The main goal we adress in this work is text comprehension in social web through Natural Language Processing. The type of content we are interested in includes noisy sentences for the most part and thus standard topic extraction is generally not applicable. We present here a semi-automatic method based on machine learning and human defined linguistic features that performs points of interest extraction in consumer reviews. We first run a classification of subjective adjectives with a high precision and then use this extracted lexicon to identify co-occurring common nouns. While there is still room for improvement on several levels of our process, the results we obtain with this first method are very promising.