WorksheetsWeb Mining 7
Total questions: 116
Worksheet time: 58mins
Name
Class
Date
1.
In the slides, information extraction is defined as the process of finding what in a text?
a)
Entities and relationships between entities
b)
Only document topics
c)
Only sentiment polarity
d)
Only word frequencies
e)
Only paragraph boundaries
2.
Which statement is an assumption of information extraction mentioned in the slides?
a)
Information is presented explicitly and requires no inference
b)
Every document contains a knowledge graph
c)
All entities are unique in each sentence
d)
All relationships are numeric
e)
Entities never overlap
3.
Which of the following is a type of information that can be extracted according to the slides?
a)
Attributes of an entity
b)
Only document length
c)
Only font styles
d)
Only page layout
e)
Only file metadata
4.
In the high-level information extraction architecture, which step comes after parsing?
a)
Field analysis
b)
Word tokenize
c)
Partitions
d)
Pos-tag
e)
Shallow parsing
5.
Named Entity Recognition (NER) does what according to the slides?
a)
Detects named entities in text and classifies them into predefined classes
b)
Clusters documents by topic
c)
Summarizes documents into one sentence
d)
Translates text between languages
e)
Predicts document popularity
6.
Phrase chunking is described in the slides as detecting what?
a)
Noun and verb phrases in sentences
b)
Only named entities
c)
Only dependency arcs
d)
Only stopwords
e)
Only punctuation tokens
7.
Relation extraction aims to extract what according to the slides?
a)
Relationships between entities (such as attributes or events)
b)
Only character embeddings
c)
Only sentence boundaries
d)
Only part-of-speech tags
e)
Only spelling corrections
8.
Coreference resolution is defined in the slides as detecting what?
a)
Occurrences of the same entity as different references
b)
All synonyms in a dictionary
c)
All possible relations in a knowledge base
d)
All noun phrases in a sentence
e)
All topics in a document
9.
Which limitation of dictionary-based NER is explicitly mentioned in the slides?
a)
It cannot handle ambiguity
b)
It always requires a dependency parser
c)
It cannot tokenize text
d)
It cannot output BIO tags
e)
It requires image features
10.
In machine-learning-based NER, what is the output for a sentence as described in the slides?
a)
A label for each word in the sentence
b)
A single label for the whole document
c)
A ranked list of websites
d)
A set of paragraphs with no labels
e)
A dependency tree only
11.
In the BIO labeling scheme, what does the label 'B' mean?
a)
Begin
b)
Between
c)
Background
d)
Boundary
e)
Backoff
12.
Which feature is included in the NER feature set described in the slides?
a)
Words in a window around the current word
b)
Pixel intensity around a token
c)
Font size of each word
d)
Slide background color
e)
Audio pitch of the speaker
13.
In the CRF-based NER slide, one setting is described as using what type of PoS and chunking information?
a)
Golden PoS and chunking
b)
No word segmentation and no labels
c)
Only character CNN outputs
d)
Only topic clusters
e)
Only dictionary matches
14.
In the evaluation slides, one comparison concerns which type of PoS tags?
a)
Golden PoS versus automatic PoS from tools
b)
Random PoS versus hand-written PoS
c)
Image-based PoS versus text-based PoS
d)
HTML PoS versus PDF PoS
e)
None of these
15.
Another evaluation comparison in the slides concerns which preprocessing step?
a)
Golden word segmentation versus automatic tokenizers
b)
Stopword removal versus stemming
c)
Language detection versus translation
d)
HTML parsing versus CSS parsing
e)
Compression versus encryption
16.
Which model choice in the evaluation slides avoids word tokenization?
a)
A syllable-based model
b)
A word-based model with golden segmentation
c)
A word-based model with automatic segmentation
d)
A CRF model with chunk tags
e)
A dictionary-based model
17.
Which feature type is mentioned in the evaluation slide as coming from a clustering representation?
a)
Brown clustering representation
b)
K-means over pixels
c)
Hierarchical clustering of fonts
d)
DBSCAN over URLs
e)
Spectral clustering of slides
18.
