Abstractive Document Summarization with Word Embedding Reconstruction
20162Authors
nan
Abstract
AbstractNeural sequence-to-sequence (Seq2Seq) models and BERT have achieved substantial improvements in abstractive document summarization (ADS) without and with pre-training, respectively. However, they sometimes repeatedly attend to unimportant source phrases while mistakenly ignore important ones. We present reconstruction mechanisms on two levels to alleviate this issue. The sequence-level reconstructor reconstructs the whole document from the hidden layer of the target summary, while the word embedding-level one rebuilds the average of word embeddings of the source at the target side to guarantee that as much critical information is included in the summary as possible. Based on the assumption that inverse document frequency (IDF) measures how important a word is, we further leverage the IDF weights in our embedding-level reconstructor. The proposed frameworks lead to promising improvements for ROUGE metrics and human rating on both the CNN/Daily Mail and Newsroom summarization datasets.
Keywords
Full Content
Proceedings of Recent Advances in Natural Language Processing, pages 138–146 Sep 1–3, 2021. https://doi.org/10.26615/978-954-452-072-4_017 138 Cross-Lingual Wolastoqey-English Definition Modelling Diego Bear Faculty of Computer Science University of New Brunswick diego.bear@unb.ca Paul Cook Faculty of Computer Science University of New Brunswick paul.cook@unb.ca Abstract Definition modelling is the task of automat- ically generating a dictionary-style definition given a target word. In this paper, we consider cross-lingual definition generation. Specif- ically, we generate English definitions for Wolastoqey (Malecite-Passamaquoddy) words. Wolastoqey is an endangered, low-resource polysynthetic language. We hypothesize that sub-word representations based on byte pair encoding (Sennrich et al., 2016) can be lever- aged to represent morphologically-complex Wolastoqey words and overcome the challenge of not having large corpora available for train- ing. Our experimental results demonstrate that this approach outperforms baseline methods in terms of BLEU score. 1 Introduction Definition modelling, introduced by Noraset et al. (2017), is the task of automatically generating a dictionary-style definition for a given target word. Definition modelling can provide more-transparent, human-interpretable representations of the infor- mation in embeddings. Definition modelling could also potentially be applied to automate, or semi- automate, the constructing or updating of dictio- naries, for example, by generating draft definitions for newly-emerged words that are not yet listed. Although there has been a range of work on defini- tion modelling (e.g., Ni and Wang, 2017; Gadetsky et al., 2018; Chang and Chen, 2019) the focus has been on monolingual definition modelling, with the target word and generated definition being in the same language. Malecite-Passamaquoddy (also Maliseet- Passamaquoddy, Passamaquoddy-Maliseet) is an Eastern Algonquian language spoken in regions of what is now New Brunswick and Quebec, Canada, and Maine, United States. Malecite and Passamaquoddy are dialects of this language. However, Malecite is a Mi’kmaq exonym, with Wolastoqey being the term this speech community uses to refer to their language. We therefore use the term Wolastoqey throughout this paper. Wolastoqey is an endangered language, with roughly 300 remaining first language speakers in Canada (Statistics Canada, 2017). Moreover, chil- dren are typically not learning the language pro- ficiently. Wolastoqey is also a low-resource lan- guage, with no large corpora or annotated datasets available for training natural language processing (NLP) systems. However, the Passamaquoddy- Maliseet Dictionary (Francis and Leavitt, 2008) is available online through the Passamaquoddy- Maliseet Language Portal.1 This dictionary in- cludes roughly 19k entries with Wolastoqey head- words and English definitions. Many entries also include parallel Wolastoqey-English example sen- tences. There has been very little prior computa- tional work on Wolastoqey, with Farber (2015) pre- senting a preliminary finite-state model of nouns. Wolastoqey, like other Algonquian languages, is polysynthetic. Verbs in particular have rich mor- phological structure, and often include several roots (Leavitt, 1996). Consider the example gloss below for paskoloqessu: pask-oloq-ess-u breaking-ice-move.quickly-s/he ‘She or he moves quickly across ice as it cracks’ The root oloq can be seen in various other words, such as ’ketoloqtehmon ‘s/he chips it out of ice’, sahsoloqe ‘it is slippery, is icy’, and supoloqe ‘there is smooth ice (on lake, etc.)’. …
Content truncated to 500 words
Labeled Text
AbstractNeural sequence-to-sequence (Seq2Seq) models and BERT have achieved substantial improvements in abstractive document summarization (ADS) without and with pre-training, respectively. However, they sometimes repeatedly attend to unimportant source phrases while mistakenly ignore important ones. We present reconstruction mechanisms on two levels to alleviate this issue. The sequence-level reconstructor reconstructs the whole document from the hidden layer of the target summary, while the word embedding-level one rebuilds the average of word embeddings of the source at the target side to guarantee that as much critical information is included in the summary as possible. Based on the assumption that inverse document frequency (IDF) measures how important a word is, we further leverage the IDF weights in our embedding-level reconstructor. The proposed frameworks lead to promising improvements for ROUGE metrics and human rating on both the CNN/Daily Mail and Newsroom summarization datasets.
Publication Details
- Paper ID:
- 20162
- Year:
- 2021
- Publication:
- TIPSTER
- Publication Type:
- conf
- DOI:
- nan