The trained model is already provided and can be used to run predictions on a csv file containing suggestions. An example file is available at data/test.csv If you wish to train your own model, you can either take advantage of a GPU and set "cuda_device": 0 param in your jsonnet config file to run on a GPU. It is currently set "cuda_device": -1 and will train on CPU. This task can take very long to run, but can be cut short by reducing the number of epochs, currently set to 50.

Requires Python 3\

allennlp train --include-package=my_library cnn_bert.jsonnet -s ser_dir

The trained model is already provided and can be used to run predictions on a csv file with the folliwing command: allennlp predict ser_dir/model.tar.gz data/test.csv --include-package=my_library --predictor=suggestion_mining

The above command will output confident scores between 0 and 1 for a given text.

The model classifies a given text in the form of a csv file into a suggestion or not a suggestion. The term 'suggestions' refers to the expressions of tips, advice, recommendations, etc. Specifically, we will focus on classifying whether a comment is a suggestion or not.

For hotel companies, one such application could be used to extract room tips from online hotel reviews, which further will benefit future customers. Another example would be collecting suggestions for hotel improvements. In general, this would be greatly helpful to the service industry. A suggestion mining component can extend the capabilities of traditional opinion mining systems, which can then cater to additional applications. Such systems can empower both public and private sectors by extracting the suggestions which are spontaneously expressed on various online platforms, enabling organizations to collect suggestions from much larger and varied sources of opinions than the traditional suggestion box or online feedback forms.

We will use Distilled BERT which smaller and faster than the existing tools. People will care about the difference because in production, smaller model and faster modeling serving will reduce the cost significantly.

We anticipate that BERT and AllenNLP will effective for this task.

SemEval 2019 Task 9 - SubTask A - Suggestion Mining from Online Reviews and Forums https://competitions.codalab.org/competitions/19955#learn_the_details

Suggestion Mining from Online Reviews using ULMFiT https://arxiv.org/pdf/1904.09076.pdf