What kind of approach is introduced by the slide titled 'RNN-based NER'?
a)
Neural sequence labeling for NER
b)
Rule-based web scraping
c)
Hand-crafted regular expressions only
d)
Database indexing for IR
e)
Graph shortest-path search
19.
In the neural NER input layer described in the slides, character representation is learned using what?
a)
A bidirectional LSTM network
b)
A decision tree
c)
A Naive Bayes classifier
d)
A k-nearest neighbors model
e)
A page-rank model
20.
The slide titled 'Learn to represent characters' indicates the model should learn representations of what?
a)
Characters
b)
Web pages
c)
Database tables
d)
Audio signals
e)
Image pixels
21.
According to the slides, the purpose of a bidirectional LSTM is to use what information for predictions?
a)
Both previous and following words in the sentence
b)
Only previous words
c)
Only following words
d)
Only word length
e)
Only punctuation marks
22.
In the neural NER output layer slide, what labels are predicted?
a)
BIO labels for entity types
b)
Page categories for websites
c)
Cluster IDs for documents
d)
URL hashes
e)
HTML tag types only
23.
According to the BiLSTM-CRF evaluation slide, BiLSTM-CRFs use additional information such as what?
a)
PoS and clustering information
b)
Only slide templates
c)
Only named entity dictionaries
d)
Only website anchor text
e)
Only image features
24.
Which trade-off between supervised and unsupervised relation extraction is stated in the slides?
a)
Supervised learning is accurate but requires training data
b)
Unsupervised learning is always more accurate
c)
Supervised learning needs no labeled data
d)
Unsupervised learning cannot use large corpora
e)
They are identical in accuracy
25.
In the Snowball method for relation extraction, what is the initial input mentioned in the slides?
a)
Seed tuples
b)
Dependency trees
c)
Topic models
d)
Word embeddings only
e)
HTML templates
26.
In Snowball, seed tuples are described as being provided by whom?
a)
The user
b)
A web crawler
c)
A POS tagger
d)
A database index
e)
A translation system
27.
After defining seed tuples, the slides describe searching for what in the corpus?
a)
Occurrences of the seed tuples
b)
All stopwords
c)
All rare characters
d)
All file headers
e)
All hyperlinks
28.
In the Snowball pipeline, what is the purpose of entity labeling as shown in the slides?
a)
Replace specific entities with entity-type tags (for example ORG and LOC)
b)
Convert all words to uppercase
c)
Remove all punctuation
d)
Translate the corpus to English
e)
Generate random features
29.
The slides define a 5-tuple for Snowball as which structure?
a)
Left context, first entity tag, middle context, second entity tag, right context
b)
Sentence, paragraph, document, corpus, web
c)
Subject, verb, object, adjective, adverb
d)
Prefix, stem, suffix, lemma, root
e)
Token, POS, chunk, dependency, topic
30.
In the 5-tuple representation, what does the 'middle' part refer to?
a)
Words between the two entity tags
b)
Words before the first entity only
c)
Words after the second entity only
d)
Only punctuation tokens
e)
Only entity types with no context
31.
In a Snowball 5-tuple, which component captures words to the right of the second entity?
a)
Right context
b)
Left context
c)
Tag 1
d)
Middle context
e)
Tag 2
32.
According to the slides, extraction patterns in Snowball are produced by doing what with 5-tuples?
a)
Clustering similar 5-tuples and using cluster centroids as patterns
b)
Sorting 5-tuples alphabetically
c)
Replacing all weights with zeros
d)
Keeping only the longest sentences
e)
Randomly sampling half the tuples
33.
In the Snowball tuple generation algorithm, what is tau_sim used for?
a)
A similarity threshold to decide if an occurrence matches a pattern
b)
A maximum sentence length
c)
A learning rate for gradient descent
d)
A stopword probability
e)
A tokenization delimiter
34.
