Filters








9,376 Hits in 5.6 sec

Controlling Hallucinations at Word Level in Data-to-Text Generation [article]

Clément Rebuffel, Marco Roberti, Laure Soulier, Geoffrey Scoutheeten, Rossella Cancelliere, Patrick Gallinari
2021 arXiv   pre-print
Data-to-Text Generation (DTG) is a subfield of Natural Language Generation aiming at transcribing structured data in natural language descriptions.  ...  Our model is able to reduce and control hallucinations, while keeping fluency and coherence in generated texts.  ...  Keywords Data-to-Text Generation · Hallucinations · Controlled Text Generation Introduction Data-to-Text Generation (DTG) is the subfield of Computational Linguistics and Natural Language Generation  ... 
arXiv:2102.02810v2 fatcat:waiacytsubczvfqaesqlzpaglu

A Token-level Reference-free Hallucination Detection Benchmark for Free-form Text Generation [article]

Tianyu Liu, Yizhe Zhang, Chris Brockett, Yi Mao, Zhifang Sui, Weizhu Chen, Bill Dolan
2022 arXiv   pre-print
Existing work usually attempts to detect these hallucinations based on a corresponding oracle reference at a sentence or document level.  ...  However ground-truth references may not be readily available for many free-form text generation applications, and sentence- or document-level detection may fail to provide the fine-grained signals that  ...  Acknowledgments The authors would like to thank the anonymous reviewers for their thoughtful and constructive comments.  ... 
arXiv:2104.08704v2 fatcat:ozw6ioh32zhh3mmlu32s7vwxoy

Detecting Hallucinated Content in Conditional Neural Sequence Generation [article]

Chunting Zhou, Graham Neubig, Jiatao Gu, Mona Diab, Paco Guzman, Luke Zettlemoyer, Marjan Ghazvininejad
2021 arXiv   pre-print
Codes and data available at https://github.com/violet-zct/fairseq-detect-hallucination.  ...  We also apply our method to word-level quality estimation for MT and show its effectiveness in both supervised and unsupervised settings.  ...  Acknowledgements The work in this paper was supported in part by a Facebook SRA Award. We thank the anonymous reviewers for the insightful feedback that helps us improve the paper.  ... 
arXiv:2011.02593v3 fatcat:p4ldz52bhvgqnn2fpfntgqop5u

Survey of Hallucination in Natural Language Generation [article]

Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Yejin Bang, Andrea Madotto, Pascale Fung
2022 arXiv   pre-print
This advancement has led to more fluent and coherent NLG, leading to improved development in downstream tasks such as abstractive summarization, dialogue generation and data-to-text generation.  ...  downstream tasks, namely abstractive summarization, dialogue generation, generative question answering, data-to-text generation, and machine translation.  ...  Current works treat the hallucination level as a controllable attribute in order to remains the hallucination in outputs at a low level.  ... 
arXiv:2202.03629v4 fatcat:s6c26a7orncrffis55q5swo5ue

Hallucinated n-best lists for discriminative language modeling

K. Sagae, M. Lehr, E. Prud'hommeaux, P. Xu, N. Glenn, D. Karakos, S. Khudanpur, B. Roark, M. Saraclar, I. Shafran, D. Bikel, C. Callison-Burch (+7 others)
2012 2012 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)  
This paper investigates semi-supervised methods for discriminative language modeling, whereby n-best lists are "hallucinated" for given reference text and are then used for training n-gram language models  ...  We perform controlled experiments on a very strong baseline English CTS system, comparing three methods for simulating ASR output, and compare the results with training with "real" n-best list output from  ...  Machine translation methods Above we discussed methods for generating hallucinated ASR nbest lists using phone-based finite-state transducers. We now turn to generating confusions at the word level.  ... 
doi:10.1109/icassp.2012.6289043 dblp:conf/icassp/SagaeLPXGKKRSSBCCHHKLPR12 fatcat:emi5pcwldrgdrlvi26wtyh6aqy

Text Generation with Text-Editing Models [article]

Eric Malmi, Yue Dong, Jonathan Mallinson, Aleksandr Chuklin, Jakub Adamek, Daniil Mirylenka, Felix Stahlberg, Sebastian Krause, Shankar Kumar, Aliaksei Severyn
2022 arXiv   pre-print
In contrast, seq2seq models generate outputs word-by-word from scratch thus making them slow at inference time.  ...  We discuss challenges related to productionization and how these models can be used to mitigate hallucination and bias, both pressing challenges in the field of text generation.  ...  These advantages have generated a substantial and continued level of interest in text-editing research.  ... 
arXiv:2206.07043v1 fatcat:erjlc3paandfjn4w33kibfqgze

Controlled Hallucinations: Learning to Generate Faithfully from Noisy Data [article]

Katja Filippova
2020 arXiv   pre-print
Neural text generation (data- or text-to-text) demonstrates remarkable performance when training data is abundant which for many applications is not the case.  ...  Our contribution is a simple but powerful technique to treat such hallucinations as a controllable aspect of the generated text, without dismissing any input and without modifying the model architecture  ...  Word Overlap When the source and the target are similar on the token level, one can use word overlap between them to estimate how many words unsupported by the source are present in the target.  ... 
arXiv:2010.05873v1 fatcat:xarkkuubkvelbjejy46kakvwei

Learning to Revise References for Faithful Summarization [article]

