Affect inTweets: A Transfer Learning Approach

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 1511–1516 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 1511 Affect in Tweets: A Transfer Learning Approach Linrui Zhang, Hsin-Lun Huang, Yang Yu, Dan Moldovan Lymba Corporation 901 Waterfall Way Building 5, Richardson, TX 75080 {lzhang, hhuang, yyu, moldovan}@lymba.com Abstract People convey sentiments and emotions through language. To understand these affectual states is an essential step towards understanding natural language. In this paper, we propose a transfer-learning based approach to inferring the affectual state of a person from their tweets. As opposed to the traditional machine learning models which require considerable effort in designing task specific features, our model can be well adapted to the proposed tasks with a very limited amount of fine-tuning, which significantly reduces the manual effort in feature engineering. We aim to show that by leveraging the pre-learned knowledge, transfer learning models can achieve competitive results in the affectual content analysis of tweets, compared to the traditional models. As shown by the experiments on SemEval-2018 Task 1: Affect in Tweets, our model ranking 2nd, 4th and 6th place in four of its subtasks proves the effectiveness of our idea. Keywords: Natural Language Processing, Deep Learning, Transfer Learning 1. Introduction In recent years, the interest in analyzing Twitter has grown exponentially among the NLP community and a substantial amount of related events and workshops (Rosenthal et al., 2015) (Nakov et al., 2016) (Rosenthal et al., 2017) (Barbi- eri et al., 2018) (Van Hee et al., 2018) have been organized. An essential part towards analyzing Twitter is to detect the emotions and the intensity of these emotions that are con- tained or can be inferred from tweets. For example, from the following two tweets, (1) I’m in tears. This is so heart- breaking, (2) You don’t know how to love me when you’re in sober. we should know that both tweets convey sadness, and the second tweet implies sadness to a lesser extent. SemEval-2018 Task 1: Affect in Tweets (Mohammad et al., 2018) presents an array of subtasks where participat- ing systems need to automatically determine the (intensity of) emotions and (intensity of) sentiments from the corpora provided by the organizers. Meanwhile, the organizers also summarized the methods and resources used by the par- ticipating teams. From their summarization, we observed that most of the participants chose to solve the problems with feature-based machine learning algorithms, which im- plement systems with a tremendous amount of linguistic features, pre-learned vectors and extra corpora (Park et al., 2018) (Baziotis et al., 2018) (Meisheri and Dey, 2018). For instance, the top one performer SeerNet (Duppada et al., 2018) implemented its system with pre-trained DeepMoji (Felbo et al., 2017) vectors, Skip-Thought vectors (Kiros et al., 2015) and Sentiment Neuron vectors (Radford et al., 2017), as well as a substantial amount of linguistic fea- tures, such as AFINN (Nielsen, 2011), NRC Affect Inten- sities (Mohammad, 2018) and Emotion Lexicon (Moham- mad and Turney, 2010). Such …

Content truncated to 500 words