AI 기술발표 영어 스피킹 실전 가이드 — Agentic Reasoning over Hyper-Relational KG

Part 0 · Why This Document

문서의 목적

이 문서는 AI 분야 연구자와 엔지니어가 영어로 기술 세미나, 국제학회 발표, 논문 발표, 사내 연구 발표를 보다 명료하고 설득력 있게 수행하도록 돕는 스피킹 훈련 자료다.

핵심 원칙은 단순하다. 발표 영어의 성패를 가르는 것은 어휘의 화려함이 아니라 논리의 가시성이다. 청중은 문장을 감상하러 오지 않는다. 논리를 따라오려고 앉아 있다. 따라서 학습의 목표도 원어민 발표자의 문장을 그대로 암기하는 데 있지 않다. 다음 흐름을 영어로 자연스럽게 연결하는 능력을 만드는 데 있다.

Problem → Why it matters → Key idea → Method → Evidence → Interpretation → Limitation → Takeaway → Q&A

이를 위해 이 문서는 다음 여덟 가지를 하나로 통합한다.

따라 말하기 좋은 AI 발표자와 YouTube 자료
AI 기술발표에서 반복적으로 사용하는 핵심 문장과 문구
Shadowing보다 한 단계 발전된 Technical Retelling 훈련법
하루 30~45분 훈련 루틴
8주 학습 계획
Agentic Reasoning using Hyper-Relational Knowledge Graphs를 주제로 한 12~15분 영어 세미나 예제
예상 Q&A와 즉답 훈련법
Hyper-relational KG와 agentic reasoning의 기술적 배경 자료

Part I · Role Models

어떤 발표자를 따라 연습할 것인가

모방은 창조의 어머니라는 말이 있지만, 아무나 모방한다고 다 어머니가 되는 것은 아니다. 발표 영어의 훈련에서는 순서가 중요하다. 명료함을 먼저 익히고, 내러티브를 배우고, 그다음에 즉흥적 기술 설명으로 난도를 올린다. 아래 네 사람이 그 순서다.

단계	발표자	학습 포인트	추천 이유
1	Andrew Ng	명료성, 짧은 문장, 핵심 강조	기술발표 영어의 기본 리듬을 익히기 좋다
2	Fei-Fei Li	연구 동기, 문제의식, 학술적 스토리텔링	연구의 의미를 설명하는 영어에 강하다
3	Andrej Karpathy	기술적 직관, LLM 설명, 즉흥적 설명	실제 AI 엔지니어식 기술 설명을 익히기 좋다
4	Pieter Abbeel	수식, 알고리즘, 강화학습 설명	강의형 기술 설명과 알고리즘 발표에 적합하다

한 사람만 먼저 선택해야 한다면 Andrew Ng을 권한다. 그다음 Fei-Fei Li → Andrej Karpathy 순서로 난도를 높이는 것이 좋다.

STEP 1 Andrew Ng 명료한 기술발표 영어

AI Dev 25 x NYC — Andrew Ng: Opening Keynote

Andrew Ng의 발표에서 주목할 것은 발음이 아니라 구조다. 그는 언제나 주장 → 이유 → 예 → 결론의 순서로 말한다. 발음을 그대로 흉내 내기보다 다음 논리 문장들을 하나의 덩어리로 익힌다.

One of the things we're seeing is...
The reason this matters is...
What this means is...
The key point is...

첫 2주의 목표

문장을 짧게 끊어 말하기
핵심 명사와 동사에만 강세 주기
불필요한 um, uh 줄이기
한 문장에 한 메시지만 넣기

STEP 2 Fei-Fei Li 학술적 스토리텔링

What we see and what we value — AI with a human perspective

Fei-Fei Li의 발표에서는 기술 그 자체보다 연결에 주목한다. 그의 발표는 motivation → problem → research direction → implication의 사슬로 짜여 있다. 특히 다음과 같은 연구 내러티브용 표현을 수집한다.

The question we want to ask is...
This brings us to...
What we have learned from this is...
Why does this matter?

연구의 "무엇"뿐 아니라 왜 중요한가를 설명하는 연습에 적합하다.

STEP 3 Andrej Karpathy 기술적 직관과 설명력

State of GPT — BRK216HFS, Microsoft Build

Karpathy는 복잡한 개념을 개념 → 그림 → 구조/수식 → 직관 → 예외/주의점의 순서로 풀어낸다. 그 순서를 관찰하면서 다음 표현을 연습한다.

So basically, what is happening here is...
You can think of this as...
Intuitively,...
The important thing to understand is...
In other words,...

