ganeshjawahar / mem_absa

Aspect Based Sentiment Analysis using End-to-End Memory Networks
MIT License
199 stars 104 forks source link

Task Definition #6

Open jurukode opened 7 years ago

jurukode commented 7 years ago

Hi @ganeshjawahar,

I would like to ask about task definition for this program. Is it want to infer the polarity of aspect category or aspect term?

Thank you.

joeybose commented 7 years ago

It's the polarity of the aspect term as outlined in this paper https://arxiv.org/pdf/1605.08900.pdf

jurukode commented 7 years ago

Hi @0220joey,

So, the aspect term should be annotate first as input alongside with the review text? Or the model can extract it automatically?

Thank you

joeybose commented 7 years ago

The aspect term with its polarity is provided as part of the training data. When you're testing you select the aspect term along with the input and model predicts the polarity.

jurukode commented 7 years ago

Hi @0220joey,

i see, so everytime we would like to predict unseen review data, we should know the list of aspect term first, right? I thought the model can automatically extract the aspect terms.

joeybose commented 7 years ago

that makes no sense. Why would a model be expected to extract aspect terms? The question you're trying to answer with this model is given a sentence, you pick a word in the sentence and you would like to know that words polarity in the context of the sentence.

jurukode commented 7 years ago

Because it is aspect-based sentiment analysis. So, instead of predicting overall sentiment for given text input. The model should be able to extract valid aspects and define its polarity.

Back to the original model, if we want to use this model in the real world, we should annotate each aspect word that we want to know which makes no sense, right?

joeybose commented 7 years ago

I think you have a fundamental misunderstanding of the actual problem, reading the paper will help but to answer your question. Aspect level sentiment analysis refers to the polarity of an aspect in the context of the sentence it appears in. For example, the sentence "Great food but the service was bad", the polarity of the word food is positive but that of service is negative. In this case it doesnt make too much sense to ask the polarity of the overall sentence, instead we ask the polarity of a chosen word in the sentence. A simple extension for determining the polarity of all words in a sentence would be to sample from the sentence a word at a time and use that as an aspect word for the sentence which might help detail which words are good "aspect words". Also what is the real world situation you are envisioning, as the interesting question is to know the polarity of a word in a sentence.

jurukode commented 7 years ago

Hi @0220joey,

Thanks for detail explanation. So, if i can infer from your statement, the model explained on the paper cannot automatically extract predefined category of aspect (e.g.: food, service) and its polarity which need to be assigned for different task?