How is a pattern confidence conf(P) computed in the slides?
a)
Positive matches divided by positive plus negative matches
b)
Negative matches divided by positive matches
c)
Total matches divided by unknown matches
d)
Unknown matches divided by total matches
e)
Positive matches multiplied by negative matches
35.
According to the slides, a tuple is included in the pattern example set when what condition holds?
a)
Its confidence exceeds a threshold
b)
It appears in the title slide
c)
It has zero left context
d)
It has no entity tags
e)
It contains only numbers
36.
Which drawback of the Snowball approach is listed in the 'Pros, Cons' slide?
a)
The iterative process can lead to quality degradation
b)
It cannot use unlabeled data
c)
It requires no manual input
d)
It never produces new tuples
e)
It works only on images
37.
In distant supervision, what resource is described as a large knowledge base of relationships between entities?
a)
Freebase
b)
WordNet
c)
PageRank
d)
TF-IDF
e)
BLEU
38.
According to the slides, Freebase is built from what source?
a)
Wikipedia
b)
Twitter
c)
GitHub
d)
Stack Overflow
e)
Common Crawl
39.
According to the distant supervision slide, labeled data is created by combining what?
a)
A knowledge base and a text corpus
b)
Two tokenizers only
c)
A POS tagger and a stemmer
d)
A dependency parser and an image classifier
e)
A web crawler and a browser
40.
In the distant supervision workflow, what does the classifier predict for an entity pair based on extracted features?
a)
A relation label
b)
A file format
c)
A color palette
d)
A font family
e)
A time zone
41.
Which item is included in the feature set for relation classification under distant supervision?
a)
The dependency-tree path between two entities
b)
The slide number where entities appear
c)
The background image of the page
d)
The speaker volume in audio
e)
The PDF page margins
42.
The dependency tree slide indicates that features can be derived from what kind of syntactic structure?
a)
A dependency tree
b)
A decision stump
c)
A confusion matrix
d)
A topic hierarchy
e)
A regex library
43.
According to the slides, coreference resolution is important for information extraction because it helps do what?
a)
Link different mentions that refer to the same entity
b)
Detect HTML tags
c)
Compute PageRank scores
d)
Compress documents
e)
Generate word clouds
44.
Which type of coreference is explicitly listed in the slides?
a)
Possessive pronoun coreference
b)
Rhyming-word coreference
c)
Color-based coreference
d)
Font-style coreference
e)
Audio-timbre coreference
45.
Which coreference type is described as using a first name to refer to a previously mentioned person?
a)
First-name reference
b)
Hyperlink reference
c)
Acronym expansion
d)
Date normalization
e)
Token splitting
46.
Which coreference type links an entity to one of its components, as listed in the slides?
a)
Part-whole
b)
Synonymy
c)
Antonymy
d)
Homophony
e)
Metonymy
47.
Traditional coreference methods in the slides eliminate candidates using properties such as what?
a)
Gender and singular/plural agreement
b)
Background color and font size
c)
File hash and checksum
d)
CPU speed and memory
e)
Network latency and bandwidth
48.
The neural coreference slide states that coreference resolution is essentially what kind of problem within text?
a)
A hard clustering problem
b)
A sorting problem
c)
A shortest-path problem
d)
A compression problem
e)
A matrix factorization problem
49.
In the end-to-end neural coreference model architecture, what component provides contextual information over the sentence?
a)
A bidirectional LSTM
b)
A unigram language model
c)
A decision tree
d)
A k-means clusterer
e)
A regular expression engine
50.
In the coreference scoring diagram, what does epsilon represent?
a)
No antecedent (the span does not refer to any previous span)
b)
The first entity in the document
c)
The sentence boundary
d)
A dependency relation label
e)
The document title
51.
In the input layer of the end-to-end coreference model, how are out-of-vocabulary words represented?
a)
As a zero vector
b)
As a random URL string
c)
As a POS tag only
d)
As a dependency path only
e)
As a topic ID only
52.
According to the character representation slide, what neural component is used to build character-level features?
a)
A CNN over characters
b)
A k-means clusterer
c)
A suffix dictionary
d)
A PageRank graph
e)
A TF-IDF vectorizer
53.
