[text generation] recommended in the collection of papers - Stanford researchers introduce time control methods to make long text generation more smooth

In a recent study , A research group at Stanford University proposed time control (TC), This language model plans implicitly through potential random processes , And generate text consistent with the potential plan , To improve the performance of long text generation . Let's take a look at the text generation ：

Text generation is an important research field in natural language processing , It has a broad application prospect . At home and abroad, there have been such as Automated Insights、Narrative Science as well as “ Xiaonan ” Robot and “ Xiao Ming ” Text generation systems such as robots are put into use . These systems generate news based on formatted data or natural language text 、 Financial report or other explanatory text .

AMiner Prepared for you 【 The text generated 】 Collection of papers , Click the link to view ba ：https://www.aminer.cn/topic/61cc80ef42323c8f84232085?f=cs

The following are some high-quality papers sorted out from it ：

1.SALSA-TEXT : self attentive latent space based adversarial text generation.

“ We propose a novel architecture based on self attention , To improve the performance of adversarial potential code schemes in text generation … In this paper , We took a step to strengthen the architecture used in these settings , especially AAE and ARAE. We discuss two potential code approaches based on adversarial settings （AAE and ARAE） Benchmarking … Experiments show that , The proposed （ Self - ） Attention based models are superior to the most advanced models in text generation based on adversarial code …”

PDF Download link ：https://www.aminer.cn/pub/5bdc31c217c44a1f58a0cc6c/?f=cs

2.Text Generation Based on Generative Adversarial Nets with Latent Variable.

“ In this paper , We propose a generation countermeasure network （GAN） Generate realistic text models … We proposed VGAN Model , The generated model is composed of recurrent neural network and VAE form . The discriminant model is a convolutional neural network . We use the strategy gradient training model . We apply the proposed model to the text generation task , And compare it with other recent models based on Neural Network , For example, recursive neural network language model and Seq-GAN…”

PDF Download link ：https://www.aminer.cn/pub/5a73cbcc17c44a0b3035f5f0/?f=cs

3.Content preserving text generation with attribute controls.

“ In this work , We solved the problem of modifying the text attributes of sentences . Given an input sentence and a set of attribute tags , We try to generate sentences compatible with conditional information . To ensure that the model generates content compatible sentences , We introduced a reconstruction loss , This loss is interpolated between the automatic encoding and the reverse translation loss components …”

PDF Download link ：https://www.aminer.cn/pub/5c8fb7d04895d9cbc65d4525/?f=cs

4.Optimizing Referential Coherence in Text Generation

“ This paper describes an implemented system , The system uses centering theory to plan the selection of coherent text and quotation expressions . We think , Text and sentence planning needs to be driven in part by the goal of maintaining referential continuity , So as to promote the resolution of pronouns ： Get favorable ordering of parameters in clauses and clauses , It may increase the chance of using unambiguous pronouns …”

PDF Download link ：https://www.aminer.cn/pub/53e9b0b2b7602d9703b1e972/?f=cs

5.Text generation from keywords

“ We describe a method from “ keyword ” or “ Title words ” Methods of generating sentences . The method consists of two main parts , Candidate text construction and evaluation … The model considers not only words n-gram Information , Also consider the dependency information between words . Besides , It also considers string information and morphological information …”

PDF Download link ：https://www.aminer.cn/pub/53e99adcb7602d9702360b5a/?f=cs

Click the link to enter AMiner Official website , See more good papers ~https://www.aminer.cn/?f=cs