이 단계의 목표는 문장을 그대로 따라 하는 것이 아니다. 익숙한 AI 개념을 슬라이드 없이 영어로 설명하는 것이다.

STEP 4 Pieter Abbeel 수식·알고리즘 설명

Foundations of Deep RL — 6-lecture series (Playlist)

수식이나 알고리즘을 단계적으로 설명해야 하는 발표자에게 특히 좋다. 강의형 기술 설명의 호흡, 즉 정의를 제시하고 직관을 붙이고 예외를 언급하는 리듬을 배울 수 있다.

Part II · Sentence Frames

AI 기술발표의 핵심 Sentence Frames

발표를 잘하는 사람은 발표 중에 문장을 만들지 않는다. 미리 자동화해 둔 문장 골격에 그날의 내용을 얹을 뿐이다. 아래는 발표의 각 국면에서 반복적으로 쓰이는 골격들이다.

06발표 시작 — 청중의 관심을 잡기

단순히 Today, I'm going to talk about...으로 시작하기보다 문제나 질문을 먼저 던지는 편이 강하다. 청중의 머릿속에 물음표를 심어 놓으면, 그다음부터는 청중이 스스로 답을 찾아 발표를 따라온다.

Let me start with a simple question.
I'd like to start with a simple observation.
Consider the following scenario.
Imagine that we want to...
One of the fundamental challenges in AI is...
A key question we want to address is...
The problem we're interested in is...
What I'd like to show you today is...
The main message of this talk is simple.
There is one idea I'd like you to take away from this talk.

07발표 구조 안내

Let me first give you some background.
I'll start by motivating the problem.
Then I'll introduce our approach.
After that, I'll walk you through the experimental results.
Finally, I'll discuss some limitations and future directions.
The talk is organized into three parts.
I'll focus on three key questions.
Before getting into the details, let me give you the big picture.

가장 재사용성이 높은 문장

Before getting into the details, let me give you the big picture.

08연구 문제 설명

The key challenge is that...
The fundamental problem is that...
One major limitation of existing approaches is...
Current methods struggle with...
Existing approaches typically assume that...
This assumption does not always hold in practice.
The difficulty comes from the fact that...
The problem becomes even more challenging when...
What makes this problem difficult is...
This raises an important question.

추천 조합

Existing approaches work well under controlled settings.
The problem, however, is that this assumption does not always hold in practice.

09왜 중요한가 설명하기

Why does this matter?
Why is this important?
This matters because...
The reason this is important is that...
This becomes particularly important when...
From a practical perspective,...
From a deployment perspective,...
From a systems perspective,...
This has important implications for...
If we can solve this problem, we can...

추천 조합

Why does this matter?
Because inference cost quickly becomes a bottleneck when the system scales to millions of queries.

10기존 연구의 한계

Previous work has primarily focused on...
Most existing approaches rely on...
A common approach is to...
However, these methods have several limitations.
One limitation is that...
Another issue is that...
This approach works well when..., but it becomes problematic when...
The main bottleneck is...
This creates a trade-off between A and B.
There is still a significant gap between...
This leaves an important question unanswered.

추천 조합

Existing methods achieve high accuracy.
However, they come at the cost of significantly higher inference latency.

11핵심 아이디어 소개

Our key idea is simple.
The basic idea behind our approach is...
The intuition is that...
Our approach is based on a simple observation.
Instead of doing A, we do B.
Rather than treating A and B separately, we...
The key insight is that...
What we propose is...
This leads us to our approach.
Based on this observation, we developed...

가장 강력한 구조

Our key idea is simple. Instead of A, we do B.

예 — Our key idea is simple. Instead of retrieving documents independently, we jointly reason over the graph structure.

12연구 기여 설명

Our work makes three main contributions.
Our first contribution is...
Second, we introduce...
Third, we demonstrate that...
To the best of our knowledge, this is the first...
We introduce a new framework for...
We develop an efficient method for...
We provide an extensive empirical evaluation of...
We show that our approach consistently outperforms...
More importantly, we show that...

주의. To the best of our knowledge는 실제 문헌 검토로 뒷받침할 수 있을 때만 사용한다.

13모델·시스템 아키텍처 설명

Let me walk you through the architecture.
At a high level, the system consists of three components.
The architecture has three main stages.
The first component is responsible for...
The second component takes this representation and...
The output is then passed to...
These two modules interact through...
The entire pipeline works as follows.
Given an input query, the model first...
This representation is then used to...
Finally, the model produces...

재사용 예

At a high level, the pipeline consists of three stages: retrieval, reasoning, and generation.

