Try Wonders for freeHow the literature is organised
The field has a clear starting point. Pang, Lee and Vaithyanathan (2002), "Thumbs up?", applied Naive Bayes, maximum entropy and support vector machines to film reviews: the classifiers beat human-produced baselines, but did not do as well on sentiment as on topic classification, and the paper closes by examining what makes sentiment harder. Pang and Lee's monograph (2008) is the survey that set out the applications, challenges, categorisation, extraction and summarisation — and, unusually for the field, the broader questions of privacy, vulnerability to manipulation and economic impact. Liu's book (2012) is the other standard reference for definitions.
Two methodological families run through everything. Lexicon-based methods score text using dictionaries of words annotated with polarity and strength, handling negation and intensification; Taboada et al. (2011) is the canonical account, with SO-CAL performing consistently across domains and on unseen data. Machine-learning methods learn from labelled examples, and since the mid-2010s that has meant deep learning (Zhang, Wang and Liu, 2018). Medhat, Hassan and Korashy (2014), Ravi and Ravi (2015), Feldman (2013) and Cambria et al. (2013) are the general surveys of techniques and applications that students cite for background.
Main debates
Accuracy on a benchmark is easier than validity in use, and two of the listed papers say so with numbers. Van Atteveldt et al. (2021) compared manual annotation, crowd coding, numerous dictionaries and machine learning on a validation set of Dutch economic headlines: trained human or crowd coding performed best, none of the dictionaries came close to acceptable validity, and machine learning — deep learning especially — beat the dictionaries substantially but still fell short of human performance. Hartmann et al. (2023), across 272 datasets and 12 million labelled documents, found transfer learning models best but performing worse than leaderboard benchmarks suggest, and showed the result depends on how many sentiment classes you want and how long the texts are. Their shared conclusion is the one to act on: validate the method on your own data before using it, and report that validation.
Within computer science the debate is about granularity: a single polarity score per document hides that people like one aspect of a product and dislike another. That is the motivation Zhang et al. (2023) give for aspect-based sentiment analysis, which separates the aspect term, aspect category, opinion term and sentiment polarity.
Where recent work is heading
Aspect-based sentiment analysis and multimodal analysis of video are the most active areas among the recent reviews. Zhang et al. (2023) propose a taxonomy organised by which sentiment elements a task concerns, summarise how pre-trained language models changed performance, and discuss cross-domain and cross-lingual systems; Gandhi et al. (2023) and Das and Singh (2023) cover the multimodal side, and Liang et al. (2022) and Tan et al. (2022) are examples of the model work underneath.
Applied work in the pandemic years produced benchmark datasets and studies of public discussion: COVIDSenti (Naseem et al., 2021), a large-scale Twitter dataset built for an after-the-fact assessment of early information flows, and Lyu, Han and Luli (2021), who combined topic modelling with sentiment analysis on COVID-19 vaccine discussion to track how topics and sentiments changed over time. Work in languages other than English, such as Elgeldawi et al. (2021) on Arabic, is where students with the right language skills can contribute most readily — and note that both Zhang et al. (2023) and the validity studies treat cross-lingual transfer as unsolved, so the contribution is real.