Introduction to Large Language Models

Assignment- 9

Number of questions: 10 Total mark: 10 X 1 = 10

_________________________________________________________________________

QUESTION 1: [1 mark]

Which of the following statement best describes why knowledge graphs (KGs) are
considered more powerful than a traditional relational knowledge base (KB)?

a. KGs require no schema, whereas KBs must have strict schemas.

b. KGs store data only in the form of hypergraphs, eliminating redundancy.
c. KGs allow flexible, graph-based connections and typed edges, enabling richer
relationships and inferences compared to KBs.
d. KGs completely replace the need for textual sources by storing all possible facts.

Correct Answer: c

Explanation:

• Traditional relational knowledge bases enforce strict schemas (rows, columns,

tables). In contrast, KGs store entities as nodes with typed edges (relations) between
them, allowing richer and more flexible relationships.

• KGs can represent complex connections (e.g., multi-edges, different relation types)
and support inference by traversing these connections.

• While some knowledge graphs can be partially schema-less or schema-flexible,

choice (c) specifically highlights the flexibility in graph connections and typed edges,
which is what makes KGs generally more powerful for representing rich, interlinked
data.

_______________________________________________________________________

QUESTION 2: [1 mark]

Entity alignment and relation alignment are crucial between KGs of different languages.
Which of the following factors contribute to effective alignment?

a. Aligning relations solely by their lexical similarity, ignoring semantic context

b. Transliteration or language-based string matching for entity labels
c. Ensuring all language aliases are represented identically in each KG
d. Matching neighbours, or connected entities, across different KGs

Correct Answer: b, d

Explanation:
 • Transliteration or language-based string matching (b): For multilingual KGs,
matching entity labels across languages (e.g., transliterating names or matching
synonyms) is key to identifying equivalent entities.

• Matching neighbours (d): Alignment goes beyond just matching labels; you also
look at the graph structure - if two entities are neighbours or connected to the same
concept in each KG, that supports they might be the same or related.

• Why not a or c?

(a) Lexical similarity alone (ignoring context) is insufficient, as terms might match
lexically but differ in meaning across languages.

(c) It is often not possible or necessary to have identical labels for all language
aliases if they mean the same thing in context; the important aspect is linking them,
not forcing identical representations.

_________________________________________________________________________

QUESTION 3: [1 mark]

In the context of knowledge graph completion (KGC), which statement best describes the
role of the scoring function 𝑓(𝑠, 𝑟, 𝑜)?

a. It determines whether two entities refer to the same real-world concept.

b. It produces a raw confidence score indicating how plausible a triple (𝑠, 𝑟, 𝑜) is.
c. It explicitly encodes only the subject’s embedding, ignoring the relation and object
embeddings.
d. It ensures that every negative triple gets a higher score than any positive triple.

Correct Answer: b

Explanation:

• A raw confidence score for plausibility (b): In KGC, the scoring function evaluates
a triple (𝑠, 𝑟, 𝑜) (subject, relation, object) to determine how likely it is to be true
according to the learned embeddings. A higher (or lower, depending on the
convention) score indicates higher plausibility.

• Why not the others?

(a) That relates to entity resolution, not necessarily the KGC scoring function.

(c) The scoring function typically factors in all three embeddings (subject, relation,
and object).
 (d) While in training we often prefer valid triples to have higher scores than negative
ones, the scoring function itself just produces a plausibility value. It doesn’t guarantee
all negative triples score higher or lower unconditionally.

_________________________________________________________________________

QUESTION 4: [1 mark]

One key difference between the differentiable KG approach and the semantic interpretation
approach to KGQA is:

a. Differentiable KG approaches are fully rule-based, while semantic interpretation is

purely neural.
b. Differentiable KG approaches do not require any graph embeddings, relying instead
on explicit logical forms.
c. Semantic interpretation is more transparent or interpretable, whereas differentiable
KG is end-to-end trainable but less interpretable.
d. Both approaches use logical forms; the primary difference is the type of question they
can answer.

Correct Answer: c

Explanation:

• Semantic interpretation is more interpretable, differentiable KG is more end-to-

end (c):
o Semantic interpretation methods typically rely on building an explicit logical
form of the question, which is transparent and can be easily explained.
o Differentiable KGQA uses neural embeddings and end-to-end
backpropagation over the graph, making it powerful but less human-
interpretable.

• Why not the others?

(a) Differentiable KG approaches are not fully rule-based — in fact, they’re more
neural and less rule-based.

(b) Differentiable approaches definitely do use graph embeddings.

(d) While both can handle complex questions, the key difference in c is the
interpretability vs. end-to-end trainability.

________________________________________________________________________

QUESTION 5: [1 mark]

Considering the differentiable KG approach, which elements are typically learned jointly
when training an end-to-end KGQA model?
 a. The textual question representation (e.g., BERT embeddings)
b. The graph structure encoding (e.g., GCN or transformer-based graph embeddings)
c. Predefined logical forms to ensure interpretability
d. The final answer selection mechanism that identifies which node(s) in the graph
satisfy the question

Correct Answer: a, b, d

Explanation:

• Textual question representation (a): The system learns how to embed the input
question (often with a neural model like BERT).
• Graph structure encoding (b): It also learns how to encode nodes and relations in
the knowledge graph (using a graph neural network, attention, etc.).
• Final answer selection (d): Finally, the model learns how to map from the question
and graph embeddings to the correct node(s) in the KG.
• Why not (c)? Predefined logical forms are more typical of semantic-parsing-based
KGQA, not differentiable KGQA approaches. Differentiable KGQA is usually end-to-
end and does not require manually crafted logical forms.