14그림·아키텍처 슬라이드 설명

What you're seeing here is...
This figure illustrates...
On the left, you can see...
On the right, we show...
The blue boxes represent...
The arrows indicate...
The important part here is...
I'd like to draw your attention to...
Notice that...
If you look closely, you can see that...
The key thing to notice here is...
Let me highlight one important detail.

슬라이드의 모든 요소를 읽지 않는다. 청중이 봐야 할 부분을 지정한다.

15수식·알고리즘 설명

You don't need to worry about all the details of this equation.
The important part is...
Intuitively, this equation says that...
In simple terms,...
You can think of this as...
Essentially, what we're doing is...
The first term captures...
The second term encourages...
We optimize this objective with respect to...
The intuition behind this loss function is...

좋은 발표자는 수식을 읽지 않는다. 수식의 의미를 영어로 번역한다.

16어려운 AI 개념을 쉽게 설명하기

You can think of this as...
A useful way to think about this is...
Conceptually,...
Intuitively,...
Essentially,...
Roughly speaking,...
In other words,...
Put differently,...
In simple terms,...
Another way of looking at this is...

예시

A useful way to think about retrieval-augmented generation is that
the language model gets an external memory system.

17실험 설정 설명

We evaluate our method on...
For evaluation, we use...
We compare our method against...
Our baselines include...
We report results on...
We use the same experimental setting across all methods.
All experiments were repeated...
The main evaluation metric is...
We measure both accuracy and efficiency.
For a fair comparison, we...

18결과 그래프 설명

그래프 안내

Let me walk you through this result.
The x-axis represents...
The y-axis shows...
Each line corresponds to...
Higher is better.
Lower is better.

결과 강조

As you can see,...
We observe a clear improvement.
Our method consistently outperforms the baselines.
The improvement is particularly significant when...
Interestingly, we see that...
What is surprising here is that...
The most important result is...
The key takeaway from this figure is...

숫자를 읽는 것보다 결론을 먼저 말하는 것이 중요하다.

19결과 해석

This suggests that...
This indicates that...
This tells us that...
One possible explanation is...
We believe this is because...
This result supports our hypothesis that...
This behavior can be explained by...
One interpretation is that...
Importantly, this improvement does not come at the cost of...
Taken together, these results suggest that...

예시

Accuracy improves by 7%.
This suggests that explicitly modeling the graph structure provides useful information that cannot be captured by text retrieval alone.

20비교 표현

Compared with the baseline,...
In contrast to existing methods,...
Unlike previous approaches,...
The main difference is that...
The advantage of our approach is...
The benefit becomes more apparent when...
We achieve comparable accuracy with significantly lower cost.
We improve A without sacrificing B.
This provides a better trade-off between A and B.

시스템 연구에 특히 유용한 표현

We obtain a better accuracy–efficiency trade-off.

21중요도 강조

The key point here is...
The important thing to remember is...
What matters here is...
Most importantly,...
More importantly,...
Crucially,...
In particular,...
It is worth emphasizing that...
I want to emphasize one point.
There is one thing I'd like you to remember.

very important를 반복하기보다 The key point here is...를 활용한다.

22Transition — 논리 연결

다음 주제로 이동

Now let's move on to...
With that in mind, let's look at...
This brings us to...
This leads naturally to...
Now that we've seen A, let's look at B.

다시 핵심 질문으로

Let me come back to the main question.
Going back to our original question,...
Recall that our goal was to...

앞 내용을 기반으로

Building on this idea,...
Based on this observation,...
Given these results,...
With this background in mind,...

23한계 설명

Our approach has several limitations.
There are still several open questions.
One limitation of our current approach is...
We have not yet evaluated...
Our experiments are currently limited to...
This method may not work well when...
We should interpret these results with some caution.
This remains an important direction for future work.
We see this as an opportunity for future research.

24과도한 주장 피하기

단정은 학문의 언어가 아니다. This proves that... 같은 문장은 반박당하는 순간 발표 전체의 신뢰를 무너뜨린다. 대신 다음을 사용한다.

This suggests that...
This provides evidence that...
Our results indicate that...
The results are consistent with the hypothesis that...
We observe that...
We find that...
This appears to...
In our experiments,...
Under this setting,...

25데모 설명

Let me show you a quick demo.
Here's what happens when...
Suppose we give the model...
As you can see, the system first...
Notice what happens here.
Now let's change the input.
This is where things get interesting.
In real time, the model...
This example illustrates...

26결론 시작

Let me wrap up.
Let me conclude with three key points.
To summarize,...
To recap,...
So what have we learned?
Let me leave you with three takeaways.
The main takeaway is...

