Sentiment Analysis for Talent Attraction

Sammanfattning: The reputation of a company on employer review platforms can have a significant impact on its ability to attract talented workers. Companies use sentiment analysis to learn how their employer brand is perceived online. Furthermore, sentiment analysis can detect strengths and weaknesses in their employer brand, indicating which areas need improvement. The proposed methods for improving word embeddings for sentiment analysis commonly involve combining several pre-trained word embeddings, or concatenating vector representations of non-textual elements (e.g., emojis and images) to word embeddings. These methods involve training complex neural networks, which is usually computationally expensive. This thesis investigates if adding features prior to tokenization, instead of concatenating embeddings, increases the accuracy of word embeddings, thereby improving the results of the fine-tuned BERT model for classifying sentiment of employer reviews on the online platform Glassdoor. It also investigates the impact of the BERT Next Sentence Prediction objective on the models’ ability to learn more accurate word embeddings. Testing three different models and comparing their performance indicates that the suggested approach can improve the model’s accuracy. However, additional research is needed to investigate the impact of the chosen features on the observed results. The study hasn’t found enough evidence that addition of the Next Sentence Prediction objective results in higher accuracy of the model, but it shows that it significantly improves model ability to understand the sentiment of the reviews.