Filters








212 Hits in 2.2 sec

Neural Text Generation with Unlikelihood Training [article]

Sean Welleck, Ilia Kulikov, Stephen Roller, Emily Dinan, Kyunghyun Cho, Jason Weston
2019 arXiv   pre-print
Neural text generation is a key tool in natural language applications, but it is well known there are major problems at its core.  ...  We show that both token and sequence level unlikelihood training give less repetitive, less dull text while maintaining perplexity, giving superior generations using standard greedy or beam search.  ...  THE UNLIKELIHOOD TRAINING OBJECTIVE We now describe unlikelihood training for neural language models, then in Section 6 demonstrate empirically that our proposal substantially improves neural text degeneration  ... 
arXiv:1908.04319v2 fatcat:kr32627xcncaro6nwtjppnziae

Implicit Unlikelihood Training: Improving Neural Text Generation with Reinforcement Learning [article]

Evgeny Lagutin and Daniil Gavrilov and Pavel Kalaidin
2021 arXiv   pre-print
We apply this approach to minimizing repetition in generated text, and show that, when combined with unlikelihood training (Welleck et al., 2020), our method further reduces repetition without impacting  ...  We also evaluate other methods for improving generation at training and decoding time, and compare them using various metrics aimed at control for better text generation output.  ...  Neural text generation with unlikelihood training. In International Conference on Learning Representa- tions. Table 3 : 3 Repetition with Beam Search.  ... 
arXiv:2101.04229v1 fatcat:gmnuvcjm6je2hgefc7tfcwedyq

Implicit Unlikelihood Training: Improving Neural Text Generation with Reinforcement Learning

Evgeny Lagutin, Daniil Gavrilov, Pavel Kalaidin
2021 Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume   unpublished
We apply this approach to minimizing repetition in generated text, and show that, when combined with unlikelihood training (Welleck et al., 2020), our method further reduces repetition without impacting  ...  We also evaluate other methods for improving generation at training and decoding time, and compare them using various metrics aimed at control for better text generation output.  ...  Neural text generation with unlikelihood training. In International Conference on Learning Representa- tions. Table 3 : 3 Repetition with Beam Search.  ... 
doi:10.18653/v1/2021.eacl-main.123 fatcat:ulx5poizazek7j27jk25hagbfe

Language Model Evaluation in Open-ended Text Generation [article]

An Nguyen
2021 arXiv   pre-print
in open-ended text generation.  ...  In this work, we study different evaluation metrics that have been proposed to evaluate quality, diversity and consistency of machine-generated text.  ...  To find out how sensitive unlikelihood training is to the number of training epochs, we also train another model using unlikelihood training with 1 epoch only.  ... 
arXiv:2108.03578v1 fatcat:foqvdu4i3rgovoqnxin72bt6e4

Diverse Keyphrase Generation with Neural Unlikelihood Training [article]

Hareesh Bahuleyan, Layla El Asri
2020 arXiv   pre-print
Our findings show that repetition of keyphrases is a major issue with MLE training. To alleviate this issue, we adopt neural unlikelihood (UL) objective for training the S2S model.  ...  Our version of UL training operates at (1) the target token level to discourage the generation of repeating tokens; (2) the copy token level to avoid copying repetitive tokens from the source text.  ...  Thus neural unlikelihood training (Welleck et al., 2020) is well suited to our problem.  ... 
arXiv:2010.07665v1 fatcat:kwwttfgxdbgx7nnuazqhkhhtky

A Simple Contrastive Learning Objective for Alleviating Neural Text Degeneration [article]

Shaojie Jiang, Ruqing Zhang, Svitlana Vakulenko, Maarten de Rijke
2022 arXiv   pre-print
Comprehensive experiments on language modeling and open-domain dialogue generation tasks show that the proposed contrastive token objective yields less repetitive texts, with a higher generation quality  ...  than unlikelihood training, achieving the new state-of-the-art performance.  ...  Thanks to Maartje ter Hoeve and Mohammad Aliannejadi for help with the MTurk task design.  ... 
arXiv:2205.02517v1 fatcat:ntvsej6iqbahjh6oekscjxqiby

Beyond [CLS] through Ranking by Generation [article]

Cicero Nogueira dos Santos, Xiaofei Ma, Ramesh Nallapati, Zhiheng Huang, Bing Xiang
2020 arXiv   pre-print
However, with the advent of modern deep neural networks, attention has shifted to discriminative ranking functions that model the semantic similarity of documents and queries instead.  ...  Recently, deep generative models such as GPT2 and BART have been shown to be excellent text generators, but their effectiveness as rankers have not been demonstrated yet.  ...  The usual approach to train an LM using a neural network with parameters θ consists on performing maximum likelihood estimation (MLE) by minimizing the negative log-likelihood over a large text corpus  ... 
arXiv:2010.03073v1 fatcat:bbhobahdgje4xbgjtzo5g3nviy

A Contrastive Framework for Neural Text Generation [article]

