In this work, we propose a rule based method to identify the event type and create a frame for that event. With the help of Natural Language Toolkit (NLTK) and preloaded SpaCy models, we have tried to define certain methods to identify the event (for task 1) and create the event frame from the given article (for task 2).
The use of language processing modules such as parser, chunkers, stemmers has been growing steadily in recent years. The event extraction is a significant and necessary aspect of the natural language processing (NLP) & computational linguistics.
The identification[
Event detection is the mechanism by which event types are analysed so as to identify collections of events relating to event patterns in the sense of an event. The patterns of the occurrence and context describe forms of events. If during the study a series of events following the sequence of an event type is detected, then the event type perpetrators should be alerted. Typically, the analysis involves filtering and integrating the events.
Event frame creation[
In order to detect a disaster, whether manmade or natural and their subtypes, we store certain terms and categories from the train datasets that define the type of disaster. Then using the bag-of-words approach, we find the similar words in test datasets. For each similar word in a ifle, we increase counter for the type & subtype that the word is representing. We denote those that has maximum occurrence with the word.
SpaCy is an open source software library available for advanced natural language processing.
It has pre-built neural network models to perform many operations such as parts-of-speech
tagging, named entity recognition tagging, dependency parsing etc. such works on mainly
European languages. We use preloaded SpaCy models to perform named entity recognition[
To find the reason data, we search for all possible substrings that has a common keyword from disaster data, then we continue until any specified stopword occurs. If we get the reason substring that contains only very few number of words i.e. not be understandable by human being, we discard those and move forward until we find any such specific sentence.
For casualty statement, we find such substrings that contains a keyword such as injure or kill or dead along with a cardinal to represent the number of casualties occurred during that disaster. If we can find only some cardinal data, but no such specified keywords can be found within that, we generally discard that substring and looking forward for another substring with similar kind of data.
We ran our program for both event detection (task 1) & event frame creation (task 2) for English
Dataset and only event detection task for Bengali Dataset. The precision, recall & f1_score
for our results are given by EDNIL 2020 organizers[
Despite of our method being rule based, we faced some constraints during the implementation. And as our results are not a state-of-art being such simple technique. Here are some of our restrictions• As we use SpaCy models, it is very dificult for this model to identify arguments in regional Indian languages. • As we use bag-of-words approach, so if there exists any other type of disaster in the given data, our method may not capture that. • For generating the reason argument string, we have to define certain stopwords, otherwise it will be impossible for the program to select those strings.