Transforming Multi-Conditioned Generation from Meaning Representation

Proceedings of Recent Advances in Natural Language Processing, pages 805–813 Sep 1–3, 2021. https://doi.org/10.26615/978-954-452-072-4_092 805 Transforming Multi-Conditioned Generation from Meaning Representation Joosung Lee Kakao Enterprise Corp., South Korea rung.joo@kakaoenterprise.com Abstract Our study focuses on language generation by considering various information repre- senting the meaning of utterances as multi- ple conditions of generation. Generating an utterance from a Meaning representation (MR) usually passes two steps: sentence planning and surface realization. However, we propose a simple one-stage framework to generate utterances directly from MR. Our model is based on GPT2 and gener- ates utterances with flat conditions on slot and value pairs, which does not need to de- termine the structure of the sentence. We evaluate several systems in the E2E dataset with 6 automatic metrics. Our system is a simple method, but it demonstrates com- parable performance to previous systems in automated metrics. In addition, using only 10% of the dataset without any other techniques, our model achieves compara- ble performance, and shows the possibility of performing zero-shot generation and ex- panding to other datasets. 1 Introduction In many conversation systems, generating sen- tences with specific information is useful. For ex- ample, it can be used in chatbot systems or spoken dialogue systems to generate utterances that con- tain meaning representations (MRs) corresponding to a user’s query. In order to train the NLG sys- tem that reflects this variety of information, a large amount of labeled data is required. At the 2017 E2E challenge (Duˇsek et al., 2019), a large dataset was released, which consisted of pairs of MRs rep- resenting restaurant reviews and corresponding ut- terances. Table 1 shows an example. MRs can be regarded as the multi-conditions type of utterance MR (slot[value]) name[Giraffe], eatType[pub], food[Fast food], area[riverside], familyFriendly[yes] Utterance On the riverside the Giraffe is a Fast food, kid friendly pub. Table 1: An example of an E2E dataset consists of pairs of a MR and an utterance. generation, which consists of slots and values, and the corresponding utterances are references writ- ten by humans. We focus on training the model to generate utterances directly from MRs. Many of the previous NLG research are a two- stage approach through sentence planning and sur- face realization. Sentence planning determines the overall sentence structure and surface realization is the process of flattening the sentence structure. In recent studies (Konstas and Lapata, 2013; Duˇsek and Jurˇc´ıˇcek, 2016; Juraska et al., 2018), these two stages are processed at once by learning end-to-end without aligned data with a neural network. When generating sentences from an input (flat or structured MR), there are a template-based ap- proach and a neural network-based approach. Smi- ley et al. (2018); Puzikov and Gurevych (2018); Wiseman et al. (2018) generate sentences based on template. The template method is to obtain the structural sets of sentences corresponding to MRs from training data and apply the template appro- priate to the test data to generate the sentences. In Smiley et al. (2018); Puzikov and Gurevych (2018), a template is formed based …

Content truncated to 500 words