The slides state that the final contextual word representation is formed by doing what with forward and backward LSTM outputs?
a)
Concatenating them
b)
Averaging them into a single scalar
c)
Sorting them alphabetically
d)
Replacing them with zeros
e)
Removing one direction entirely
54.
Which element is included in the span representation gi described in the slides?
a)
A representation of the span length
b)
Only the document ID
c)
Only the slide title
d)
Only punctuation counts
e)
Only URL features
55.
How is the soft representation of the main word in a span computed according to the slides?
a)
Using attention weights learned by a feed-forward network
b)
Using a fixed stopword list
c)
Using a regular expression
d)
Using manual rules only
e)
Using document-level clustering
56.
What does the mention score sm(i) represent in the neural coreference model as described in the slides?
a)
A score indicating how likely span i is a mention
b)
The number of sentences in the document
c)
The file size of the PDF
d)
The number of web pages crawled
e)
The number of entity types
57.
In pairwise scoring between spans i and j, the slides state that phi(i, j) can include what information?
a)
Speaker info, gender, and distance between spans
b)
Font color and page margin
c)
URL length and domain
d)
Image resolution and aspect ratio
e)
Keyboard layout and locale
58.
The loss function slide describes using what kind of probabilities over antecedent choices?
a)
Marginal probabilities
b)
Only uniform probabilities
c)
Only binary probabilities
d)
Only random probabilities
e)
Only deterministic rules
59.
In an ablation evaluation, what is the purpose of removing one model component at a time?
a)
To measure each component's contribution to performance
b)
To reduce the number of slides
c)
To change the PDF theme
d)
To increase font size
e)
To remove all entities
60.
The span-width evaluation chart in the slides analyzes how span width relates to what?
a)
Precision measures and how often spans occur
b)
Network throughput and latency
c)
GPU utilization and memory
d)
File compression ratio and speed
e)
Browser cache hits and misses
61.
The highlighted-text examples in the evaluation slides illustrate which task?
a)
Coreference resolution across mentions in a document
b)
Image classification in photos
c)
Web page ranking by links
d)
Topic modeling of documents
e)
Tokenization of URLs only
62.
Which scenario most directly violates an information extraction assumption stated in the slides?
a)
A relation must be inferred from world knowledge rather than explicitly stated
b)
Entities are written in uppercase letters
c)
The text contains punctuation
d)
A document contains more than one sentence
e)
A word appears multiple times
63.
An information extraction system expects that necessary information appears locally. What does 'locally' imply in this context?
a)
Relevant cues are near the mentions in the text rather than scattered across distant sections
b)
All evidence must be in the document title
c)
Evidence must be in a separate database
d)
Evidence must be in images only
e)
Evidence must be in footnotes only
64.
Dictionary-based NER can detect common entities but cannot handle ambiguity. Which example best reflects that limitation?
a)
A single surface form could refer to different entity types or entities depending on context
b)
A sentence has no punctuation
c)
A token contains digits
d)
A document has multiple paragraphs
e)
A named entity spans multiple words
65.
Why does the slides' feature window idea (words in window around a token) help sequence labeling for NER?
a)
Nearby words provide contextual cues that influence the token's entity label
b)
It eliminates the need for any training data
c)
It guarantees perfect precision
d)
It replaces the BIO labeling scheme
e)
It removes the need for tokenization
66.
In BIO tagging, what is the key difference between B-ORG and I-ORG for a multiword organization mention?
a)
B-ORG marks the first word of the organization, while I-ORG marks subsequent words inside it
b)
B-ORG marks words outside the organization, while I-ORG marks inside
c)
B-ORG is used only for locations, while I-ORG is only for people
d)
B-ORG indicates punctuation, while I-ORG indicates digits
e)
They are interchangeable labels with no difference
67.
