Font size
S
M
L
XL
WorksheetsWeb Mining 6.2
Total questions: 102
Worksheet time: 51mins
Name
Class
Date
1.
In ABSA, what is the goal of aspect extraction?
a)
Identify the aspect mentioned in a text (for example, voice quality)
b)
Assign a sentiment label to a whole document
c)
Detect spam and advertisements
d)
Translate a sentence into another language
e)
Summarize multiple reviews into one sentence
2.
In the slides, what approach is described as using dependency syntax to extract syntactic features?
a)
Supervised approach
b)
Classical rule-only approach
c)
Unsupervised clustering approach
d)
Reinforcement learning approach
e)
Graph neural network approach
3.
Which item is listed in the agenda of this lecture?
a)
Sentiment Ontology Tree
b)
Image segmentation
c)
PageRank computation
d)
TCP congestion control
e)
Compiler optimization
4.
According to the slides, what does a Sentiment Ontology Tree (SOT) demonstrate within a domain?
a)
Ancestor-descendant relationships among aspects
b)
Exact numeric ratings for each review
c)
The layout of a web page
d)
The shortest path between two URLs
e)
A list of stopwords for a language
5.
In the example SOT, what kind of relationship is organized hierarchically?
a)
Aspects of a product and their sub-aspects
b)
Random words in alphabetical order
c)
User accounts and their passwords
d)
Web pages grouped by file size
e)
IP addresses grouped by country
6.
In the SOT definition, what does the root node represent?
a)
A property or aspect being modeled
b)
A training instance index
c)
A stopword list
d)
A web domain name
e)
A sentiment classifier margin
7.
In HL-SOT, what values can each entry of the label vector take?
a)
0 or 1
b)
-1, 0, or 1
c)
Any real number
d)
Any integer from 1 to N
e)
Only the value 1
8.
In the hierarchical labeling constraint for HL-SOT, what must be true if a node is labeled positive (1)?
a)
All of its ancestor nodes must also be labeled 1
b)
All of its descendant nodes must be labeled 1
c)
All other nodes must be labeled 0
d)
The root node must be labeled 0
e)
The classifier thresholds must all be negative
9.
In HL-SOT, what role does the threshold for a node's classifier play?
a)
It is the cutoff used to decide whether the node is labeled 1
b)
It forces all nodes to be labeled 0
c)
It encodes the stopword list
d)
It selects the target object in a tweet
e)
It measures the number of retweets
10.
In the learning setup, how are the weight matrix and threshold vector initialized?
a)
Both are initialized to 0
b)
Weights are random and thresholds are 1
c)
Weights are 1 and thresholds are random
d)
Weights are learned without thresholds
e)
Thresholds are fixed and weights are unused
11.
When does the algorithm update the weight of a node during learning (as stated in the slides)?
a)
Only when the parent node is set positive
b)
Only when the node is a leaf
c)
Only when the tweet is a retweet
d)
Only when the sentence is objective
e)
Only when there are no negative words
12.
In threshold updating, what does the small value epsilon control?
a)
Update speed
b)
Number of classes
c)
Size of the corpus
d)
Number of stopwords removed
e)
Maximum tweet length
13.
In the learning algorithm description, what is explicitly initialized to the zero vector before training begins?
a)
The weight vectors and the threshold vector
b)
The retweet graph and reply graph
c)
The stopword list and stemming dictionary
d)
The set of target objects and hashtags
e)
The maximum entropy constraints and priors
14.
According to the slides, what is the maximum length of a tweet (as stated for Twitter)?
a)
140 characters
b)
280 characters
c)
500 characters
d)
1000 characters
e)
No character limit
15.
Compared with product reviews, tweets are described as being:
a)
Shorter and more ambiguous
b)
Longer and more formal
c)
Always about a known target object
d)
Written only in one language
e)
Always accompanied by star ratings
16.
In the Twitter sentiment problem definition, what are the sentiment categories used for output?
a)
Neutral, positive, negative
b)
Objective, subjective, sarcastic
c)
Angry, happy, sad
d)
Spam, ham, unknown
e)
Relevant, irrelevant, duplicate
17.
