"I studied at Stanford" yields a positive result, while "I studied at Harvard" it's neutral. So I guess it works fine .
:)))))))
Unfortunately, "I studied at Leland Stanford Junior University" yields a negative result, which suggests that its positive assessment of "I studied at Stanford" is just an artifact of incomplete information.
Seriously? Is this a little subjectively biased?
As opposed to being objectively biased?
It can finally be concluded: Bilbo Baggins' statement, "I don't know half of you half as well as I should like, and I like less than half of you half as well as you deserve," is indeed primarily an insult.
That isn't right... The first part means "I'd like to know most of you better", and the second part means "I dislike most of you much more than you deserve".
For the second part: not most, less than half. So "I don't like some of you nearly as much as you deserve." Either way neither is an insult.
"I like you half as well as you deserve" means "I should like you at least twice as much". So the sentence means "I like less than half of you half as much as you deserve" which means "I should like most of you at least twice as much."
It's a compliment expressed as a negative sentiment. Saying that he doesn't like them as much as he should
Since "none" is less than half, it could also mean that none of you deserve to be liked twice as much as I like you.
I think.
"Seeing your face makes me reconsider living on this planet."
This is classified as mainly positive.
Well, if it's said by someone who is just about to leave the planet, it would indeed be quite positive.
Will that classify as sarcasm?
"I'm eagerly looking forward to Mr. Wiseau's future work." gets a negative rating - either this is a mistake or the algorithm is far more sophisticated than I had imagined.
You said "work".
because of the corpus
I think it's mainly caused by "eagerly". But it confuses me that " 'm" has some negative effects.
How exactly does one build from the sentiment model used in the source code? For example, let's say I have a list of sentences and each one is scored 0,1,2,3,4 corresponding to very negative, negative, neutral, positive, very positive. The list of each sentence would be my datasetSentences.txt. Is it possible to just train a model from a list of sentences and scores for each sentence or must I score each individual phrase within the sentence?
Moreover, It's unclear from both your paper and the read me provided with the sample dataset how you go from a dataset of sentences to phrases. The parser returns trees but I'm not sure how to convert those trees into phrases.
It would be even more interesting to know how I can add to the model in a similar process you have by scoring each node of my own sentence (on your "help the model" page). How could one go about performing that process locally?
and i do not know how it is feasible to store random and different vector for each word!
"The movie was too good" -- It gives me a negative result
Fascinating example. Let me try to defend the model's negative judgment briefly. I would say that the sentence is in fact semantically negative, just as "The movie was too long" is semantically negative. However, it is possible for the speaker to act as if the movie's degree of goodness was unacceptably high, in order to create a specific pragmatic effect. For this to come across clearly, it helps to emphasize 'too' and act exclamatively: 'The movie was TOO good!' To get this right, we really need to know what the speaker values (compare 'She is TOO smart' with 'She is too smart for her own good').
"Too good" should be positive as "not good" acts as a negation like that "too good" should increase the positivity of the phrase..Detecting sarcasm depends on the context of the document and one liners are most difficult ones to identify...
too bad...
Its the word "long" that changes the sentiment; too should only magnify the sentiment of the adjective
Which still does not make the result correct. In the above case, the sentence on its own can never be taken as negative unless it's surrounded by negative sentences itself.
I think in a quest to identify negative and sarcastic sentences, you have biased the model and it considers each sentence as negative until proven otherwise.
Or
It has pre-associated "too" with a negative meaning by default and cannot correct outsell now.
too itself is neutral but "is too good" has a negative sentiment associated with it. I wonder why.
"I am going too" "I am hurting too" "I am celebrating too" -- "too" seems essentially neutral in this construction.
"too" (something) implies excessively, more than is desirable, and seems likely to be negative.
Correct "itself"
May it be that it is considering "too" to mean "overly" and therefore inferring a negative connotation? For example "This food is too hot".
I agree. I understand 'too' as something analogous to 'troppo' (in italian) or 'très' (in french) both meaning 'something that exceeds a desirable state'. But in this case, considering 'good' is infinetely desirable, it seems that we end up interpreting 'too good' as something positive.
Using live-demo for an in-class lab exercises for information design research methods class, Mount Royal University, Calgary AB. Exceedingly useful, richly explanatory, intuitive presentation--an effective teaching aid. Thank you so much for all your hard work.
Hi, Can you give me the link, so that it will be helpful for me
Nice example from ranjeetbhatia from https://news.ycombinator.com/i...
Sentiment analysis has never been this good. Sentiment analysis has never been good.
Both were classified correctly :)
Just released a quick Ruby module to parse & import the sentiment treebank, see "http://blog.someben.com/2013/0...".
Should we expect the same results in the live demo as in the downloaded package?
The following sentence is tagged as "very negative" in the offline version yet "very positive" in the online demo:
"Wishing you a very happy holiday from all of your friends at Ameristar!"
running into this as well. what's different?
I re-trained the model with the downloadable training dataset for 24h (~750 epochs out of the default 1,000) and the sentiment tag for my example sentence changed from "very negative" to "positive", much closer to the online demo's "very positive" tag. Maybe the trained model included in the NLP Core package is outdated?
If I want to re-train the model with some other dataset, does it have to labelled training data?
Jose Tagarro, I have trained the model with the training dataset. However, I am having trouble figuring out how to run sentiment analysis from the command line with the new model. Could you point me in the right direction?
After training the model you end up with your own model.ser.gz. Then you just need to pass it with the -sentimentModel parameter. Adjust the class path and model path as needed. For example:
echo 'Wishing you a very happy holiday!' | java -cp '*' -mx5g edu.stanford.nlp.sentiment.SentimentPipeline -sentimentModel model.ser.gz -stdin
how to get PDB format file from original sentences, for examples tweets, amazon reviews? thanks!
We published a java class in the code that transforms text lines into the PTB format.
Thanks a lot。Which class can transform original sentences to PTB format file?
How to training sentiment model for other language?
The class name is BuildBinarizedDataset
You can ask more detailed questions on the Stanford CoreNLP mailinglist.
Best,
Richard
Jose Tagarro , thanks. I did eventually get to the javadocs here http://nlp.stanford.edu/nlp/ja... and find that as well.
If you don't mind, I'm looking through all the docs and am wondering about barrier to entry for training this on a new model, such as Twitter tweets.
Given that I have stanford-corenlp-3.3.1, a Twitter POS tagger such as http://gate.ac.uk/wiki/twitter... or http://www.ark.cs.cmu.edu/Twee..., how can I generate the train dev and test ptb trees?
What about the model englishRNN.ser.gz? If I decide to use a Twitter tagger, will I lose many of the benefits of this approach? Is it possible to create a englishTwitterRNN.ser.gz?
thanks
Jose Tagarro how big was your trained model ? do you mind sharing it if it isn't huge. thanks
It's 4MB. I don't mind sharing it, please give me an e-mail address.
@ Jose Tagarro Would you share your training set with me too? would be very appreciative!