The slides compare CRF-based NER with and without PoS and chunking. What is the most direct reason those features can matter?
a)
They provide syntactic structure that can help distinguish entity boundaries and types
b)
They convert Vietnamese to English
c)
They ensure all words are lowercase
d)
They make BIO tags unnecessary
e)
They replace the need for evaluation metrics
68.
In the neural NER architecture, why combine word embeddings with character-level representations?
a)
Character-level signals can capture morphology and handle unseen or rare word forms
b)
Character-level models remove the need for any labels
c)
Word embeddings cannot be used with LSTMs
d)
Character models only help with punctuation tokens
e)
It forces the model to ignore context
69.
The output layer slide notes that a CRF can be placed on top of neural outputs. What benefit does the CRF provide?
a)
It models dependencies between adjacent labels using transition probabilities
b)
It learns word embeddings without any text
c)
It converts sequence labeling into document classification
d)
It performs web crawling
e)
It removes entity types from the label set
70.
The slides state that supervised relation extraction is accurate but requires training data. In contrast, unsupervised methods trade accuracy for what advantage?
a)
They can exploit large amounts of unlabeled text
b)
They always use a knowledge base
c)
They require less computation than rule systems
d)
They never need entity labeling
e)
They guarantee no manual effort
71.
In Snowball, why is entity labeling (replacing names with ORG/LOC tags) an important step before pattern induction?
a)
It generalizes contexts so patterns can match many different entity pairs
b)
It hides the relation type from the system
c)
It prevents tokenization errors entirely
d)
It forces all tuples to be unique
e)
It converts extraction into topic modeling
72.
The Snowball similarity function combines left, middle, and right context matches. What does a high match score indicate?
a)
Two occurrences share similar contextual evidence around the entity pair
b)
Two entities have identical names
c)
Two documents share the same author
d)
Two sentences have the same length
e)
Two corpora use the same file format
73.
The Snowball algorithm uses a similarity threshold tau_sim. If tau_sim is set too high, what is the most likely outcome?
a)
Fewer occurrences will match patterns, reducing the number of candidate tuples found
b)
More noisy tuples will be accepted, increasing drift
c)
The system will stop tokenizing text
d)
Entity labeling will be skipped automatically
e)
Confidence will always become 1.0
74.
Pattern confidence conf(P) uses positive and negative matches. What does a low confidence value mean?
a)
The pattern often matches incorrect tuples relative to correct ones
b)
The pattern matches only unknown examples
c)
The pattern never matches any example
d)
The pattern can only match people entities
e)
The pattern is guaranteed to be correct
75.
The Snowball 'Pros, Cons' slide warns about quality degradation over iterations. Which mechanism best explains that risk?
a)
Early errors can introduce bad tuples that generate bad patterns in later rounds
b)
Adding more text always reduces accuracy
c)
Using weight vectors prevents learning
d)
Clustering eliminates all useful patterns
e)
Manual labeling guarantees drift
76.
In distant supervision, how does the system obtain labels for training examples without manual annotation?
a)
It matches entity pairs in text to known triples in a knowledge base
b)
It uses only regular expressions for labeling
c)
It labels every sentence as positive
d)
It uses slide titles as labels
e)
It infers labels from file extensions
77.
Why is the order of two entities included as a feature in distant supervision relation classification?
a)
Some relations are directional, so swapping entity order can change the relation interpretation
b)
Entity order determines tokenization quality
c)
Entity order determines the PoS tag set
d)
Entity order removes the need for dependency parsing
e)
Entity order guarantees perfect recall
78.
Among the distant supervision features, what is the role of the dependency-tree path between two entities?
a)
It captures the syntactic relationship connecting the entities in the sentence
b)
It provides the document topic distribution
c)
It encodes the PDF page layout
d)
It lists all entities in the document
e)
It replaces word embeddings
79.
Coreference resolution is important for information extraction. Which IE outcome is most directly improved by resolving coreferences?
a)
Linking relations mentioned with pronouns back to the correct entity
b)
Changing the language of the document
c)
Compressing the corpus size
d)
Removing stopwords from text
e)
Generating POS tags
80.
