force mission. The high-order weighted compact scheme and non-reflecting boundary conditions in a curvilinear coordinate are the original work conducted by us which is a significant contribution to the CFD community.

This is the third time for worldwide DNS/LES experts to get together to introduce research progress, exchange ideas, and discuss future research directions since the US Air Force Office of Scientific Research (AFOSR) sponsored the First AFOSR. International Conference on DNSILES (FAICDL) at Louisiana Tech University in 1997. The second conference was held at Rutgers - The State University of New Jersey in 1999. I am very pleased that many scientists and engineers are interested in this conference. I have received over 90 regular papers, in addition to 19 invited papers. It clearly shows that the DNS/LES community is much larger than ever before. Significant progress has been achieved during the past decade. It is difficult to give a complete review in this area. People may take a reference from reviews by Moin & Mehesh (1998) for DNS and Leseiur & Metais (1996) and Piomelli (1999) for LES. Sagaut (2000) just published a LES book, which gives discussions on many issues related to LES. It is also very difficult to invite all of the important contributors to give invited lectures for the conference. During this conference, Germano will give a review on LES; Visbal and Gatski will talk about Air Force and NASA needs on DNSILES and their applications; Sagaut, Kato, and Shan will talk about complicated configurations and industrial application; Jolsin will talk about DNS for flow control; Adams and Thong about high speed flow transitions; Givi about LES for reacting flow, Shu and Jiang about high order discretization, Karniadakis about DNS for unstructured grids; Vasilyev about some important LES constrains. In addition, Knight, Piomelli, Satofuka, Verstappen will also give lectures about their own distinguished work., ISBN 1-57074-488-2. For individual articles, see ADP013620 thru ADP013707

In this work, the high-order central compact scheme is introduced to improve the accuracy of the IFLOW solver. The high-order compact filter is used to reduce non-physical oscillations. To achieve the conservation property of the scheme, the compact scheme is combined with the ENO reconstruction method. The numerical flux at the cell interface is approximated by the derivative of its primitive function which can be exactly calculated. At the boundary, ghost points are constructed. The high-order extrapolation is used to obtain values at the ghost points. Instead of using the one-sided compact scheme, the explicit central difference scheme is applied at the boundary. The results from spatial simulations of trailing vortices have shown obvious improvement in accuracy., Presented at AFOSR International Conference on DNS/LES (3rd) held at Univ. of Texas, Arlington, TX on 5-9 Aug 2001. Prepared in collaboration with Dept. of Mathematics, Univ. of Texas, Arlington, TX. This article is from ADA412801 DNS/LES Progress and Challenges. Proceedings of the Third AFOSR International Conference on DNS/LES

In this paper a new class of finite difference schemes - the Weighted Compact Schemes are proposed. According to the idea of the WENO schemes, the Weighted Compact Scheme is constructed by a combination of the approximations of derivatives on candidate stencils with properly assigned weights so that the non-oscillatory property is achieved when discontinuities appear. The primitive function reconstruction method of ENO schemes is applied to obtain the conservative form of the Weighted Compact Scheme. This new scheme not only preserves the characteristic of standard compact schemes and achieves high order accuracy and high resolution using a compact stencil, but also can accurately capture shock waves and discontinuities without oscillation. Numerical examples show the new scheme is very promising and successfill., Presented at AFOSR International Conference on DNS/LES (3rd) held at Univ. of Texas, Arlington, TX on 5-9 Aug 2001. This article is from ADA412801 DNS/LES Progress and Challenges. Proceedings of the Third AFOSR International Conference on DNS/LES

In this paper, the Laplace equation and the algebraic transformation were presented for the domains in 2D and 3D physical space. The second order finite difference scheme was used for the discretization of the grid generating equations. The linear system is solved by the ADI method and the convergence was accelerated by the multigrid FAS scheme. The performance characteristic of the algorithm was discussed and illustrations were made., Presented at AFOSR International Conference on DNS/LES (3rd) held at Univ. of Texas, Arlington, TX on 5-9 Aug 2001. This article is from ADA412801 DNS/LES Progress and Challenges. Proceedings of the Third AFOSR International Conference on DNS/LES

A new so-called truncation error reduction method (TERM) is developed in this work. This is an iterative process which uses a coarse grid (2h) to estimate the truncation error and then reduces the error on the original grid (h). The purpose is to use multigrid and simple stencils for high-order accuracy., Presented at AFOSR International Conference on DNS/LES (3rd) held at Univ. of Texas, Arlington, TX on 5-9 Aug 2001. This article is from ADA412801 DNS/LES Progress and Challenges. Proceedings of the Third AFOSR International Conference on DNS/LES

A traditional local approach is applied for simulation of a statistically steady flow near a flat plate. A local flow domain is placed far from the leading edge of the plate. A coupled system of the Prandtl and the Navier Stokes equations for disturbances is chosen as the governing system of equations. Fully spectral approach is used for space discretization. This helps to avoid prescribing inflow and outflow boundary conditions. Acceptable particular solutions are selected to describe physically adequate phenomenon. A new spectral method for integrating the Cauchy problem is adopted to study the flow evolution., Presented at AFOSR International Conference on DNS/LES (3rd) held at Univ. of Texas, Arlington, TX. Prepared in cooperation with Dept. of Mathematics, Univ. of Texas, Arlington, TX. This article is from ADA412801 DNS/LES Progress and Challenges. Proceedings of the Third AFOSR International Conference on DNS/LES