27마지막 메시지

The key takeaway is...
If you remember one thing from this talk, it should be this: ...
Ultimately, our goal is to...
We believe this opens up new opportunities for...
We hope this work provides a step toward...
There is still a lot of work to be done.
And with that, I'd be happy to take questions.

28Q&A 시작

That's a good question.
That's an important question.
Thanks for bringing that up.
There are two aspects to this question.
Let me answer that in two parts.

모든 질문에 That's a great question.을 반복하면 기계적으로 들린다. 표현을 다양화한다.

29질문 이해 확인

If I understand your question correctly, you're asking whether...
Just to make sure I understood the question,...
Are you asking about A or B?
Let me restate the question.
I think the question is whether...

30생각할 시간 확보

That's an interesting point.
Let me think about that for a moment.
There are a couple of ways to look at this.
I think the short answer is yes, but...
The short answer is no. The longer answer is...
It depends on how we define...

이런 문장들을 자동화하면 um, uh가 줄어든다. 침묵을 메우는 소리 대신 침묵을 버는 문장을 쓰는 것이다.

31모르는 질문 대응

We haven't tested that yet.
We haven't looked at that systematically.
I don't have a definitive answer to that yet.
That's something we're currently investigating.
I don't want to speculate too much, but...
My intuition is that..., although we would need to verify this experimentally.
That's an excellent direction for future work.

연구자에게 특히 유용한 문장

My intuition is that X, although we would need to verify this experimentally.

32비판적 질문 대응

That's a fair point.
I agree that this is an important limitation.
You're right that...
That's certainly one way to interpret the result.
Our interpretation is slightly different.
The reason we made this design choice is...
There is indeed a trade-off here.
We chose A primarily because...

추천 조합

That's a fair point. Our interpretation is slightly different.

Part III · Automate First

먼저 자동화할 핵심 문장 — Essential 20

백 문장을 아는 것과 스무 문장을 말할 수 있는 것 중에 후자가 낫다. 처음에는 아래 20개부터 완전히 자동화한다. 생각하지 않고도 입에서 나오는 상태가 자동화다.

01 — Let me start with a simple question.
02 — The key challenge is that...
03 — Why does this matter?
04 — The reason this is important is that...
05 — The key idea behind our approach is...
06 — Our approach is based on a simple observation.
07 — Instead of A, we do B.
08 — At a high level, the system consists of three components.
09 — Let me walk you through how this works.
10 — The key thing to notice here is...
11 — Intuitively, what this means is...
12 — You can think of this as...
13 — The key takeaway from this figure is...
14 — This suggests that...
15 — One possible explanation is...
16 — With that in mind, let's move on to...
17 — There are, however, several limitations.
18 — If I understand your question correctly,...
19 — Let me answer that in two parts.
20 — If you remember one thing from this talk, it should be this: ...

Part IV · Training Method

어떻게 연습할 것인가

34핵심 원칙 — Shadowing보다 Technical Retelling

AI 기술발표에서는 30분짜리 영상을 처음부터 끝까지 계속 따라 말하는 것보다 하루에 1~3분 분량을 완전히 자기 것으로 만드는 편이 효과적이다. 넓게 훑는 학습은 안심을 주고, 깊게 파는 학습은 실력을 준다.

단순 shadowing은 다음까지만 훈련한다.

듣기 → 발음 → 리듬

실제 기술발표 능력으로 연결하려면 다음까지 가야 한다.

듣기 → 이해 → 논리 기억 → 자기 영어로 재설명 → 자신의 연구에 적용

이 마지막 단계가 Technical Retelling이다. 남의 문장을 흉내 내는 것이 아니라 남의 논리를 내 언어로 재구성하는 훈련이다.

356단계 35~45분 훈련법

그냥 듣기 5분

2~3분짜리 발표 구간을 자막 없이 본다. 목표는 세 가지다. 전체 논리 파악, 어디에서 강세를 주는지 관찰, 어디에서 멈추는지 관찰. 100% 이해하려고 하지 않는다.

영어 자막으로 분석 5분

모르는 모든 단어를 찾지 않는다. 다음 세 종류만 표시한다 — technical phrase, transition phrase, explanation phrase. 예를 들어 The key idea here is..., What this allows us to do is..., In practice, however,... 같은 문장들이다.

Chunk Shadowing 10분

한 문장을 의미 단위로 끊는다. What we found / is that / larger models / tend to... 처음에는 필요하면 0.75~0.9배속으로 듣고, 최종적으로 1.0배속에서 따라간다.

Pause & Reproduce 5분

한 문장을 듣고 화면을 멈춘다. 자막을 보지 않고 같은 의미를 영어로 다시 말한다. 똑같은 문장을 재현하는 것이 목표가 아니다. 의미가 유지되면 성공이다.

