With the rapid growth of user-generated content on the internet, automatic sentiment analysis of textual reviews has become a hot research topic recently among researchers of data mining and natural language processing. Sentiment analysis or opinion mining is the computational study of people’s opinions, appraisals, attitudes, and emotions toward entities, individuals, issues, events, topics and their attributes. One of the main problems in sentiment analysis is aspect-based sentiment analysis which its aim is to detect aspects and the sentiments expressed on the aspects.As the number of product and service reviews expands, it is essential to develop an efficient sentiment analysis model that is capable of extracting aspects and determining the sentiments for these aspects. Up to now, different aspect-based sentiment analysis approaches have been developed in the field of relation and frequency-based and model-based approaches. Most of presented methods are domain dependent and produce many unrelated aspects. But, due to variety and wide range of products and services being reviewed on the internet, the supervised and domain-specific models are often not practical.Additionally, with the lack of labeled training dataset, building sufficient labeled data is often expensive, time consuming and needs much human labor. Hence, a model trained on labeled data in one domain often performs poorly in another domain. Other main challenges in aspect-based sentiment analysis are jointly detecting aspects and sentiments, discovering implicit aspects, detecting multi-word aspects and language-dependency of models. This thesis follows aiming to improve previous works to overcome these challenges to present unsupervised and weakly supervised models in that they require no labeled training data for the aspect-based sentiment analysis. To this matter, in this thesis we proposed two domain-independent weakly supervised models for the aspect-based sentiment analysis on online customer reviews. Unsupervised aspect-sentiment detection model is the first proposed model which requires no training data and detectsboth explicit and implicit aspects. Joint aspect-sentiment detection model, JASE, is the second proposed model which is presented based on topic modeling and tries to detect sentiment and aspect words simultaneously. To mine sentiments, the JASE model uses advantages of both frequency-based approaches and semantic features of text. In experimentsection, the proposed models are evaluated based on standard measures of information retrieval on English and Persian review datasets in document and sentence level. In the experiments, the proposed models have been compared with other approaches such as frequency and relation-based and supervised approaches. Experimental results show considerable improvements of the proposed models over conventional models including unsupervised and supervised approaches. By comparing the results, the best values for precision and recall measure in the proposed unsupervised aspect-sentiment model are %90 and %71 respectively. The best performance of the proposed JASE model in accuracy measure for English reviews is %86.79 and for Persian reviews is %79.58. Keywords: Sentiment analysis, natural language processing, opinion mining, aspect-based sentiment analysis, data mining, weakly supervised model.