_________________________________________________________________________

QUESTION 6: [1 mark]

Uniform negative sampling can have high variance and may require large number of
samples. Why is that the case?

a. Because the margin-based loss cannot converge without big mini-batches.

b. Because randomly picking negative entities does not guarantee close or challenging
negatives, causing unstable training estimates.
c. Because negative sampling must ensure every possible negative triple is covered.
d. Because the number of relations in the KG is too large for small number of samples.

Correct Answer: b

Explanation:

• High variance arises when negatives are not challenging (b): If negative
examples are chosen completely at random, many will be too easy for the model to
distinguish, providing limited learning signal. The model sees less “borderline” cases,
so estimates of how well the model can separate real vs. fake facts fluctuate
significantly.

• Why not the others?

(a) Margin-based losses can converge with or without large mini-batches if sampling
is done carefully.
 (c) We don’t need to cover every possible negative triple, just enough meaningful
ones for training.

(d) The number of relations in a KG might be large, but that alone doesn’t necessarily
drive the variance issue.
_________________________________________________________________________

QUESTION 7: [1 mark]

In testing embedding and score quality for KG completion, mean rank and hits@K are typical
metrics. What does hits@K specifically measure in this context?

a. The percentage of queries for which the correct answer appears in the top-K of the
ranked list.
b. The reciprocal of the rank of the correct answer.
c. The probability of the correct answer appearing as the highest scored candidate.
d. The margin of the correct triple score relative to all negative triples.

Correct Answer: a

Explanation:

• Hits@K = The percentage of queries for which the correct entity (or triple) is in
the top-K predictions (a). This means if the correct answer is within the first K
results in the ranking, we call it a “hit.” We then compute how many queries achieve
this, divided by the total.

• Why not b, c, or d?

(b) That is more like Mean Reciprocal Rank (MRR), not hits@K.

(c) If K=1, that might coincide with hits@1, but the metric hits@K is about the top-K in
general, not exclusively the top candidate.

(d) That describes a margin-based idea, not hits@K.

_________________________________________________________________________

QUESTION 8: [1 mark]

In the TransE model, the scoring function for a triple (𝑠, 𝑟, 𝑜) is typically defined as

𝑓(𝑠, 𝑟, 𝑜) =∥ 𝑒! + 𝑒" − 𝑒# ∥

where 𝑒! , 𝑒" , 𝑒# are embeddings of the subject, relation, and object, respectively. Which
statement best explains what a low value of 𝑓(𝑠, 𝑟, 𝑜) indicates in this context?

a. That (𝑠, 𝑟, 𝑜) is an invalid triple according to the learned embeddings.

 b. That 𝑒! and 𝑒# must be orthogonal.
c. That the relation embedding 𝑒" is zero.
d. That (𝑠, 𝑟, 𝑜) has a high likelihood of being a true fact in the knowledge graph.

Correct Answer: d

Explanation:

• A low distance = a high plausibility (d): In TransE, the model is trained such that
𝑠 + 𝑟 ≈ 𝑜 for a valid triple. If the norm ∥ 𝑠 + 𝑟 − 𝑜 ∥ is small, it means the subject,
relation, and object embeddings line up well, indicating that triple is likely true.

• Why not a, b, c?

(a) A high value would correspond to an invalid triple.

(b) Orthogonality is not directly indicated by a small distance in TransE.

(c) Zero relation embedding is not required for plausibility; the relation embedding
can be non-zero and still yield a low distance.

_________________________________________________________________________

QUESTION 9: [1 mark]

In RotatE, if a relation 𝑟 is intended to be symmetric, how would that typically manifest in the
complex plane?

a. The relation embedding 𝑒" must always equal zero.

$
b. The angle of 𝑒" must be .
%
c. The relation embedding 𝑒" is its own inverse (i.e., a 180° rotation when squared).
d. The magnitude of 𝑒" must be greater than 1.

Correct Answer: c

Explanation:

• Relation embedding is its own inverse (c): In RotatE, each relation is modeled as
a rotation in the complex plane. For a relation to be symmetric, applying that relation
twice would yield the original entity, so 𝑟 % = 1. A 180° rotation (i.e., 𝜋 radians) is its
own inverse because rotating twice by 180° brings you back to the same orientation.

• Why not a, b, or d?

(a) A zero embedding is not characteristic of symmetry in RotatE.

(b) An angle of 𝜋 radians alone indicates an inverse relation, but the question
specifically references the notion of the embedding being its own inverse, which
implies 𝑟 % = 1.
 (d) The magnitude constraint (often magnitude = 1 in RotatE) is not specifically about
symmetry.

_________________________________________________________________________

QUESTION 10: [1 mark]

Which main advantage do rotation-based models (like RotatE) have over translation-based
ones (like TransE) when it comes to complex multi-relational patterns in a KG?

a. Rotation-based models cannot model any symmetry or inverse patterns, so they are
simpler.
b. Rotation-based models handle a broader set of relation properties (symmetry, anti-
symmetry, inverses, composition) more naturally.
c. Rotation-based models have no hyperparameters to tune, unlike TransE.
d. Rotation-based models are guaranteed to yield perfect link prediction.

Correct Answer: b

Explanation:

• Rotation-based models can capture more complex relational properties (b): By

representing relations as rotations in the complex plane, RotatE naturally supports
symmetric relations (𝑟 % = 1) anti-symmetric relations, inverses (rotations in the
opposite direction), and composition (cumulative rotations).

• Why not a, c, or d?

(a) In fact, they can model symmetry, inverses, etc.

(c) They do have hyperparameters (e.g., embedding dimension, learning rate); it is

not hyperparameter-free.

(d) No model is guaranteed to be perfect for all link prediction tasks.

_________________________________________________________________________