Semantic Role Labeling for Generating Template of Indonesian News Sentences

Template-based approach is commonly used in automated journalism, including for Indonesian news article generation based on hand-crafted templates. Nowadays, there is still no research in automated template generation. This paper describes a method to generate templates automatically using semantic role labeling (SRL). Since there is no public dataset for Indonesian SRL, we also create the dataset for Indonesian SRL using PropBank annotation (Palmer et al., 2005). For the SRL model, we adapt the model that was introduced by He et al. (2017). Result from the model will be saved and compared to the templates that were manually defined by Indrayani and Khodra (2018). Our 2-layer BiLSTM model achieves 0.92 F1 on token level and 0.84 on sentence level. The automatically generated templates have similar quality to the ones that Indrayani and Khodra (2018) defined.