Biographically Relevant Tweets – a New Dataset, Linguistic Analysis and Classification Experiments

Proceedings of the 29th International Conference on Computational Linguistics, pages 3669–3679 October 12–17, 2022. 3669 Biographically Relevant Tweets – A New Dataset, Linguistic Analysis and Classification Experiments Michael Wiegand Digital Age Research Center (D!ARC) Alpen-Adria-Universität Klagenfurt AT-9020 Klagenfurt, Austria michael.wiegand@aau.at Rebecca Wilm Institute of Computational Linguistics Heidelberg University D-69120 Heidelberg, Germany wilm@cl.uni-heidelberg.de Katja Markert Institute of Computational Linguistics Heidelberg University D-69120 Heidelberg, Germany markert@cl.uni-heidelberg.de Abstract We present a new dataset comprising tweets for the novel task of detecting biographically rele- vant utterances. Biographically relevant utter- ances are all those utterances that reveal some persistent and non-trivial information about the author of a tweet, e.g. habits, (dis)likes, fam- ily status, physical appearance, employment information, health issues etc. Unlike previ- ous research we do not restrict biographical relevance to a small fixed set of pre-defined relations. Next to classification experiments employing state-of-the-art classifiers to estab- lish strong baselines for future work, we carry out a linguistic analysis that compares the pre- dictiveness of various high-level features. We also show that the task is different from estab- lished tasks, such as aspectual classification or sentiment analysis. 1 Introduction Since its beginning, the web has been a valuable data source for natural language processing (NLP) applications. Particularly social media contain data complementary to what previous (news) text cor- pora could offer (Farzindar and Inkpen, 2015), such as information on ordinary people. This work focuses on biographical information on ordinary users of Twitter. The task is to identify tweets that contain biographically relevant infor- mation on the author of a tweet. We frame this as a binary text classification task: We distinguish between biographically relevant tweets (1)-(5) and biographically irrelevant tweets (6)-(10). (1) relevant: I’m a grown man in my mid-30s with no children (2) relevant: Today was the first day at Deckers my manager told me good job lmao. (3) relevant: one of the reasons i regularly wear @Ameri- canOutlaws kit is because it makes a political statement (4) relevant: i’m actually rather shy in real life (5) relevant: Brenda Fricker will always be my favourite Irish actress (6) irrelevant: It’s after 1 a.m. here and I am stupidly awake (7) irrelevant: @USER i wanna buy your mum a beverage. (8) irrelevant: I heard they’re changing a dollar for corn at the Tampa Bay stadium, that’s a buck an ear. (9) irrelevant: I’m glad I stayed home last night but I’m sad I was alone lmfaoooooooo (10) irrelevant: Headache? I’ve had a headache all day xx Unlike previous work (Sekine and Artiles, 2009; Ji et al., 2010; Garcia and Gamallo, 2015), we do not use a pre-specified list of biographical rela- tions. Instead, we consider everything as a bio- graphically relevant relation if it reveals some persistent and non-trivial information on the user authoring a particular tweet. This may in- clude simple information on age and family status (1), (major) life events (2), habits (3), personality traits (4), (dis)likes (5), etc. Tweets may offer a plethora of biographical information about their users and we …

Content truncated to 500 words