Technical Retelling 10분

영상을 끄고 발표자가 설명한 내용을 1~2분간 자신의 영어로 설명한다. Let me briefly explain how a GPT model is trained.으로 시작해 핵심 논리를 재구성한다.

녹음·자가 평가 5분

네 가지를 확인한다. 불필요한 pause, um·uh의 빈도, 문장 끝이 작아지거나 흐려지는 현상, 기술용어에만 지나치게 힘을 주는 현상. 문법 오류를 한 번에 전부 고치려 하지 않는다.

36Slot Substitution

문장 골격 하나를 선택하고 자기 연구로 5~10개 변형한다. 이 훈련은 phrase를 암기 대상에서 즉석 생성 도구로 바꾼다.

Frame

The key challenge is that ~~______~~.

The key challenge is that LLMs can generate plausible but unsupported answers.
The key challenge is that retrieval quality decreases as the knowledge base grows.
The key challenge is that temporal information changes over time.
The key challenge is that the model has limited access to external knowledge.

37Slide Drill

자신의 논문 슬라이드 한 장을 띄우고 60초 동안 영어로 설명한다. 반드시 다음 세 문장을 한 번씩 사용한다.

Mandatory Three

~~What you're seeing here is...~~
~~The key thing to notice here is...~~
~~What this tells us is...~~

예 — What you're seeing here is the overall architecture of our system. The key thing to notice here is the interaction between the retriever and the reasoning module. What this tells us is that retrieval and reasoning are performed iteratively rather than independently.

38즉석 Q&A Drill

질문을 스스로 만든다. 예컨대 Why didn't you use a larger language model? 그리고 다음 패턴으로 답한다.

Answer Pattern

That's a good question. / There are ~~two reasons~~ for this. / ~~First,~~... / ~~Second,~~... / So, ~~in our current setting,~~...

39하루 30분 압축 루틴

시간	훈련
5분	AI 발표 1~2분 듣기
5분	핵심 phrase 5개 shadowing
5분	phrase 하나당 여러 문장 변형
5분	슬라이드 1장 영어 설명
5분	영상 내용을 자신의 말로 retelling
5분	즉석 Q&A 3문제

하루 100문장을 공부하는 것보다 5문장을 실제로 말할 수 있게 만드는 것이 더 중요하다.

Part V · 8-Week Plan

8주 학습 계획

Week 1–2 — Andrew Ng: Clarity

목표. 짧고 명료한 문장, 문장 끊기, 핵심어 강조, 그리고 Problem → Why it matters 구조의 자동화.

연습. 하루 1~2분 구간을 선택해 같은 구간을 반복한다. 마지막 날에는 영상 없이 3분간 재설명한다.

Week 3–4 — Fei-Fei Li: Research Narrative

목표. motivation → problem → approach → implication의 서사 구조를 몸에 익힌다. 특히 다음 네 문장을 집중 연습한다.

The question we want to ask is...
This brings us to...
Why does this matter?
What we have learned from this is...

자신의 연구도 같은 구조로 3분간 설명한다.

Week 5–7 — Andrej Karpathy: Technical Explanation

60~120초 분량만 선택한다. 예를 들어 다음 개념을 슬라이드 없이 설명한다.

tokenization · pretraining · fine-tuning
retrieval · inference
agentic reasoning · knowledge graph

개념을 아는 것 → 영어로 직관을 설명하는 것

Week 8 — Own Your Talk

이 단계에서는 모방을 줄인다. 자신의 연구 주제로 5분 영어 기술발표를 한다. 가능하면 5장 정도의 슬라이드를 사용한다.

변화의 목표

시점	도달 상태
Week 1	발표자가 말한 문장을 따라 말한다
Week 4	발표자가 설명한 내용을 자신의 영어로 다시 설명한다
Week 7	기술 개념을 자신이 강연자라고 생각하고 설명한다
Week 8	자신의 연구를 자신의 영어로 발표한다

Part VI · Five-Sentence Formula

설득력 있는 기술발표의 5문장 공식

기술발표를 다음 다섯 단계로 생각하면 논리가 강해진다. 숫자를 읽는 발표가 아니라 결론을 설득하는 발표로 바뀌는 핵심 구조다.

Claim

Our method significantly improves retrieval accuracy.

Evidence

As you can see, it outperforms all baselines across the three datasets.

Interpretation

This suggests that graph structure provides complementary information to semantic similarity.

Qualification

However, the improvement becomes smaller on very small datasets.

Takeaway

So the key takeaway is that structural information becomes particularly useful at scale.

