WUT at SemEval-2019 Task 9: Domain-Adversarial Neural Networks for Domain Adaptation in Suggestion Mining

Proceedings of the 13th International Workshop on Semantic Evaluation (SemEval-2019), pages 1262–1266 Minneapolis, Minnesota, USA, June 6–7, 2019. ©2019 Association for Computational Linguistics 1262 WUT at SemEval-2019 Task 9: Domain-Adversarial Neural Networks for Domain Adaptation in Suggestion Mining Mateusz Klimaszewski Piotr Andruszkiewicz Institute of Computer Science Warsaw University of Technology, Poland mk.klimaszewski@gmail.com, P.Andruszkiewicz@ii.pw.edu.pl Abstract We present a system for cross-domain sugges- tion mining, prepared for the SemEval-2019 Task 9: Suggestion Mining from Online Re- views and Forums (Subtask B). Our submit- ted solution for this text classification problem explores the idea of treating different sugges- tions’ sources as one of the settings of Trans- fer Learning - Domain Adaptation. Our ex- periments show that without any labeled tar- get domain examples during training time, we are capable of proposing a system, reach- ing up to 0.778 in terms of F1 score on test dataset, based on Target Preserving Domain- Adversarial Neural Networks. 1 Introduction Suggestion mining is an emerging task in a natu- ral language processing (NLP) field. Definition of suggestion mining task differs in NLP’s commu- nity. Close areas of study like opinion mining or sentiment analysis get a lot of attention not only from academic, but also industrial researchers. From a linguistic point of view, while these areas treat neutral polarity of a statement as an absence of opinion (Liu, 2009), suggestion does not have to be connected with positive or negative emotion and can be treated as complementary information (Negi and Buitelaar, 2015). Lack of sensitivity to statement’s sentiment and various suggestions’ realization strategies (Mart´ınez Flor, 2005) make suggestion mining task interesting and challeng- ing from a NLP’s standpoint. In this work, we present a system for cross- domain suggestion mining, ranked in the 7th place in SemEval-2019 Task 9 Subtask B. The train- ing data for this task was collected from feedback posts on Universal Windows Platform. On the other hand, the test dataset comes from the dif- ferent domain of hotel reviews from the TripAd- visor website (Negi et al., 2019). In this work we will refer to those datasets’ domain as source do- main and target domain accordingly. For sugges- tion mining task in this context, we employ en- semble of Domain-Adversarial Neural Networks (DANN) where we use Structured Self-Attentive Sentence Embedding (Lin et al., 2017) as a fea- ture extractor. Moreover, to achieve better adap- tation towards target domain, we follow the ap- proach of Gui et al. (2017) for the part-of-speech tagging and extend DANN with a Target Preserv- ing component in a form of words decoder for tar- get domain sentences. We train all of the parts of the described system using modified domain ad- versarial training procedure than the one proposed in (Ganin et al., 2016). 2 Data preparation 2.1 Dataset augmentation Training dataset for Subtask B was built using only sentences from source domain. In order to train DANN we take advantage of an additional set of unlabeled sentences from the same domain as a test dataset. We use …

Content truncated to 500 words