Referring Expression Generation Using Entity Profiles

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 3163–3172, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 3163 Referring Expression Generation Using Entity Profiles Meng Cao Jackie Chi Kit Cheung School of Computer Science, McGill University, Montreal, QC, Canada MILA, Montreal, QC, Canada {meng.cao@mail, jcheung@cs}.mcgill.ca Abstract Referring Expression Generation (REG) is the task of generating contextually appropriate references to entities. A limitation of exist- ing REG systems is that they rely on entity- specific supervised training, which means that they cannot handle entities not seen during training. In this study, we address this in two ways. First, we propose task setups in which we specifically test a REG system’s ability to generalize to entities not seen during training. Second, we propose a profile-based deep neu- ral network model, PROFILEREG, which en- codes both the local context and an external profile of the entity to generate reference real- izations. Our model generates tokens by learn- ing to choose between generating pronouns, generating from a fixed vocabulary, or copy- ing a word from the profile. We evaluate our model on three different splits of the WebNLG dataset, and show that it outperforms competi- tive baselines in all settings according to auto- matic and human evaluations. 1 Introduction Entities can be expressed by various types of lin- guistic expressions, including by their names (e.g., Barrack Obama), a definite description (e.g., the former president of the United States), a pronoun (e.g., he, him, his), or a demonstrative (e.g., that person). Many factors play a role in determining what type of expression is appropriate in a partic- ular context (Henschel et al., 2000), including in- formation status, familiarity with the entity, and referential clarity. In this study, we aim to design a model that can generate appropriate referring ex- pressions of entities in an extended passage. Such a system is useful in a variety of natural language generation settings, from dialogue systems to au- tomatic summarization applications (Reiter and Dale, 2000; Krahmer and Van Deemter, 2012). Referring expression generation (REG) can be broken into two steps. The first is to decide the form of referring expression; that is, what type of reference should be used (e.g., a proper name). The second is to determine the content of the re- ferring expression (e.g., Ernest Hemingway). Many computational approaches, both rule- based and machine-learning-based, have been proposed for REG. Rule-based models use pre- defined heuristics and algorithms to determine referential form (Reiter and Dale, 2000; Hen- schel et al., 2000; Callaway and Lester, 2002). Machine learning-based approaches require train- ing on samples to predict referring expressions (Nenkova and McKeown, 2003; Greenbacker and McCoy, 2009; Ferreira et al., 2016, 2017), often as a classification task. The common limitation of previous REG sys- tems is that they are incapable of generating refer- ring expressions for new, unseen entities. Previous REG setups have tended to focus on form selection rather than content …

Content truncated to 500 words