Claim → Evidence → Interpretation → Qualification → Takeaway

Part VII · Final Goal

최종적으로 즉답할 수 있어야 하는 10개 질문

다음 질문 각각에 1분 동안 영어로 답하는 것을 최종 목표로 삼는다. 열 개의 답이 준비되어 있으면, 어떤 발표든 뼈대는 이미 완성되어 있는 셈이다.

01 — What problem are you trying to solve?
02 — Why is this problem important?
03 — What is wrong with existing approaches?
04 — What is your key idea?
05 — How does your method work?
06 — What is your main contribution?
07 — What are your main results?
08 — Why do you think your method works?
09 — What are the limitations?
10 — What is the main takeaway?

Part VIII · Practice Seminar

실전 세미나 — Agentic Reasoning using Hyper-Relational Knowledge Graphs

아래 대본은 12~15분 정도의 영어 기술 세미나 연습용 가상 발표다. 12장의 슬라이드로 구성되어 있으며, 각 슬라이드마다 스피킹 포커스를 붙였다.

중요. 이 대본은 특정 논문의 실제 실험 결과를 재현한 발표가 아니다. 결과 슬라이드의 비교와 관찰은 영어 스피킹 연습용 시나리오다. 기술적 배경은 hyper-relational KG, StarE, ReAct, KG-Agent 등의 연구 흐름을 참고한 일반화된 구성이다.

SLIDE 01Opening — Start with a Question

Good afternoon, everyone.

Let me start with a simple question.

Can an AI agent reason reliably when a fact is true only under certain conditions?

Consider a simple example. Suppose a knowledge base tells us that a particular drug treats a particular disease. That sounds like a useful fact. But what if the treatment is valid only for a certain patient population? What if the evidence comes from a specific clinical trial? And what if that evidence is valid only during a certain time period?

The key challenge is that real-world knowledge is rarely just a collection of simple facts. Facts usually come with context, conditions, evidence, time, and provenance.

The main message of this talk is simple. If we want AI agents to reason reliably over structured knowledge, we need to reason not only about facts, but also about the context surrounding those facts. And this is where hyper-relational knowledge graphs become particularly useful.

Speaking Focus — 끊어읽기와 강세

~~Let me start with~~/a simple question.

~~The key challenge is~~/that real-world knowledge/is rarely/just a collection of ~~simple facts~~.

~~The main message~~/of this talk/is ~~simple~~.

SLIDE 02The Problem — Why Ordinary Triples Are Not Enough

Let me first motivate the problem. A conventional knowledge graph typically represents knowledge as triples: subject, relation, object. For example: Drug A — treats — Disease B. This representation is simple and powerful.

The problem, however, is that important contextual information can be lost. We may also want to know:

who reported the fact,
when the fact is valid,
which population it applies to,
what type of evidence supports it,
and under which conditions it holds.

Why does this matter? Because an agent may retrieve a technically correct fact and still reach the wrong conclusion if it ignores the conditions attached to that fact.

So the problem is not simply: "Can the agent retrieve the right fact?" The more important question is: "Can the agent retrieve the right fact under the right conditions?"

Speaking Focus

~~The problem, however, is that...~~

~~Why does this matter?~~

~~The more important question is...~~

SLIDE 03Hyper-Relational Knowledge Graphs

So, how can we represent this additional context? This brings us to hyper-relational knowledge graphs.

A hyper-relational knowledge graph extends an ordinary relation with additional qualifier information. Let me give you a simple, fictional example. Suppose we have the main fact:

Drug A — treats — Disease B

We may attach qualifiers such as:

population — adults
evidence type — randomized trial
valid after — 2022
source — Study X

The important point is that these qualifiers are not independent facts in this representation. They describe the conditions under which the main fact should be interpreted. In other words, we move from "A is related to B." to something closer to "A is related to B, under conditions C, D, and E."

The key thing to notice here is that the qualifiers change how we should interpret the relation. That additional structure can be valuable for reasoning.

Speaking Focus

~~This brings us to...~~·~~Let me give you a simple example.~~

~~The important point is that...~~·~~In other words,...~~

~~The key thing to notice here is...~~

SLIDE 04From Retrieval to Agentic Reasoning

Now let us move from representation to reasoning. A conventional retrieval system might receive a question, retrieve several facts, and pass them directly to a language model.

Our key idea is simple. Instead of treating knowledge retrieval as a one-shot operation, we treat it as an iterative reasoning process. The agent repeatedly asks:

What do I know?
What information am I missing?
Which relation should I explore next?
Which qualifiers should I inspect?
Is the evidence consistent with the question?
Do I have enough information to answer?