In the Twitter sentiment algorithm, what happens if a tweet is classified as objective?
a)
It is labeled neutral
b)
It is labeled positive
c)
It is labeled negative
d)
It is removed as an outlier
e)
It is treated as a retweet
18.
Which tool is used for part-of-speech tagging in the preprocessing step?
a)
OpenNLP
b)
Weka
c)
PageRank
d)
PorterStemmer
e)
MaxEnt
19.
Which is listed as a content feature for sentiment classification of tweets?
a)
Hashtags
b)
TF-IDF topic vectors only
c)
Web page titles
d)
IP addresses
e)
HTML tag depth
20.
Which type of phrase is listed as an 'extra object' candidate in tweets?
a)
Noun phrases
b)
Regular expressions
c)
URL hostnames
d)
CSS selectors
e)
Binary trees of web pages
21.
In the slides, PMI is used to relate extra objects to the original object using:
a)
Top K nouns and noun phrases
b)
Only emoticons and punctuation
c)
Only stopwords
d)
Only retweet counts
e)
Only language identifiers
22.
When a noun phrase is selected as related (by PMI), what can be chosen as its representative term?
a)
The central word of the noun phrase
b)
The first stopword in the phrase
c)
The longest punctuation sequence
d)
The URL embedded in the tweet
e)
The user's screen name
23.
In object-dependent features, what does a feature like 'love_arg2' indicate?
a)
A transitive verb takes the target object as its object
b)
A transitive verb takes the target object as its subject
c)
An adjective modifies a hashtag
d)
A stopword is removed by stemming
e)
A reply link connects two tweets
24.
In object-dependent features, what does a copula-based feature describe?
a)
A noun or adjective links to the target using a copula like 'to be'
b)
A retweet preserves original content
c)
A hashtag indicates the target object
d)
A classifier uses only unigrams
e)
A tweet is filtered as spam
25.
How are features handled when they are modified by a negative word, according to the slides?
a)
A 'neg-' prefix is added
b)
They are deleted from the feature set
c)
They are duplicated without change
d)
They are converted into hashtags
e)
They are forced to be neutral
26.
In graph optimization for tweets, retweets are described as:
a)
Keeping the original content
b)
Always reversing the sentiment label
c)
Removing the target object
d)
Replacing all nouns with generic labels
e)
Being ignored as outliers
27.
In the tweet graph description, what does a dashed line represent?
a)
A retweet link
b)
A reply link
c)
Tweets from the same user
d)
A synonym relation
e)
A dependency parse edge
28.
In the model evaluation slide, what is used as the query set for collecting tweets?
a)
A set of target entities (for example, named people or products)
b)
A set of stopwords for each language
c)
A set of HTML tags from web pages
d)
A set of web domains to crawl
e)
A set of emoticons only
29.
The feature evaluation slide compares performance across different feature sets. Which feature set name appears on that slide?
a)
Target-dependent features
b)
Only page-level link features
c)
Only IP geolocation features
d)
Only HTML layout features
e)
Only audio features
30.
In the forum sentiment analysis section, what type of data is analyzed?
a)
Product reviews in forums
b)
Network packets
c)
Image captions
d)
Source code repositories
e)
Sensor readings
31.
In vocabulary features, what is a unigram in this context?
a)
A single word or token in a sentence
b)
A pair of consecutive words
c)
A dependency tree path
d)
A topic distribution vector
e)
A set of retweet edges
32.
In the syntax features slide, what is one way to compute distance between word/entity features in a dependency tree (as stated)?
a)
Distance measured by BFS in the dependency tree
b)
Distance measured by tweet length
c)
Distance measured by hashtag count
d)
Distance measured by number of web links
e)
Distance measured by page rank score
33.
The example of syntax features includes which kind of relationship between a word and an entity?
a)
Depth and path distance in a dependency tree
b)
Pixel distance in an image
c)
Time difference between posts
d)
Geographic distance between users
e)
File size difference between web pages
34.
Which is listed as a language-specific phenomenon that affects sentiment analysis?
a)
Compound nouns
b)
HTTP redirects
c)
GPU memory limits
d)
SQL joins
e)
TLS handshakes
35.
