Volume 18, No. 6, 2021
Weighted Focal Loss Function For Evaluating Effectiveness Of Word Embeddings On Suggestion Mining From Opinion Reviews
Naveen Kumar Laskari , Suresh Kumar Sanampudi
Abstract
Opinion reviews are significant in the purchase and decision-making process. Opinion Mining(OM) approaches are used to detect the sentiment at varying levels of granularity. It is worth noting that suggestions appear in opinion reviews, and mining these suggestions is regarded as suggestion mining. Suggestion mining seems to be a text classification technique that includes a variety of methodologies, ranging from traditional machine learning methods to deep models. For any of these models to work, the input text must be represented as a vector. Word embeddings are the process for encoding the input into a vector representation. This study proposes the effect of several widely used pre-trained word vectors combined with neural network architectures to utilize the weighted focus loss function. With the unique loss function established after examining the dependability of numerous models for the job, FasText and Glove embedding approaches fared substantially better with CNN and multi-layer LSTM architectures. The FastText embedding grabbed the information more effectively attributed to the sub-word information-based method, and the 1-D convolution operation captured the sequential information more effectively. Consequently, the CNN model can learn faster and achieve the best outcome. The models are implemented and tested using the datasets given by the SemEval-2019 organizers.
Pages: 1464-1483
Keywords: Opinion Reviews, Suggestion Mining, Neural Networks, Word Embeddings, and Loss function.