So the agent does not simply retrieve knowledge. It plans, acts on the graph, observes the result, updates its reasoning state, and then decides what to do next.

Plan → Act → Observe → Update → Verify

You can think of this as evidence-driven navigation through a structured knowledge space.

Speaking Focus

~~Our key idea is simple.~~

~~Instead of doing A, we do B.~~

~~You can think of this as...~~

SLIDE 05Proposed Architecture

At a high level, the system consists of five components. Let me walk you through the architecture.

COMPONENT 1Planner

Given a user question, the planner identifies the target entity, the required relations, important constraints, and the information that is still missing.

COMPONENT 2Graph Explorer

It performs actions over the knowledge graph, such as retrieving neighboring entities, relations, and candidate statements.

COMPONENT 3Qualifier-Aware Retriever

Its job is not only to retrieve relevant facts, but also to examine the qualifiers associated with those facts.

COMPONENT 4Evidence Memory

The retrieved evidence is stored here. This allows the agent to keep track of what it has already established and what still needs to be verified.

COMPONENT 5Verifier

It checks whether the available evidence is sufficient and consistent with the original question. If insufficient, the agent continues reasoning; if sufficient, the system generates the final answer.

The important thing to remember is that retrieval and reasoning are not separate stages. They interact with each other throughout the process.

SLIDE 06A Concrete Reasoning Example

Let me show you how this works with a simple example. Suppose the user asks:

"Which treatment is supported for Condition X in adult patients after 2022, and what evidence supports that conclusion?"

This is a fictional example, but it illustrates the reasoning process.

Understand the Constraints

The agent first identifies three important constraints: treatment, adult population, evidence after 2022. The key point here is that the query is not asking for any treatment. It is asking for a treatment satisfying several conditions simultaneously.

Retrieve Candidate Facts

The agent retrieves candidate treatment relations from the graph. Suppose it finds three possible treatments. At this stage, all three may appear relevant.

Inspect the Qualifiers

Treatment A may apply only to children. Treatment B may have evidence from 2018. Treatment C may apply to adults and have supporting evidence after 2022. Notice what happens here. All three facts may be valid in isolation. But only one satisfies the constraints of the question.

Follow the Evidence

The agent now follows the supporting evidence relations. It may inspect the study, the evidence type, the publication date, and the source.

Verify

Finally, the verifier asks: Does the evidence actually support the proposed answer? If yes, the agent stops. If not, it continues exploring the graph.

This example illustrates why qualifier-aware reasoning can be more informative than retrieving triples alone.

SLIDE 07What Makes the Reasoning "Agentic"?

At this point, you may ask: What exactly makes this approach agentic? That is an important question. The answer is adaptive decision making.

The system does not follow exactly the same retrieval path for every query. Instead, its next action depends on what it has already discovered. For example:

If temporal information is missing, the agent may search for a time qualifier.
If the evidence source is unclear, it may follow a provenance relation.
If two facts conflict, it may retrieve additional evidence.
If the evidence is sufficient, it stops.

In other words, the reasoning path is dynamically constructed from the current evidence state. This gives us a simple loop:

reason about the current state → choose an action → observe new evidence → revise the plan

The process continues until the stopping condition is satisfied.

SLIDE 08How Would We Evaluate the System?

Now that we understand the approach, let us look at how we would evaluate it. I would focus on three questions.

FIRSTAnswer Correctness

Does the system produce the correct answer? This is the most obvious metric. But it is not sufficient.

SECONDEvidence Faithfulness

Is the answer actually supported by the retrieved graph evidence? This is particularly important because a fluent answer is not necessarily a well-supported answer.

THIRDReasoning Efficiency

How much graph exploration is required before the agent reaches a reliable conclusion? We could measure the number of reasoning steps, the number of graph queries, retrieval cost, and end-to-end latency.

So our evaluation should consider not only "Did the system get the answer right?" but also "Did it reach that answer using relevant and consistent evidence?" and "How efficiently did it get there?"

SLIDE 09Experimental Results — Speaking Practice Version

Important. 이하의 결과 서술은 가상이며, 영어 발표 연습만을 위해 포함했다. StarE, ReAct, KG-Agent 또는 다른 인용 논문의 측정 결과로 보고된 것이 아니다.

Let me walk you through the main result. We compare four types of systems:

an LLM without graph access,
an LLM with ordinary triple-based retrieval,
an LLM with passive hyper-relational retrieval,
and our hypothetical agentic hyper-relational reasoning system.

The key takeaway from this figure is not simply that the agent produces more correct answers. More importantly, the advantage becomes larger on questions involving temporal constraints, provenance, multiple conditions, and conflicting evidence.