Which classifier is described as maximizing the margin between two classes?
a)
Support Vector Machines (SVMs)
b)
Multinomial Naive Bayes (MNB)
c)
Maximum Entropy (ME)
d)
K-means clustering
e)
PageRank
36.
In the cascade model diagram, how many levels are shown?
a)
Three levels
b)
One level
c)
Two levels
d)
Four levels
e)
Five levels
37.
What is the stated target of active learning?
a)
Learn a good quality model with minimal labeled data
b)
Eliminate all neutral examples from the dataset
c)
Guarantee perfect accuracy without training
d)
Avoid using any features beyond unigrams
e)
Replace all entities with random labels
38.
In uncertain sampling, how is uncertainty measured for SVMs according to the slides?
a)
Distance to the hyperplane
b)
Number of hashtags
c)
Number of stopwords
d)
Sentence length in characters
e)
Number of retweets
39.
Which is listed as a possible source of datasets for sentiment analysis?
a)
Blog sites
b)
Only printed newspapers
c)
Only television transcripts
d)
Only private emails
e)
Only academic textbooks
40.
In the dataset preprocessing described, what is done with target object names like specific car or movie titles?
a)
They are replaced with a generic label such as CAR or MOVIE
b)
They are translated into French
c)
They are deleted and ignored
d)
They are converted into emoticons
e)
They are used as stopwords
41.
In experiment settings, which feature set is stated as the first level of the cascade model?
a)
Unigram features
b)
Parsed features only
c)
Hashtag features only
d)
Retweet graph features only
e)
Topic model features only
42.
One experimental results slide highlights the 'Impact of features'. Which other impact is named alongside it on the same slide?
a)
Impact of bagging
b)
Impact of HTML parsing
c)
Impact of web crawling depth
d)
Impact of encryption
e)
Impact of image resolution
43.
Another experimental results slide is titled 'Impact of cascade model'. What is being studied there according to the title?
a)
Effect of using a cascade model
b)
Effect of changing tweet length limits
c)
Effect of removing all neutral labels
d)
Effect of adding more languages beyond three
e)
Effect of replacing all unigrams with bigrams only
44.
Which experimental result topic is explicitly mentioned on the slide titled 'Efficiency on neutral example'?
a)
Efficiency on neutral examples
b)
Efficiency on retweet extraction
c)
Efficiency of HTML parsing
d)
Efficiency of web crawling
e)
Efficiency of spam filtering
45.
Which experimental result topic is explicitly mentioned on the slide titled 'Impact of domain'?
a)
Impact of domain
b)
Impact of font choice
c)
Impact of screen resolution
d)
Impact of keyboard layout
e)
Impact of file format
46.
On an experimental results slide, what is reported as being impacted by syntax features?
a)
Ambiguous examples
b)
Server response time
c)
HTML rendering
d)
Disk I/O throughput
e)
Network routing
47.
Which experimental result topic is explicitly mentioned on the slide titled 'Impact on outliers'?
a)
Impact on outliers
b)
Impact on hyperlinks
c)
Impact on page titles
d)
Impact on CPU usage
e)
Impact on log files
48.
Which section title introduces the discussion of model mistakes and failure cases?
a)
Error analysis
b)
Feature engineering
c)
Web crawling
d)
Index compression
e)
Link prediction
49.
According to the error analysis list, which is one cause of errors?
a)
Ambiguous sentences
b)
Perfectly labeled training data
c)
Fully objective sentences labeled neutral always
d)
No metaphors in language
e)
No need for context or background knowledge
50.
In the extended error analysis list, which is mentioned as a source of difficulty?
a)
Emotions expressed through idioms
b)
Emotions always expressed with explicit sentiment words only
c)
No language-specific differences
d)
All errors come only from spam
e)
All errors come only from stemming
51.
The error analysis examples include sentences showing which challenging phenomenon?
a)
Metaphors and indirect sentiment expression
b)
Only exact star ratings
c)
Only HTML tag errors
d)
Only network protocol errors
e)
Only camera sensor noise
52.
