Bitte benutzen Sie diese Referenz, um auf diese Ressource zu verweisen:
Volltext verfügbar? / Dokumentlieferung
doi:10.22028/D291-38656
Titel: | Neural Data-to-Text Generation with LM-based Text Augmentation |
VerfasserIn: | Chang, Ernie Shen, Xiaoyu Zhu, Dawei Demberg, Vera Su, Hui |
Sprache: | Englisch |
Verlag/Plattform: | arXiv |
Erscheinungsjahr: | 2021 |
DDC-Sachgruppe: | 400 Sprache, Linguistik |
Dokumenttyp: | Sonstiges |
Abstract: | For many new application domains for data-to-text generation, the main obstacle in training neural models consists of a lack of training data. While usually large numbers of instances are available on the data side, often only very few text samples are available. To address this problem, we here propose a novel few-shot approach for this setting. Our approach automatically augments the data available for training by (i) generating new text samples based on replacing specific values by alternative ones from the same category, (ii) generating new text samples based on GPT-2, and (iii) proposing an automatic method for pairing the new text samples with data samples. As the text augmentation can introduce noise to the training data, we use cycle consistency as an objective, in order to make sure that a given data sample can be correctly reconstructed after having been formulated as text (and that text samples can be reconstructed from data). On both the E2E and WebNLG benchmarks, we show that this weakly supervised training paradigm is able to outperform fully supervised seq2seq models with less than 10% annotations. By utilizing all annotated data, our model can boost the performance of a standard seq2seq model by over 5 BLEU points, establishing a new state-of-the-art on both datasets. |
URL der Erstveröffentlichung: | https://arxiv.org/abs/2102.03556 |
Link zu diesem Datensatz: | urn:nbn:de:bsz:291--ds-386560 hdl:20.500.11880/34847 http://dx.doi.org/10.22028/D291-38656 |
Datum des Eintrags: | 4-Jan-2023 |
Bemerkung/Hinweis: | Preprint |
Fakultät: | MI - Fakultät für Mathematik und Informatik |
Fachrichtung: | MI - Informatik |
Professur: | MI - Prof. Dr. Vera Demberg |
Sammlung: | SciDok - Der Wissenschaftsserver der Universität des Saarlandes |
Dateien zu diesem Datensatz:
Es gibt keine Dateien zu dieser Ressource.
Alle Ressourcen in diesem Repository sind urheberrechtlich geschützt.