Sentiment analysis is used to extract opinions expressed in product reviews. Aspect level sentiment analysis extracts opinions about features of a product. However, such analysis cannot infer the underlying reason for the opinions expressed. The following useful applications can be realized, if reasons for opinions are inferred: (i) To perform root cause analysis of negative opinions. (ii) Sentiment inference of causes helps in building a phrase-level sentiment lexicon, and (iii) Creation of causality-aware word embeddings to enhance the accuracy of sentiment analyzers. To realize above-cited use-cases, we have proposed to use a deep neural sequence model to extract cause phrases and designed a novel hybrid model based on graph neural networks and Bayesian reasoning for inferring the sentiments implied by cause phrases. We tested our models on three annotated datasets and observed mean accuracies of 96.34%, 96.12%, 97.14% and 82.14%, 85.23%, 87.21% for the cause phrase extraction and sentiment inference tasks respectively. We have also investigated the impact ofover smoothing in graph neural network through an ablation study and reported the results.