Yixuan Su and Tian Lan and Yan Wang and Dani Yogatama and Lingpeng Kong and Nigel Collier
2022 arXiv   pre-print
Existing approaches introduce stochasticity via sampling or modify training objectives to decrease probabilities of certain tokens (e.g., unlikelihood training).  ...  However, maximization-based decoding methods (e.g. beam search) of neural language models often lead to degenerate solutions -- the generated text is unnatural and contains undesirable repetitions.  ...  A Related Work Neural Text Generation. Neural text generation is a core component in many NLP applications.  ... 
arXiv:2202.06417v1 fatcat:v66u5bnagzdi3cnhniaphvfrvi

Understanding by Understanding Not: Modeling Negation in Language Models [article]

Arian Hosseini, Siva Reddy, Dzmitry Bahdanau, R Devon Hjelm, Alessandro Sordoni, Aaron Courville
2021 arXiv   pre-print
To improve language models in this regard, we propose to augment the language modeling objective with an unlikelihood objective that is based on negated generic sentences from a raw text corpus.  ...  By training BERT with the resulting combined objective we reduce the mean top~1 error rate to 4% on the negated LAMA dataset. We also see some improvements on the negated NLI benchmarks.  ...  Neu- ral text generation with unlikelihood training. In ICLR. OpenReview.net. Adina Williams, Nikita Nangia, and Samuel Bowman. 2018.  ... 
arXiv:2105.03519v1 fatcat:j6xzqndrfva3lf3bt2gwhluql4

Straight to the Gradient: Learning to Use Novel Tokens for Neural Text Generation [article]

Xiang Lin, Simeng Han, Shafiq Joty
2021 arXiv   pre-print
With the simplicity in architecture, our method can serve as a general training objective that is applicable to most of the neural text generation tasks.  ...  Advanced large-scale neural language models have led to significant success in many language generation tasks.  ...  To address the issue with MLE, Welleck et al. (2020) propose the neural unlikelihood (UL) training method.  ... 
arXiv:2106.07207v1 fatcat:3nhktiziwnewzbc2zvquxrqfmi

Don't Say That! Making Inconsistent Dialogue Unlikely with Unlikelihood Training [article]

Margaret Li, Stephen Roller, Ilia Kulikov, Sean Welleck, Y-Lan Boureau, Kyunghyun Cho, Jason Weston
2020 arXiv   pre-print
Generative dialogue models currently suffer from a number of problems which standard maximum likelihood training does not address.  ...  models with greater reasoning ability.  ...  Introduction Open-ended tasks such as dialogue reveal a number of issues with current neural text generation methods.  ... 
arXiv:1911.03860v2 fatcat:bwogo2yijjcmtp6p3oj3vbhgge

Reinforcement Learning for Few-Shot Text Generation Adaptation [article]

Cheng Pengsen, Dai Jinqiao, Liu Jiayong
2021 arXiv   pre-print
To address this shortcoming, we frame the adaptation of text generation systems as a reinforcement learning problem and provide a new approach to make text generation models easily adaptable to target  ...  However, the texts generated by few-shot learning are typically devoid of linguistic diversity.  ...  The goal of unlikelihood training is to improve neural text degeneration [29] .  ... 
arXiv:2111.11030v1 fatcat:a66p54623bgyhaogegvfhu64zm

Diversifying Neural Dialogue Generation via Negative Distillation [article]

Yiwei Li, Shaoxiong Feng, Bin Sun, Kan Li
2022 arXiv   pre-print
First, we introduce a negative teacher model that can produce query-wise generic responses, and then the student model is required to maximize the distance with multi-level negative knowledge.  ...  Recently, an interesting approach, namely negative training, has been proposed to alleviate this problem by reminding the model not to generate high-frequency responses during training.  ...  However, with the MLE-based training, this phenomenon will cause the model to produce generic responses.  ... 
arXiv:2205.02795v1 fatcat:bbdl7l4f4fhnbca4yqlwafaz7m

Learning to Generate Code Comments from Class Hierarchies [article]

Jiyang Zhang, Sheena Panthaplackel, Pengyu Nie, Raymond J. Mooney, Junyi Jessy Li, Milos Gligoric
2021 arXiv   pre-print
of the overriding method; and (3) unlikelihood training to discourage predictions which do not conform to invariant characteristics of the comment corresponding to the overridden method.  ...  Our experiments show that the proposed approach is able to generate comments for overriding methods of higher quality compared to prevailing comment generation techniques.  ...  To do this, we use unlikelihood training with synthetically generated sequences that do not conform to invariant characteristics of the overridden comment.  ... 
arXiv:2103.13426v2 fatcat:5wp7tc3q2nfizau5qhnh4w7h2y

Tailor: Generating and Perturbing Text with Semantic Controls [article]

Alexis Ross, Tongshuang Wu, Hao Peng, Matthew E. Peters, Matt Gardner
2022 arXiv   pre-print
However, current techniques rely on training a model for every target perturbation, which is expensive and hard to generalize. We present Tailor, a semantically-controlled text generation system.  ...  Controlled text perturbation is useful for evaluating and improving model generalizability.  ...  Next, we explain how to embed them within inputs ( §2.2) to the generator. We train the generator to follow control codes with unlikelihood training ( §2.3).  ... 
arXiv:2107.07150v2 fatcat:bhayvtxw5rfvdbvxlatzdymcpq
« Previous Showing results 1 — 15 out of 212 results