I have a dataset with a number of houses, for each house, I have a description. For example "The house is luxuriously renovated" or "The house is nicely renovated". My aim is to identify for each house whether it is luxuriously, well or poorly renovated. I am new to NLP so any tips on how to approach this problem would be much appreciated.


It all depends on what kind of annotations or other variables are present in your dataset. I see 2 possible scenarios here:

  • your dataset is made only of houses descriptions, without any indication of their luxury level.

  • you have annotations regarding the luxury level or another similar variables from which you can infer the luxury level (like the house price for example).

In the first case there's not much you can do except for trying to apply some unsupervised algorithms or transfer learning. Usually in NLP unsupervised techniques are used for tasks like Topic or language modeling, both of which are not really helpful for your specific application since they work at a really abstract level trying to learn relationships between words or documents in corpora containing huge variety of texts. The best you could try could be preprocessing the data to extract specific terms like entities (cities names, furnitures names, etc.) and adjectives from each description, and then try cluster them into n clusters were n is the number of classes you're interest in by applying for example Latent Dirichlet Allocation. Even though everything is possible with the right time and patience, I would never follow this road, especially because it relies on a perfect preprocessing, that involves already transfer learning, for the name entity recognition part for example. And even tough libraries like SpaCy offer really good models to perform these tasks, once you have these kind of annotations a rule based approach would probably be faster and easier to build than another unsupervised model, e.g. creating a simple dictionary contacting names of luxurious furnitures name and adjectives that indicate if the house is expensive would probably be sufficiently good.

In the second case, the story change completely because if you have annotations you could rely on supervised learning. If you already have explicit annotations like "luxuriously renovated" "nicely renovated" nothing stops you from trying to train whatever architecture you feel more comfortable with on this classification task. Even simple architecture like CNN that are easy and fast to train usually achieve good results in classification, and you definitely want to leverage some pre-trained embedding vectors like GloVe as an input feature (every deep learning framework like tensorflow or pytorch already implement the possibility to use them).

To conclude, if you don't have annotations you might try LDA just to check if you're lucky but if I were you I would start annotating data as quick as I can. A good practice in this case is to also ask someone else to perform the annotations, not only to be faster in the creation of the dataset but also to then calculate the Inter Annotator Agreement score, that gives an indication about the quality of the annotations (if the score is low, the dataset quality is poor and no model will be able to learn something from it).


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.