This paper carries out numerical study of the flow separations around NACA 0012 airfoil at large angles of attack. Flow separation introduces two major effects: sudden loss of lift and generation of aerodynamic noise. These two factors are highly concerned on the aircraft designation. This study gives a detail picture of flow separation. As the problems caused by flow separation are complicated, the spatial and temporal complexity makes it difficult to access by conventional experiment methods. In the presented work, the numerical investigation results from solving the time-dependent Navier Stokes equations in the generalized curvilinear coordinates. Using a fourth order centered compact scheme for spatial discretization facilitates high resolution of the flow field, which will be neglected if using low-order numerical schemes. To avoid possible non-physical wave reflection, the non-reflecting boundary conditions are used at far-field and outlet boundary. Complex flow separation, vortex shedding, vortex merging, and vortex paring are observed in the computational results. The main purpose of this paper is to provide more detailed information of the flow separation., Presented at AFOSR International Conference on DNS/LES (3rd) held at Univ. of Texas, Arlington, TX on 5-9 Aug 2001. This article is from ADA412801 DNS/LES Progress and Challenges. Proceedings of the Third AFOSR International Conference on DNS/LES

The Defense University Research Instrument Program (DURIP) grant, Upgrade of the Computing and Visualization System for DNS/LES Research, Grant No. F49620-00-1-0220 for $400,000 has been mainly spent for upgrading the existing SGI Origin 2000 from 16 CPU to 32 CPU with additional 6 GB memory and over 200 GB hard disk. A 8 cpu SGI 1200L clustered system and a number of SGI and Dell Workstations and PCs have also been purchased by this fund. These purchases and upgrading have greatly strengthen Louisiana Tech University in conducting Defense related research and education. Dr. Liu's DNS/LES group has got permission to access all theses systems to conduct their state-of-the art DNS/LES research which has been supported by AFOSR for a number of years. They have used that system to conduct DNS/LES computation for 3-D airfoils and plan to work for simplified aircrafts.

An oblique transition process of a flat-plate boundary layer at a free-stream Mach number of M (infinity) =4.5 and a Reynolds number of 10000 based on free-stream velocity and inflow displacement thickness is simulated as the validation of the code. Four transitional stages are observed: the linear and weakly nonlinear growth, the appearance of staggered A-vortex patterns, the evolution of A-vortex into hairpin vortex, the breakdown of hairpin vortical structures. The evolution of averaged quantities, such as the skin-friction coefficient, the mean velocity profile, the boundary layer thickness are all computed and found agreed very well with theoretic and experimental results. For complex geometry, 2D and 3D numerical grid generation method are used to generate high-quality meshes. Subsonic flow around 2D NACA 0012 airfoil and 3D Delta wing with a large angle of attack are investigated with large eddy simulation. Flow separations and large eddy structures are observed, the interactions between vortical structures will be studied.

This conference, sponsored by the US Air Force Office of Scientific Research (AFOSR), was held in Louisiana Tech University, Ruston, Louisiana, USA on August 4-8. 1997. The conference attracted 93 participants from 16 countries all over the world including most leading scientists in the area of DNS and LES. Direct Numerical Simulation (DNS) refers to accurate numerical solution of three dimensional, time-dependent Navier-Stokes equations for both flow transition arid turbulence. DNS is the only numerical approach dealing with the exact solution of 3-D time-dependent N-S equations without any ad hoc assumption and in which all time and space length scales are taken into account.

The three dimensional development of flow transition in both subsonic and supersonic Joukowsky airfoil boundary layers are studied by direct numerical simulation (DNS). The numerical simulation is performed by a spatial approach. A full compressible Navier Stokes system in curvilinear coordinates is employed so that we can simulate the transition around general geometric configurations. The numerical results agree very well with the linear stability theory (LST) at the linear growth stage for both primary and second modes in the flat plate boundary layers. The whole process of controlled flow transition induced by blowing/suction around airfoils is simulated by directly solving the N-S system with Reynolds number around one million. Some differences are found in comparison to the incompressible counterpart, and some new phenomena for the transition around airfoils are observed which at least qualitatively agree with physics.

The two- and three-dimensional development of leading-edge receptivity and transition in a 2-D Joukowsky airfoil boundary layer are investigated by direct numerical simulation (DNS) using the incompressible Navier-Stokes equations. The numerical investigation are based on the so-called spatial approach. A contravariant based governing system is derived so that we can simulate the receptivity and transition with complex geometries. The numerical results agreed very well with the experimental results for the flat plate case, and also agreed well with the result obtained by other researchers for the 2-D elliptic leading-edge receptivity case. Some new phenomena for the transition around Joukowsky airfoils were observed which at least qualitatively agree with physics.