Table 5.

Comparison of topic models.

	Advantages	Disadvantages
LDA	• Prior domain knowledge is not necessarily required • Finds coherent topics when correct hyperparameter tuning is applied • Can deal with sparse input • The number of topics is generally smaller than word-embedding based approaches; thus, it is easier to be interpreted • One document can contain several different topics (Mixed membership extraction) • Full generative models with multinominal distribution over topics are generated • Shows both adjectives and nouns within topics	• Detailed assumptions are required • Hyperparameters need to be tuned carefully • Results can easily produce overlapping topics as topics are soft clusters • Objective evaluation metrics are widely missing • The number of topics needs to be defined by the user(s) • Since the results are not deterministic, reliability and validity are not automatically ensured • Assumes that the topics are independent of each other; hence, only the frequency of the common occurrence of words is used • Word correlations are ignored, so no relationships between topics can be modeled
NMF	• Prior domain knowledge is not required • Supports mixed membership models; thus, one document can contain several topics • In contrast to LDA, which uses raw word frequencies, the term-document matrix can be weighted with TF-IDF • It proves to be computationally efficient and very scalable • Easy to implement	• Frequently delivers incoherent topics • The number of topics to be extracted must be defined by the user in advance • Implicit specification of probabilistic generative models
Top2Vec	• Supports hierarchical topic reduction • Allows for multilingual analysis • Automatically finds the number of topics • Creates jointly embedded word, document, and topic vectors • Contains built-in search functions (easy to go from topic to documents, search topics, etc.) • Can work on very large dataset sizes • It uses embeddings, so no preprocessing of the original data is needed	• The embedding approach might result in too many topics, requiring labor-intensive inspection of each topic • Generates many outliers • Not very suitable for small datasets (<1,000) • Each document is assigned to one topic • Objective evaluation metrics are missing
BERTopic	• High versatility and stability across domains • Allows for multilingual analysis • Supports topic modeling variations (guided topic modeling, dynamic topic modeling, or class-based topic modeling) • It uses embeddings, so no preprocessing of the original data is needed • Automatically finds the number of topics • Supports hierarchical topic reduction • Contains built-in search functions (easy to go from topic to documents, search topics, etc.) • Broader support of embedding models than Top2Vec	• The embedding approach might result in too many topics, requiring labor-intensive inspection of each topic • Generates many outliers • No topic distributions are generated within a single document; rather, each document is assigned to a single topic • Objective evaluation metrics are missing