Multi-task Pairwise Neural Ranking for Hashtag Segmentation

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 2538–2549 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 2538 Multi-task Pairwise Neural Ranking for Hashtag Segmentation Mounica Maddela1, Wei Xu1, Daniel Preot¸iuc-Pietro2 1 Department of Computer Science and Engineering, The Ohio State University 2 Bloomberg LP {maddela.4, xu.1265}@osu.edu dpreotiucpie@bloomberg.net Abstract Hashtags are often employed on social me- dia and beyond to add metadata to a tex- tual utterance with the goal of increasing dis- coverability, aiding search, or providing addi- tional semantics. However, the semantic con- tent of hashtags is not straightforward to infer as these represent ad-hoc conventions which frequently include multiple words joined to- gether and can include abbreviations and un- orthodox spellings. We build a dataset of 12,594 hashtags split into individual segments and propose a set of approaches for hash- tag segmentation by framing it as a pairwise ranking problem between candidate segmen- tations.1 Our novel neural approaches demon- strate 24.6% error reduction in hashtag seg- mentation accuracy compared to the current state-of-the-art method. Finally, we demon- strate that a deeper understanding of hash- tag semantics obtained through segmentation is useful for downstream applications such as sentiment analysis, for which we achieved a 2.6% increase in average recall on the Se- mEval 2017 sentiment analysis dataset. 1 Introduction A hashtag is a keyphrase represented as a sequence of alphanumeric characters plus underscore, pre- ceded by the # symbol. Hashtags play a cen- tral role in online communication by providing a tool to categorize the millions of posts generated daily on Twitter, Instagram, etc. They are useful in search, tracking content about a certain topic (Berardi et al., 2011; Ozdikis et al., 2012), or dis- covering emerging trends (Sampson et al., 2016). Hashtags often carry very important informa- tion, such as emotion (Abdul-Mageed and Ungar, 1Our toolkit along with the code and data are pub- licly available at https://github.com/mounicam/ hashtag_master Type Single-token Multi-token Named-entity (33.0%) #lionhead #toyotaprius Events (14.8%) #oscars #ipv6summit Standard (43.6%) #snowfall #epicfall Non-standard (11.2%) #sayin #iloveu4eva Table 1: Examples of single- (47.1%) and multi-word hashtags (52.9%) and their categorizations based on a sample of our data. 2017), sentiment (Mohammad et al., 2013), sar- casm (Bamman and Smith, 2015), and named en- tities (Finin et al., 2010; Ritter et al., 2011). How- ever, inferring the semantics of hashtags is non- trivial since many hashtags contain multiple to- kens joined together, which frequently leads to multiple potential interpretations (e.g., lion head vs. lionhead). Table 1 shows several exam- ples of single- and multi-token hashtags. While most hashtags represent a mix of standard to- kens, named entities and event names are preva- lent and pose challenges to both human and auto- matic comprehension, as these are more likely to be rare tokens. Hashtags also tend to be shorter to allow fast typing, to attract attention or to sat- isfy length limitations imposed by some social me- dia platforms. Thus, they tend to contain a large number of abbreviations …

Content truncated to 500 words