Traditional coreference methods eliminate candidates by gender and number. What is the reasoning behind this step?
a)
An antecedent should agree with the mention on key grammatical properties
b)
Antecedents must always be in the same sentence
c)
Pronouns always refer to the nearest noun
d)
Gender agreement is unrelated to coreference
e)
Number agreement is only useful for relation extraction
81.
The neural coreference slide emphasizes using pre-trained representations and limiting parsers. What challenge does it mention as a result?
a)
Finding alternative signals to capture syntax and phrase context
b)
Inability to score mentions
c)
Inability to compute word embeddings
d)
Inability to form spans
e)
Inability to output any clusters
82.
The problem formulation considers many candidate spans. What does this imply about the number of possible spans as document length grows?
a)
It grows quadratically with the number of words
b)
It grows linearly with the number of sentences
c)
It stays constant regardless of length
d)
It grows only with the number of entity types
e)
It decreases as length increases
83.
In the span representation gi, why include both the first and last word representations of the span?
a)
They help encode the span boundaries and its surface extent
b)
They allow the model to ignore the span length
c)
They force the model to use only character CNNs
d)
They remove the need for attention
e)
They guarantee the span is a valid entity
84.
The slides describe a soft representation of the main word in a span using attention. What is the purpose of this mechanism?
a)
To emphasize the most informative token(s) within the span for coreference decisions
b)
To sort spans by length
c)
To replace word embeddings with POS tags
d)
To remove all multiword mentions
e)
To enforce gender agreement rules
85.
Pairwise scoring uses inputs like gi, gj, and their interaction. What is the motivation for including an interaction term between span representations?
a)
It helps capture how compatible two spans are beyond their individual properties
b)
It prevents the model from using any distance feature
c)
It forces similarity to be zero for all pairs
d)
It replaces mention scoring entirely
e)
It encodes only document-level metadata
86.
What is the most direct interpretation of the mention score sm(i) in the end-to-end coreference model?
a)
How likely span i should be considered as a mention candidate
b)
How many entities are in the document
c)
The POS tag of span i
d)
The dependency depth of span i
e)
The topic probability of span i
87.
In phi(i, j), distance is included as a feature. Why is mention distance relevant for coreference?
a)
Mentions that are far apart may be less likely to refer to each other than nearby ones
b)
Distance determines word embedding dimensions
c)
Distance determines how many entity types exist
d)
Distance removes the need for a BiLSTM
e)
Distance is only used for tokenization
88.
The ablation table removes components like attention, character CNNs, or embeddings. What does a performance drop after removal indicate?
a)
The removed component contributed useful information to the model
b)
The removed component was redundant and harmful
c)
The model no longer has any spans
d)
The evaluation metrics are invalid
e)
The document contains no entities
89.
The span-width analysis compares precision measures with how often spans occur. What is the practical reason to examine span width in a span-based model?
a)
Span width affects both the space of candidates and how reliably spans can be predicted
b)
Span width determines the number of entity types
c)
Span width determines the size of the training corpus
d)
Span width changes the knowledge base relations
e)
Span width is unrelated to coreference resolution
90.
The highlighted evaluation examples show multiple mentions grouped together. What is the intended output structure of coreference resolution implied by this?
a)
Clusters of mentions that refer to the same entity
b)
A single label per document
c)
A ranked list of relations only
d)
A set of tokenization rules
e)
A list of stopwords
91.
Distant supervision is described as improving accuracy over unsupervised relation extraction. Based on the slides, what key additional resource enables this?
a)
A knowledge base of entity relations
b)
A larger stopword list
c)
A higher-resolution PDF
d)
A faster tokenizer
e)
A different font
92.
Which feature set element is shared by both NER and relation classification slides, indicating the importance of local context?
a)
Using words around a target (such as between entities or within a window)
b)
Using slide background colors
c)
Using file names as features
d)
Using line breaks as labels
e)
Using audio metadata
93.