In evaluating active learning, what classifier is stated as being used in the experiment?
a)
Multinomial Naive Bayes (MNB)
b)
Support Vector Machines (SVMs)
c)
K-means clustering
d)
PageRank
e)
Decision trees only
53.
In the active learning evaluation, what happens to the labeled set during sampling (as described)?
a)
Examples are sampled until a fixed number of labeled examples is reached
b)
All examples are labeled at once
c)
Only neutral examples are labeled
d)
Labels are assigned without human input
e)
Only retweets are labeled
54.
The final active learning evaluation slide mentions a comparison between uncertain sampling and which other method?
a)
Stochastic sampling
b)
Crawling-based sampling
c)
Bag-of-URLs sampling
d)
Grammar-rule sampling
e)
Random-walk sampling
55.
A sentence like 'The voice quality of this phone is amazing' illustrates that ABSA can require which two outputs?
a)
The aspect (voice quality) and its sentiment
b)
Only the author identity
c)
Only the tweet graph structure
d)
Only the language ID
e)
Only the review length
56.
Why does the supervised approach described in the slides tend to be difficult to adjust to new fields?
a)
It can rely on learned syntactic features tied to a specific domain and training data
b)
It cannot use any text features
c)
It only works on images, not text
d)
It requires no labeled data
e)
It can only output neutral sentiment
57.
In a Sentiment Ontology Tree, what is the practical implication of organizing aspects as ancestors and descendants?
a)
It supports sentiment reasoning at different aspect granularities within a domain
b)
It guarantees that all tweets have known target objects
c)
It removes the need for preprocessing
d)
It replaces supervised learning with clustering
e)
It prevents any ambiguous sentences
58.
Under the HL-SOT constraint, if a sub-aspect node is labeled 1, what must also be true?
a)
Every ancestor aspect on the path to the root must be labeled 1
b)
All unrelated aspects must be labeled 0
c)
Every descendant must be labeled 0
d)
The tweet must be objective
e)
The sentence must contain a hashtag
59.
In HL-SOT decision logic, what happens if a node's own classifier would label it 1 but one of its ancestors is not labeled 1?
a)
The node is forced to 0 to respect the hierarchy
b)
The node remains 1 and ancestors are ignored
c)
The system flips all labels to 1
d)
The node becomes neutral instead of 0
e)
The system skips thresholding
60.
The slides state that a node's weight is updated only when its parent is set positive. What is the most direct reason for this design?
a)
To keep learning consistent with the hierarchical structure of the ontology
b)
To ensure every node becomes positive at least once
c)
To make tweets longer and less ambiguous
d)
To remove the need for thresholds
e)
To avoid using any syntactic information
61.
According to the threshold update rule, if a classifier incorrectly predicts an instance as positive, what adjustment is made and why is it reasonable?
a)
Increase the threshold, making future positive predictions harder
b)
Decrease the threshold, making future positive predictions easier
c)
Set the threshold to zero to reset learning
d)
Remove the node from the tree to avoid errors
e)
Randomize the threshold to encourage exploration
62.
Why do tweets often require an extra step compared with product reviews before sentiment can be assigned to a target?
a)
The target object may not be explicitly known and must be determined
b)
Tweets always contain multiple languages in one message
c)
Tweets cannot contain any nouns or noun phrases
d)
Product reviews have no sentiment words
e)
Tweets always have star ratings
63.
The slides note that a general sentiment classifier is not appropriate for tweet classification. Which reason best matches the slides?
a)
Tweets need target-object handling and additional context beyond general features
b)
Tweets cannot be preprocessed with POS tagging
c)
Tweets cannot be represented as feature vectors
d)
Tweets cannot be labeled as neutral
e)
Tweets cannot be modeled with linear classifiers
64.
Why does the Twitter algorithm classify subjective vs objective before positive vs negative?
a)
Objective tweets are mapped to neutral, reducing confusion in later polarity classification
b)
Positive and negative labels are only used for retweets
c)
Objective tweets must be removed as spam
d)
Subjective tweets are always negative
e)
Neutral tweets cannot contain nouns
65.
