C9

Question 1

levels of sentiment analysis

Answer 1

document level: sentiments score for complete text (eg. review or Tweet)

sentence level: sentiment score per sentence (review may address multiple aspects)

entity and aspect level: relate the sentiment to features of a product, event or entity

Question 2

sentiment classes

Answer 2

negative, positive, neutral

alternatives: objective vs. subjective, joy/anger/fear/etc., stance (pro/con/neutral)

ordinal scales

Question 3

ordinal regression

Answer 3

learn a model to predict class labels on an ordinal scale

• variant of regression for ordinal variables
• a problem between regression and classification (“ordinal classification”)

P(y ≤ j | theta_j, w, X) = 1 / (1 + e^(-(theta_j - Xw))
y = target variable
theta_j = threshold for class j
X = input instances
w = weights to be learned

Question 4

aspect-based sentiment analysis

Answer 4

find quintuple (E, A, S, H, C)
E = opinion target (entity, event or topic) (given by metadata in reviews, or extract from text)
A = aspect or feature of E (aspect categorization needed (can be challenging), aspects are domain and product dependent)
S = sentiment/opinion content (sentiment score of A)
H = opinion holder (the author or extract from (news) text)
C = context; time and location of the expression (data/location stamp in Tweets or reviews, else extract from text)

Question 5

why does it help to have a product database?

Answer 5

1. To know which products exist (someone might mention a different product in the review)
2. To know which aspects a given product type has (a drill does not have cleanliness as relevant aspect)
3. This facilitates aspect extraction (know what to look for in the text)

Question 6

challenges of sentiment analysis

Answer 6

1. sentiment words do not always express sentiment (“can you tell me which camera is good?”, “If I see a good camera, I will buy it”)
2. Sentiment words are ambiguous, context- and domain dependent.
3. Sarcasm (“great headphones if you enjoy the noises of other people”)
4. objective sentences that express sentiment (“the washing machine uses a lot of water.”)

Question 7

evaluation of sentiment analysis

Answer 7

discrete labels => precision and recall, average F-score only on positive and negative labels

regression => RMSE

Question 8

repeatability

Answer 8

Same team, same experimental setup: can you find your own result again with your own hardware, code, and data?

Question 9

reproducibility

Answer 9

Different team, same experimental setup: same artifact (code, data, experimental set-up) as the original researchers.

Question 10

replicability

Answer 10

Different team, different experimental setup: someone else can find the same results (e.g. “Transformers are better for this problem than SVM!”) with their own code