The lecture content includes both NER and coreference resolution. What is a key distinction between these tasks as implied by the slides?
a)
NER labels spans with entity types, while coreference links mentions that refer to the same entity
b)
NER ranks web pages, while coreference detects spam
c)
NER extracts relations, while coreference tokenizes words
d)
NER clusters documents, while coreference summarizes text
e)
NER performs translation, while coreference does stemming
94.
In the Snowball 5-tuple representation, context words are associated with weights. What is the most likely purpose of using weights?
a)
To emphasize more informative context words when comparing occurrences
b)
To encode the PDF page number
c)
To ensure every tuple has the same length
d)
To convert entity types into numerical IDs only
e)
To eliminate the need for clustering
95.
An information extraction system assumes a small number of patterns can summarize the text. Which situation most directly challenges this assumption?
a)
The same relation is expressed in many diverse linguistic forms across the corpus
b)
All entities appear in lowercase
c)
Every sentence has exactly one verb
d)
The document is short
e)
The text is tokenized correctly
96.
If an IE pipeline relies on the 'no inference' assumption, which type of relation mention is most likely to be missed?
a)
A relation implied indirectly without explicit relational phrasing
b)
A relation stated with a clear verb between two entities
c)
A relation stated in apposition
d)
A relation stated with a fixed template
e)
A relation stated with entity-type tags
97.
Dictionary-based NER cannot handle ambiguity. Which mitigation is most aligned with the slides' alternative approaches?
a)
Use a machine-learning sequence labeling model that leverages context features
b)
Increase the font size of entity mentions
c)
Remove all multiword entities
d)
Translate the corpus to another language
e)
Replace all entities with random strings
98.
BIO labeling can fail when entity boundaries are inconsistent. Which labeling is invalid for a single entity mention spanning two words?
a)
I-ORG followed by B-ORG
b)
B-ORG followed by I-ORG
c)
B-LOC followed by I-LOC
d)
B-PER followed by I-PER
e)
B-ORG followed by O
99.
A CRF placed on top of neural outputs models label transitions. Which constraint is most naturally captured by such transitions in BIO tagging?
a)
I-ORG should not directly follow O without a preceding B-ORG
b)
All tokens must be labeled O
c)
All entities must be single-word
d)
Every sentence must contain an ORG
e)
POS tags must be perfect
100.
The CRF evaluation compares golden versus automatic PoS and segmentation. What is the most plausible explanation for why 'golden' preprocessing can improve results?
a)
It reduces upstream noise so the NER model receives cleaner input signals
b)
It increases the number of entity types
c)
It removes the need for any features
d)
It forces all words to be out-of-vocabulary
e)
It prevents the model from using context
101.
Snowball iteratively expands tuples and patterns but risks quality degradation. Which failure mode best matches this warning?
a)
Semantic drift: newly accepted tuples gradually shift patterns away from the target relation
b)
Overfitting: the model memorizes the training labels
c)
Underflow: probabilities become negative
d)
Data leakage: test labels are copied into training
e)
Mode collapse: embeddings become identical
102.
Snowball pattern confidence uses positive and negative examples. Which change would most directly increase conf(P) for a fixed pattern P?
a)
Increase the number of positive matches while keeping negative matches constant
b)
Increase the number of negative matches while keeping positive matches constant
c)
Set tau_sim to zero
d)
Remove entity labeling
e)
Delete the corpus
103.
In the tuple selection rule Conf(T) > tau_t, what is the most direct effect of increasing tau_t?
a)
Fewer tuples are accepted into the pattern example set
b)
More tuples are accepted into the pattern example set
c)
Patterns are no longer clustered
d)
Entity order features are removed
e)
Dependency paths become longer
104.
Distant supervision creates labeled data by matching knowledge-base triples to sentences containing the entity pair. Which hidden assumption is implicit in this setup?
a)
Sentences that mention the entity pair provide usable evidence about the relation
b)
All sentences have exactly two entities
c)
Every relation is symmetric
d)
All entities are people
e)
POS tags are unnecessary
105.
