Synthetic Data for English Lexical Normalization: How Close Can We Get to Manually Annotated Data?

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 6300–6309 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 6300 Synthetic Data for English Lexical Normalization: How Close Can We Get to Manually Annotated Data? Kelly Dekker, Rob van der Goot University of Groningen, IT University of Copenhagen k.dekker.5@student.rug.nl, robv@itu.dk Abstract Social media is a valuable data resource for various natural language processing (NLP) tasks. However, standard NLP tools were often designed with standard texts in mind, and their performance decreases heavily when applied to social media data. One solution to this problem is to adapt the input text to a more standard form, a task also referred to as normalization. Automatic approaches to normalization have shown that they can be used to improve performance on a variety of NLP tasks. However, all of these systems are supervised, thereby being heavily dependent on the availability of training data for the correct language and domain. In this work, we attempt to overcome this dependence by automatically generating training data for lexical normalization. Starting with raw tweets, we attempt two directions, to insert non-standardness (noise) and to automatically normalize in an unsupervised setting. Our best results are achieved by automatically inserting noise. We evaluate our approaches by using an existing lexical normalization system; our best scores are achieved by custom error generation system, which makes use of some manually created datasets. With this system, we score 94.29 accuracy on the test data, compared to 95.22 when it is trained on human-annotated data. Our best system which does not depend on any type of annotation is based on word embeddings and scores 92.04 accuracy. Finally, we perform an experiment in which we asked humans to predict whether a sentence was written by a human or generated by our best model. This experiment showed that in most cases it is hard for a human to detect automatically generated sentences. Keywords: Social Media Processing, Statistical and Machine Learning Methods, Tools, Systems, Applications 1. Introduction With the emergence of social media, a novel type of lan- guage evolved. This ‘online’ language contains many new phenomena, like slang, acronyms and shortened words. However, most natural language processing tools are de- signed to process standard texts. Due to these differences, performance of NLP tools often drop dramatically when used on social media data (Liu et al., 2011; Schulz et al., 2016) Multiple solutions have been proposed to tackle this prob- lem of domain adaptation. These solutions can be broadly divided in two categories: adapting the model to the new data, or adapting the data to the model (Eisenstein, 2013). In this work, we focus on lexical normalization, a solution which adapts the data to the model by ‘translating’ words to their standard form. This has the main advantage that the adaptation only has to be done once for multiple types of NLP applications. An example of a sentence annotated with its normalization is shown in Figure 1. However, one downside …

Content truncated to 500 words