22.2.2023 ch. 5: Logistic regression

Q: when we use logistic regression for sentiment analysis it uses the probalblity of positive and negative words
how about words like ----> terribly good, not bad [YM] [more precisely it uses the conditional probability of a word having a positive or a negative sentiment (WM)]
A: Words like 'terribly' (or very, extremely., etc) are intensifiers. They increase the sentiment of the words in their scopus (e.g. good or bad). Intensifiers can also be negative ones which reduce the sentiment of other words,  like in 'hardly interesting' or 'little benefit'. 
'not' is a case of negation. It reverses a sentiment.
Intensifiers and negation can also appear in combination ('not very expensive', 'almost not useful'). Both  have in common, that it is often difficult to correctly determine their skopus ('This book is not so much a thrilling but an entertaining one'). Moreover, words are ambiguous wrt. their sentiment. E.g. a 'large' can carry a positive sentiment in one aspect (like the capacity of a storage device) and negative in another one (like its physical dimensions).

Q: Activation function is needed for non-linearity of a model, does it mean it changes itself in every input or it is fixed? [YM]
A: No, the kind of non-linearity used is part of the system design. It is the same for every input. 

Q: what the direction of the gradiet tells about? [MW]
A: It points into the direction where the optimal combination of weights can be found.

Q: Why sigmoid is prefered for classification? [DM]
A: Is it really preferred? Maybe, it is because classifiers are very simple networks. For complex architectures, the sigmoid activation function might cause serious efficiency problems.

-- WolfgangMenzel - 22 Feb 2023
 
This site is powered by FoswikiCopyright © by the contributing authors. All material on this collaboration platform is the property of the contributing authors.
Ideas, requests, problems regarding Foswiki? Send feedback