Event Extraction Using Word Clustering and Word Embedding for Roman Urdu

: Urdu is the national language of Pakistan, and Roman Urdu is a writing system representing the Urdu language using the Latin alphabet. The extraction of events from conversations conducted in Roman Urdu is a challenging task. This study aims to recognize multi-word events from text, specifically conversations among users of social networks. The expressed work consists of two phases: data preparation and event detection. In the first phase, four different approaches were investigated to remove useless and stop words in Roman Urdu, providing a comparison of the techniques. In the second phase, two different approaches were used to detect multi-word events in the cleaned text. Among the two, one is based on word clustering and the other is based on word embedding with the help of BiLM (a bidirectional language model). The clustering model efficiently finds events from conversations without complex feature engineering or data training. The second approach uses BiLSTM to obtain word vectors combined with NER (BIO) tagging to sequence the tagged dataset. Datasets for both approaches were prepared separately from the tweets. A comparison shows that word clustering can accommodate new data more efficiently, whereas a deep learning-based approach requires a huge amount of training data. The proposed approach detects the best sequence of two-or three-word events and single-word events in Roman Urdu. Since text-based events can play a significant role in finding the polarity or intensity of a conversation, two well-known techniques have been tested and evaluated.