What is a key benefit of normalizing variants like 'gooood' to 'good' in preprocessing?
a)
It reduces feature sparsity by mapping similar surface forms to one token
b)
It increases the number of unique tokens intentionally
c)
It removes the need for stemming
d)
It replaces dependency parsing with PMI
e)
It forces all tweets to be neutral
66.
The slides separate content features from vocabulary features. Which pairing is correct?
a)
Hashtags are content features, while General Inquirer words are vocabulary features
b)
Hashtags are vocabulary features, while emoticons are syntax features
c)
General Inquirer words are content features, while words are not
d)
Punctuation is a syntax feature, while dependency paths are vocabulary features
e)
Stopwords are target-dependent features, while nouns are outliers
67.
Top-K nouns and noun phrases are related to the original object using PMI. What does a higher PMI most directly suggest in this context?
a)
The term tends to co-occur with the target object more than expected by chance
b)
The term is always a stopword
c)
The term cannot be a noun phrase
d)
The term is guaranteed to express positive sentiment
e)
The term is a retweet marker
68.
Why does selecting the central word of a related noun phrase help in modeling sentiment targets?
a)
It creates a simpler and more consistent representation of the object mention
b)
It guarantees that dependency parsing is unnecessary
c)
It turns every noun phrase into an emoticon
d)
It removes all negation effects
e)
It forces the object to be a person name
69.
Which statement best explains why object-dependent features are useful for tweet sentiment about a specific target?
a)
They encode how words like verbs and adjectives syntactically relate to the target object
b)
They only count the number of characters in the tweet
c)
They replace all words with numeric IDs without structure
d)
They ignore the target and treat all objects equally
e)
They can only be computed for retweets
70.
What is the purpose of adding a 'neg-' prefix to features when a negative word is present?
a)
To distinguish negated expressions from non-negated ones
b)
To remove negative words from the tweet
c)
To force the label to be positive
d)
To treat all adjectives as neutral
e)
To convert nouns into verbs
71.
Why can using only tweet content lead to incorrect sentiment classification, according to the slides?
a)
Tweets are short, so additional context from related tweets can be necessary
b)
Tweets always contain correct sentiment words
c)
Tweets never mention target objects
d)
Content features cannot include words
e)
Linear SVM cannot handle short texts
72.
The graph optimization assumes a user does not change sentiment about an object in a short time window. What is a consequence of this assumption?
a)
Tweets by the same user can be used to provide consistent context for classification
b)
Retweets must be discarded to avoid bias
c)
Only objective tweets are kept
d)
All neutral tweets become outliers
e)
Hashtags are ignored completely
73.
How does a reply relationship between tweets most plausibly help sentiment analysis in the slide's graph framework?
a)
It can provide conversational context that clarifies sentiment in short messages
b)
It guarantees that the target object is a product
c)
It removes the need to detect the target object
d)
It forces the sentiment to be negative
e)
It eliminates ambiguity by shortening the text
74.
In forum sentiment analysis, why might adding linguistic and syntax-based features help beyond unigrams alone?
a)
They can capture how opinion words relate to entities, including negation and structure
b)
They guarantee perfect performance across all domains
c)
They avoid the need for any labeled data
d)
They only work for neutral sentences
e)
They replace the label set with numeric ratings
75.
Removing stopwords when building unigram features mainly helps because it:
a)
Reduces common, low-information tokens that can add noise
b)
Ensures every remaining token is a sentiment word
c)
Turns unigrams into dependency paths
d)
Automatically detects metaphors
e)
Eliminates the need for cross-validation
76.
Why are syntax-distance features (such as BFS distance in a dependency tree) relevant for sentiment classification?
a)
They help model how close an opinion expression is to the entity being evaluated
b)
They measure the time between tweets
c)
They count the number of retweet edges
d)
They detect whether a sentence is multilingual
e)
They replace the need for any vocabulary features
77.
The slides highlight compound nouns and composed verbs as language-specific issues. What is the main challenge they introduce?
a)
They can change meaning and word boundaries in ways that simple token features may miss
b)
They make tweets exceed the maximum character limit
c)
They prevent using any classifier besides SVM
d)
They guarantee that all sentences are ambiguous
e)
They remove the need for stemming and normalization
78.
