Referring to what you know and do not know: Making Referring Expression Generation Models Generalize To Unseen Entities

Abstract

Data-to-text Natural Language Generation (NLG) is the computational process of generating natural language in the form of text or voice from non-linguistic data. A core micro-planning task within NLG is referring expression generation (REG), which aims to automatically generate noun phrases to refer to entities mentioned as discourse unfolds. A limitation of novel REG models is not being able to generate referring expressions to entities not encountered during the training process. To solve this problem, we propose two extensions to NeuralREG, a state-ofthe-art encoder-decoder REG model. The first is a copy mechanism, whereas the second consists of representing the gender and type of the referent as inputs to the model. Drawing on the results of automatic and human evaluation as well as an ablation study using the WebNLG corpus, we contend that our proposal contributes to the generation of more meaningful referring expressions to unseen entities than the original system and related work. Code and all produced data are publicly available 12 .

Non-neural Models Matter: A Re-evaluation of Neural Referring Expression Generation Systems

“…We evaluated all the systems described in §3 on both WEBNLG and WSJ using automatic and human evaluations. We implemented the neural models based on the code of Cunha et al (2020) and Cao and Cheung (2019) 7 . For WEBNLG, we used their original parameter setting, while for WSJ, we tuned the parameters on the development set and used the best parameter set.…”

“…ATT+Copy. Cunha et al (2020) proposed using three bidirectional LSTMs (Hochreiter and Schmidhuber, 1997) to encode a pre-context, a post-context, and the proper name of an entity (i.e., replacing underscores in entity labels with whitespaces) into three hidden vectors h (pre) , h (post) and h (r) , respectively. An auto-regressive LSTM-based decoder generates REs based on context vectors.…”

“…ATT+Meta. ATT+Meta (Cunha et al, 2020) used meta information of each entity to improve the quality of the generated REs. In each decoding step t, the context vector v…”

“…Metrics. Following Cunha et al (2020), we evaluated REG systems from 3 angles. (1) RE Accuracy and String Edit Distance (SED, Levenshtein, 1966) were used to evaluate the quality of each generated RE.…”

“…Thanks to the rapid development of deep learning techniques, recent NLG models are able to generate RE in an End2End (E2E) manner, i.e., to tackle the selection of form and content simultaneously (Castro Ferreira et al, 2018a;Cao and Cheung, 2019;Cunha et al, 2020). The task of End2End (E2E) REG was proposed by Castro Ferreira et al (2018a), who extracted a corresponding corpus from the WebNLG corpus (Castro Ferreira et al, 2018b) 1 .…”