Griffin Adams, Han-Chin Shing, Qing Sun, Christopher Winestock, Kathleen McKeown, Noémie Elhadad
2022 arXiv   pre-print
At inference, we vary style codes to over-generate revisions of unsupported reference sentences and select a final revision which balances faithfulness and abstraction.  ...  In many real-world scenarios with naturally occurring datasets, reference summaries are noisy and contain information that cannot be inferred from the source text.  ...  On a highly noisy dataset of concept-to-text generation (Lebret et al., 2016) , Filippova retains all data and addresses quality by treating hallucinations as a controllable aspect of the generated text  ... 
arXiv:2204.10290v1 fatcat:ijpetqjua5cyrol2ps3kt4q3xq

Believing is seeing in schizophrenia: The role of top-down processing

Duje Tadin, Peiyan Wong, Michael W. Mebane, Michael J. Berkowitz, Hollister Trott, Sohee Park
2005 Behavioral and Brain Sciences  
59 words Main Text: 982 words References: 359 words Total Text: 1524 words Abstract:  ...  Acknowledgments Supported in part by NARSAD Grable Investigator Award to Sohee Park.  ...  Namely, the role of perceptual deficits in RCVH is to simply increase the noise, whereas the actual generation of RCVH lies within faulty higher-level processes.  ... 
doi:10.1017/s0140525x05400138 fatcat:wpncsvvuknanper7avbdz5mj4m

Faithfulness in Natural Language Generation: A Systematic Survey of Analysis, Evaluation and Optimization Methods [article]

Wei Li, Wenhao Wu, Moye Chen, Jiachen Liu, Xinyan Xiao, Hua Wu
2022 arXiv   pre-print
abstractive summarization, dialogue generation, machine translation, and data-to-text generation.  ...  This advancement has resulted in more fluent, coherent and even properties controllable (e.g. stylistic, sentiment, length etc.) generation, naturally leading to development in downstream tasks such as  ...  Faithfulness in Data-to-Text Generation Data-to-text generation (or table-to-text generation) has been widely studied for decades.  ... 
arXiv:2203.05227v1 fatcat:q2u3ojyi6vb7pjt6ajwbinjmpa

Towards Faithfulness in Open Domain Table-to-text Generation from an Entity-centric View [article]

Tianyu Liu, Xin Zheng, Baobao Chang, Zhifang Sui
2021 arXiv   pre-print
In open domain table-to-text generation, we notice that the unfaithful generation usually contains hallucinated content which can not be aligned to any input table record.  ...  We thus try to evaluate the generation faithfulness with two entity-centric metrics: table record coverage and the ratio of hallucinated entities in text, both of which are shown to have strong agreement  ...  Acknowledgments We would like to thank the anonymous reviewers for the helpful discussions and suggestions.  ... 
arXiv:2102.08585v1 fatcat:5ofv2kustfdg3dqzjy7ufpjfne

ProtoDA: Efficient Transfer Learning for Few-Shot Intent Classification [article]

Manoj Kumar, Varun Kumar, Hadrien Glaude, Cyprien delichy, Aman Alok, Rahul Gupta
2021 arXiv   pre-print
Further, we apply data augmentation in conjunction with meta-learning to reduce sampling bias.  ...  We make use of a conditional generator for data augmentation that is trained directly using the meta-learning objective and simultaneously with prototypical networks, hence ensuring that data augmentation  ...  The encoder takes in a combination of character-level and word-level representations at the input. Character representations are learnt using a 2-D convolution neural network.  ... 
arXiv:2101.11753v1 fatcat:fmyl7vidyrdutlo2uumbhebjsq

A Neural Network Simulation of Hallucinated "Voices" and Associated Speech Perception Impairments in Schizophrenic Patients

Ralph E. Hoffman, Jill Rapaport, Rezvan Ameli, Thomas H. McGlashan, Diane Harcherik, David Servan-Schreiber
1995 Journal of Cognitive Neuroscience  
Pruning anatomic connections or reducing neuronal activation in working memory caused word "percepts" t o emerge spontaneously (i.e., in the absence of external "speech inputs"), thereby providing a model  ...  of hallucinated speech.  ...  The fact that the model could not be manipulated to fit performance data for all individual hallucinators predicts that factors other than pruning level and bias were at work, but that these factors, in  ... 
doi:10.1162/jocn.1995.7.4.479 pmid:23961906 fatcat:6cfinfgwezbe7o4af6hqphynle

Hallucinated but Factual! Inspecting the Factuality of Hallucinations in Abstractive Summarization [article]

Meng Cao, Yue Dong, Jackie Chi Kit Cheung
2021 arXiv   pre-print
State-of-the-art abstractive summarization systems often generate hallucinations; i.e., content that is not directly inferable from the source text.  ...  These factual hallucinations can be beneficial in a summary by providing useful background information.  ...  Filippova (2020) proposed a method for controlling hallucination in data-to-text generation task.  ... 
arXiv:2109.09784v2 fatcat:xeowmhalbnebpfdlvop64sezce

A Simple Recipe towards Reducing Hallucination in Neural Surface Realisation

Feng Nie, Jin-Ge Yao, Jinpeng Wang, Rong Pan, Chin-Yew Lin
2019 Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics  
at Microsoft.  ...  To mitigate this issue, we propose to integrate a language understanding module for data refinement with selftraining iterations to effectively induce strong equivalence between the input data and the  ...  The contact author of this paper, according to the meaning given to this role by Sun Yat-Sen University, is Rong Pan.  ... 
doi:10.18653/v1/p19-1256 dblp:conf/acl/NieYWPL19 fatcat:jed4iroe2nfadpc6igvfmesglm
« Previous Showing results 1 — 15 out of 9,376 results