In the cascade model, when is it most appropriate to move to a later level that uses parsed features?
a)
When earlier classification is uncertain and the sentence is parsable
b)
Whenever the tweet is a retweet
c)
Only for objective tweets
d)
Only when the target object is unknown
e)
Only when there are no negative words
79.
Active learning is described as automatically selecting examples to label. What is the intended benefit?
a)
Achieve good model quality while labeling fewer examples
b)
Increase the dataset size by duplicating examples
c)
Remove all outliers without inspection
d)
Avoid evaluating on a validation set
e)
Replace all features with random noise
80.
Uncertain sampling selects examples where the classifier is uncertain. According to the slides, what is a stated target of this strategy?
a)
Reduce redundancy in training data and focus on ambiguous examples
b)
Guarantee that all examples are neutral
c)
Eliminate the need for a classifier
d)
Only select examples with many hashtags
e)
Only select examples with long noun phrases
81.
Why does replacing specific object names with a generic label (like CAR or MOVIE) help in the dataset setup described?
a)
It standardizes mentions so models focus on sentiment patterns instead of rare names
b)
It makes dependency parsing impossible
c)
It removes all noun phrases from text
d)
It forces classifiers to output only positive labels
e)
It ensures every example becomes an outlier
82.
According to the error analysis, why can lack of training data cause non-emotional sentences to be mislabeled as positive or negative?
a)
They may contain words that frequently appear in emotional sentences, biasing the classifier
b)
They always contain explicit negation
c)
They are always written in Dutch
d)
They are always retweets
e)
They contain no words at all
83.
In the experiment settings, the cascade model adds more feature types at higher levels. What is added at Level 2 compared with Level 1?
a)
Discourse and negation features
b)
Parsed features only
c)
Only hashtags and emoticons
d)
Only retweet-graph features
e)
Only language ID features
84.
Consider a hierarchy where node B is a descendant of node A. Under the HL-SOT ancestor constraint, which labeling violates the rule?
a)
A=1 and B=1
b)
A=0 and B=1
c)
A=0 and B=0
d)
A=1 and B=0
e)
A=1 and B is unknown
85.
The slides state that if a classifier predicts incorrectly as negative, the threshold is decreased. What is the most direct effect of decreasing the threshold on future decisions?
a)
It makes it easier for the node to be labeled 1 next time
b)
It makes it harder for the node to be labeled 1 next time
c)
It forces the node to always be 0
d)
It removes the need for ancestor checks
e)
It converts the task to regression
86.
Tweets often need target detection before sentiment assignment. What is the most likely failure if the target object is not correctly determined?
a)
Sentiment can be assigned to the wrong object even if polarity detection is correct
b)
The tweet becomes longer and less ambiguous
c)
The tweet is guaranteed to be labeled neutral
d)
POS tagging cannot be run at all
e)
Negation words stop working
87.
Using the object-dependent feature definitions, which feature best matches the sentence 'I love iPhone' when the target object is iPhone?
a)
love_arg2
b)
love_arg1
c)
love_it_arg2
d)
neg-love_arg2
e)
love_cp_arg1
88.
Using the object-dependent feature definitions, which feature best matches the sentence 'iPhone works better with the CellBand' when the target object is iPhone?
a)
arg1_v_better
b)
better_arg2
c)
better_cp_arg1
d)
neg-arg1_v_better
e)
better_it_arg2
89.
According to the negation handling rule, which word would trigger adding the 'neg-' prefix if it modifies a feature?
a)
seldom
b)
often
c)
always
d)
very
e)
rather
90.
Graph optimization assumes a user does not change sentiment about an object in a short time window. If this assumption is violated, what risk is most consistent with the slide's motivation for graph context?
a)
Context from the same user can propagate an outdated sentiment and mislead classification
b)
Retweets will no longer keep original content
c)
Tweets will stop being short
d)
Stopword removal will fail completely
e)
All objective tweets will become positive
91.