The distant supervision feature set includes both words between entities and the dependency path. When might the dependency path add value beyond surface words?
a)
When syntax clarifies which words govern the relation between the entities
b)
When the sentence contains no verbs
c)
When entities are not in the same document
d)
When the corpus is unlabeled
e)
When the knowledge base is missing
106.
Coreference resolution is described as a hard clustering problem. What does 'hard' clustering imply about mention assignment?
a)
Each mention is assigned to exactly one cluster (entity) in the final output
b)
Each mention belongs to all clusters with equal probability
c)
Clusters are never merged
d)
Mentions are ignored if they are pronouns
e)
Clusters are defined only by POS tags
107.
Traditional coreference methods score candidates after eliminating some by constraints like gender and number. In what situation would these constraints be least helpful?
a)
When the language provides weak or ambiguous gender/number cues for the mentions
b)
When the document is short
c)
When the entity is a location
d)
When the mention is a proper noun
e)
When sentences are tokenized
108.
The end-to-end coreference formulation considers many candidate spans. What computational issue does this create that a practical system must manage?
a)
A very large number of span candidates that grows quickly with document length
b)
A shortage of entity types
c)
A lack of word embeddings
d)
An inability to compute POS tags
e)
A need to render images
109.
In span representation gi = [start, end, soft head, length], which component most directly helps identify the head word within a multiword span?
a)
Soft head representation computed with attention
b)
Start word representation
c)
End word representation
d)
Length feature
e)
Document embedding
110.
Pairwise scoring uses both gi and gj as well as an interaction term. Which pair would most benefit from including distance and speaker/gender features in phi(i, j)?
a)
Two pronoun mentions where speaker or gender cues help decide if they refer to the same entity
b)
Two identical proper nouns adjacent to each other
c)
Two unrelated sentences in different documents
d)
Two URLs with the same domain
e)
Two stopwords in a sentence
111.
The slides include an ablation study where components are removed. If removing character-level CNN features hurts performance, what is the most defensible interpretation?
a)
Character-level information is providing useful cues that the model cannot fully replace with word embeddings alone
b)
Character CNNs are only decorative and do not affect predictions
c)
The model no longer uses BIO tags
d)
The knowledge base becomes larger
e)
The dependency tree is no longer available
112.
The span-width analysis compares precision and frequency. Why might very long spans create difficulties for coreference models?
a)
Long spans can be less precise and much rarer, making them harder to model reliably
b)
Long spans always correspond to organizations
c)
Long spans remove the need for mention scoring
d)
Long spans guarantee perfect head-word selection
e)
Long spans eliminate ambiguity entirely
113.
In Snowball, if tau_sim is lowered too much, which risk becomes more likely during tuple generation?
a)
More weakly matching occurrences are accepted, increasing noise in candidate tuples
b)
No occurrences will ever match patterns
c)
Entity labeling will fail deterministically
d)
Clustering becomes unnecessary
e)
Confidence becomes undefined
114.
Which statement best synthesizes the lecture's pipeline from entity recognition to coreference resolution for information extraction?
a)
First identify entity mentions and their types, then extract relations and finally link mentions that refer to the same entity
b)
First build a knowledge base, then translate the text, then cluster web pages
c)
First remove stopwords, then compute PageRank, then run OCR
d)
First tokenize into syllables, then generate web hyperlinks, then compress the PDF
e)
First detect events, then ignore entities, then label topics
115.
A relation extractor uses features 'words between entities' and 'entity order'. Which case most strongly motivates including entity order as a feature?
a)
A directional relation where swapping subject and object changes the meaning
b)
A symmetric relation like synonymy
c)
A sentence with only one entity
d)
A document with no verbs
e)
A corpus with no entity types
116.
When generating extraction patterns, the slides compare 5-tuples that share the same tag1 and tag2. Why is this restriction important?
a)
It ensures patterns are learned from comparable entity-type pairs rather than mixing unrelated types
b)
It guarantees all tuples have the same sentence length
c)
It removes the need for similarity computation
d)
It forces all weights to be identical
e)
It prevents any new tuples from being generated
100 %
