Source author record

Jiayu Tang

Jiayu Tang appears in the imported research catalog. Authorship, coauthor and topic links are available while profile ownership is still unclaimed.

ResearcherUnclaimed source record

astro-ph.CO astro-ph Computation and Language gr-qc

Catalog footprint

What is connected

5works

4topics

4close collaborators

Actions

Connect this record

Open graph Browse works

Inspect adjacent papers, topics, institutions and collaborators without losing the researcher page.

Building this map preview

BZPEER is loading the nearby papers, people, topics and institutions for this page.

preprint2021arXiv

AT-BERT: Adversarial Training BERT for Acronym Identification Winning Solution for SDU@AAAI-21

Acronym identification focuses on finding the acronyms and the phrases that have been abbreviated, which is crucial for scientific document understanding tasks. However, the limited size of manually annotated datasets hinders further improvement for the problem. Recent breakthroughs of language models pre-trained on large corpora clearly show that unsupervised pre-training can vastly improve the performance of downstream tasks. In this paper, we present an Adversarial Training BERT method named AT-BERT, our winning solution to acronym identification task for Scientific Document Understanding (SDU) Challenge of AAAI 2021. Specifically, the pre-trained BERT is adopted to capture better semantic representation. Then we incorporate the FGM adversarial training strategy into the fine-tuning of BERT, which makes the model more robust and generalized. Furthermore, an ensemble mechanism is devised to involve the representations learned from multiple BERT variants. Assembling all these components together, the experimental results on the SciAI dataset show that our proposed approach outperforms all other competitive state-of-the-art methods.

preprint2015arXiv

Cosmological Structure Formation in Decaying Dark Matter Models

The standard cold dark matter (CDM) model predicts too many and too dense small structures. We consider an alternative model that the dark matter undergoes two-body decays with cosmological lifetime $τ$ into only one type of massive daughters with non-relativistic recoil velocity $V_k$. This decaying dark matter model (DDM) can suppress the structure formation below its free-streaming scale at time scale comparable to $τ$. Comparing with warm dark matter (WDM), DDM can better reduce the small structures while being consistent with high redshfit observations. We study the cosmological structure formation in DDM by performing self-consistent N-body simulations and point out that cosmological simulations are necessary to understand the DDM structures especially on non-linear scales. We propose empirical fitting functions for the DDM suppression of the mass function and the mass-concentration relation, which depend on the decay parameters lifetime $τ$ and recoil velocity $V_k$, and redshift. The fitting functions lead to accurate reconstruction of the the non-linear power transfer function of DDM to CDM in the framework of halo model. Using these results, we set constraints on the DDM parameter space by demanding that DDM does not induce larger suppression than the Lyman-$α$ constrained WDM models. We further generalize and constrain the DDM models to initial conditions with non-trivial mother fractions and show that the halo model predictions are still valid after considering a global decayed fraction. Finally, we point out that the DDM is unlikely to resolve the disagreement on cluster numbers between the Planck primary CMB prediction and the Sunyaev-Zeldovich (SZ) effect number count for $τ\sim H_{0}^{-1}$.

preprint2014arXiv

Cosmic microwave background constraints on the tensor-to-scalar ratio

One of the main goals of modern cosmic microwave background (CMB) missions is to measure the tensor-to-scalar ratio $r$ accurately to constrain inflation models. Due to ignorance about the reionization history $X_{e}(z)$, this analysis is usually done by assuming an instantaneous reionization $X_{e}(z)$ which, however, can bias the best-fit value of $r$. Moreover, due to the strong mixing of B-mode and E-mode polarizations in cut-sky measurements, multiplying the sky coverage fraction $f_{sky}$ by the full-sky likelihood would not give satisfactory results. In this work, we forecast constraints on $r$ for the Planck mission taking into account the general reionization scenario and cut-sky effects. Our results show that by applying an N-point interpolation analysis to the reionization history, the bias induced by the assumption of instantaneous reionization is removed and the value of $r$ is constrained within $5\%$ error level, if the true value of $r$ is greater than about 0.1 .

preprint2011arXiv

Likelihood reconstruction method of real-space density and velocity power spectra from a redshift galaxy survey

We develop a maximum likelihood based method of reconstructing band powers of the density and velocity power spectra at each wavenumber bins from the measured clustering features of galaxies in redshift space, including marginalization over uncertainties inherent in the Fingers-of-God (FoG) effect. The reconstruction can be done assuming that the density and velocity power spectra depend on the redshift-space power spectrum having different angular modulations of mu with mu^{2n} (n=0,1,2) and that the model FoG effect is given as a multiplicative function in the redshift-space spectrum. By using N-body simulations and the halo catalogs, we test our method by comparing the reconstructed power spectra with the simulations. For the spectrum of mu^0 or equivalently the density power spectrum P_dd(k), our method recovers the amplitudes to a few percent accuracies up to k=0.3 h/Mpc for both dark matter and halos. For the power spectrum of mu^2, which is equivalent to the density-velocity spectrum P_dv(k) in the linear regime, our method can recover the input power spectrum for dark matter up to k=0.2 h/Mpc and at both z=0 and 1, if using the adequate FoG model. However, for the halo spectrum, the reconstructed spectrum shows greater amplitudes than the simulation P_dv(k). We argue that the disagreement is ascribed to nonlinearity effect that arises from the cross-bispectra of density and velocity perturbations. Using the perturbation theory, we derive the nonlinear correction term, and find that the leading-order correction term is proportional to mu^2 and increases the mu^2-power spectrum amplitudes at larger k, at lower redshifts and for more massive halos. We find that adding the nonlinearity correction term to the simulation P_dv(k) can fairly well reproduce the reconstructed P_dv(k) for halos up to k~0.2 h/Mpc.

preprint2008arXiv

Complementarity of Future Dark Energy Probes

In recent years a plethora of future surveys have been suggested to constrain the nature of dark energy. In this paper we adapt a binning approach to the equation of state factor ``w'' and discuss how future weak lensing, galaxy cluster counts, Supernovae and baryon acoustic oscillation surveys constrain the equation of state at different redshifts. We analyse a few representative future surveys, namely DES, PS1, WFMOS, PS4, EUCLID, SNAP and SKA, and perform a principal component analysis for the ``w'' bins. We also employ a prior from Planck cosmic microwave background measurements on the remaining cosmological parameters. We study at which redshifts a particular survey constrains the equation of state best and how many principal components are significantly determined. We then point out which surveys would be sufficiently complementary. We find that weak lensing surveys, like EUCLID, would constrain the equation of state best and would be able to constrain of the order of three significant modes. Baryon acoustic oscillation surveys on the other hand provide a unique opportunity to probe the equation of state at relatively high redshifts.