In the cascade model, a later level is used only when classification certainty is below a threshold and the sentence is parsable. If certainty is low but the sentence is not parsable, what happens?
a)
The system cannot use parsed features and must output the earlier level's prediction
b)
The system always outputs positive
c)
The system always outputs neutral without any prediction
d)
The system skips directly to active learning
e)
The system retrains the classifier immediately
92.
For uncertain sampling with an SVM, which example is most uncertain based on the slide's criterion?
a)
An example closest to the decision hyperplane
b)
An example farthest from the decision hyperplane
c)
An example with the most hashtags
d)
An example with the longest tweet length
e)
An example that is a retweet
93.
For uncertain sampling with a probabilistic classifier, which probability pattern best indicates high uncertainty?
a)
Predicted class probabilities that are close to each other
b)
One class probability near 1.0 and others near 0.0
c)
All probabilities exactly 0.0
d)
Probabilities that sum to more than 1.0
e)
Probabilities assigned only to neutral
94.
The datasets slide lists outliers such as advertisements and spam. Which impact is the most reasonable implication for a sentiment classifier trained without handling these outliers?
a)
It may learn patterns unrelated to sentiment and degrade classification quality
b)
It will automatically switch to a different language
c)
It will stop producing neutral labels
d)
It will only improve accuracy on every domain
e)
It will remove the need for cross-validation
95.
The error analysis lists metaphors, idioms, and homonyms as challenges. Which improvement is least likely to solve these issues by itself?
a)
Only expanding a sentiment word list without modeling context
b)
Adding context from broader text and related posts
c)
Incorporating domain knowledge for narrow fields
d)
Handling language-specific constructions
e)
Accounting for ambiguity in expressions
96.
In the experiment settings, Level 3 of the cascade model adds parsed features. What is the main reason this can help compared with using only unigrams, discourse, and negation?
a)
It can capture structural relations between entities and opinion expressions
b)
It guarantees that every tweet has a known target object
c)
It removes the need to detect negation
d)
It turns the task into unsupervised clustering
e)
It prevents any outliers from appearing
97.
The slides describe both 'content features' and 'object-dependent features'. If a tweet mentions multiple entities, why can object-dependent features reduce confusion compared with content features alone?
a)
They tie sentiment-bearing words to a specific target via syntactic relations
b)
They remove all nouns and keep only punctuation
c)
They force every entity to share the same sentiment
d)
They ignore verbs and adjectives entirely
e)
They require no preprocessing steps
98.
From the error analysis, which case most directly describes sentiment toward a different entity than the target?
a)
A sentence contains feelings toward another entity
b)
A sentence contains compound words in Dutch
c)
A sentence contains accents in French
d)
A sentence uses a stemming algorithm
e)
A sentence is classified as objective
99.
The slides note that tweets are short and ambiguous and that relevant tweets provide context. Which combination of graph links best supports adding context without changing the original tweet text?
a)
Retweet links, because they keep the original content
b)
Reply links, because they always repeat the original text
c)
Same-user links, because they duplicate tokens exactly
d)
Hashtag links, because they provide parsing
e)
Stopword links, because they remove ambiguity
100.
In HL-SOT learning, if a parent node rarely becomes positive, what is a likely consequence for learning weights of its child nodes given the stated update condition?
a)
Child node weights may be updated rarely or not at all
b)
Child node weights will be updated more often than the parent
c)
Child node thresholds will become fixed at 1
d)
Child node labels will always be 1
e)
The hierarchy constraint will be removed automatically
101.
Active learning aims to reduce labeling effort, while uncertain sampling targets ambiguous examples. Which trade-off is most consistent with these goals when building a training set?
a)
Select fewer but more informative examples rather than labeling many redundant ones
b)
Always label only neutral examples first
c)
Always label the longest sentences first
d)
Always label examples with the highest certainty first
e)
Avoid any validation set to speed up training
102.
In the tweet-sentiment pipeline, why can 'objective' classification before polarity classification reduce errors on neutral data?
a)
Because objective tweets can be mapped to neutral before deciding positive vs negative
b)
Because objective tweets always contain negative words
c)
Because polarity classification requires no features
d)
Because only retweets can be objective
e)
Because neutrality can only be detected after parsing
Reset