One possible explanation is that these questions require the system to reason about the context of a fact, rather than the fact alone.

Another interesting observation concerns evidence quality. The agentic system produces fewer answers that are inconsistent with the retrieved qualifiers. This suggests that qualifier-aware reasoning may help constrain the reasoning process to evidence that actually matches the question.

However, there is a trade-off. The iterative agent requires additional graph queries and therefore introduces additional computational cost. So the key takeaway is that agentic reasoning may provide a better reliability–efficiency trade-off when contextual constraints matter.

Result-Language Drill

~~The key takeaway from this figure is...~~·~~More importantly,...~~

~~One possible explanation is that...~~·~~This suggests that...~~

~~However, there is a trade-off.~~·~~So the key takeaway is...~~

SLIDE 10Why Might This Approach Work?

Now let me come back to the main question. Why might agentic reasoning over a hyper-relational graph be useful? I think there are three reasons.

FIRSTExplicit Context

The graph makes contextual information explicit. The agent does not have to infer every condition from unstructured text.

SECONDSelective Exploration

The agent can retrieve additional information only when it is needed. This makes reasoning adaptive rather than completely predetermined.

THIRDExplicit Evidence Tracking

The system maintains a structured record of the evidence used during reasoning.

Taken together, these properties provide a natural way to combine structured knowledge with adaptive decision making.

The important point is not that the graph replaces the language model. Nor does the language model replace the graph. The two components play complementary roles. The language model provides flexible reasoning and planning. The graph provides structured evidence and explicit constraints.

SLIDE 11Limitations

There are, however, several important limitations.

LIMITATION 1Knowledge Graph Coverage

The system cannot reason about information that is missing from the graph. A sophisticated agent cannot compensate for a fundamentally incomplete knowledge base.

LIMITATION 2Qualifier Quality

The approach depends on qualifiers being sufficiently complete and accurate. Incorrect metadata can lead to incorrect reasoning.

LIMITATION 3Reasoning Cost

Iterative graph exploration introduces additional latency and computational cost. This may become significant for complex multi-hop queries.

LIMITATION 4Conflict Resolution

Real knowledge graphs may contain conflicting statements. Determining which source to trust may itself require another reasoning mechanism.

LIMITATION 5Agent Errors

The agent itself can make poor planning decisions. It may explore an irrelevant relation, stop too early, or fail to request an important piece of evidence.

So we should not interpret agentic reasoning as a guarantee of correctness. Rather, it provides a framework for making the reasoning process more explicit, evidence-driven, and controllable.

SLIDE 12Conclusion

Let me wrap up. I would like to leave you with three main takeaways.

Real-world facts often have context. Time, provenance, evidence, and other qualifiers can change how a fact should be interpreted.
Hyper-relational knowledge graphs provide a structured way to represent that context.
Agentic reasoning allows a system to actively decide which facts and qualifiers it needs to inspect before reaching a conclusion.

So, if you remember one thing from this talk, it should be this:

Reliable reasoning requires more than finding the right fact.It requires finding the right fact, under the right conditions, supported by the right evidence.

Ultimately, the goal is not simply to build agents that produce answers. The goal is to build agents that can explain why their answers are supported by the available evidence.

Thank you. And with that, I'd be happy to take questions.

Part IX · Q&A Practice

세미나 Q&A 실전 훈련

발표의 진짜 승부는 질의응답에서 갈린다. 준비된 대본이 끝나는 순간부터가 실력이다. 아래 여덟 개의 질문과 모범 답변으로 즉답 훈련을 한다.

Why can't we just convert qualifiers into ordinary triples?

That's an important question.

The short answer is that we could represent some qualifier information using additional triples, but we may lose the direct semantic connection between the qualifier and the statement it qualifies. In our approach, we want to preserve that connection explicitly.

The main reason is that the agent needs to know not only that a qualifier exists, but also which particular statement it modifies. So the benefit is primarily semantic clarity during reasoning.

핵심 — hyper-relational KG가 필요한 이유는 qualifier와 그것이 수식하는 statement 사이의 의미적 연결을 명시적으로 보존하기 위해서다.

Why not retrieve all relevant information in a single step?

That's a fair question.

I think the main advantage of the agent is selective exploration. For simple questions, one retrieval step may be enough. But for more complex questions, the system may not know in advance which information will become necessary.

For example, it may discover a conflicting fact and only then realize that it needs provenance information. So instead of retrieving everything upfront, the agent retrieves additional evidence as the reasoning process unfolds.

핵심 — 에이전트를 쓰는 이유는 선택적 탐색이다. 필요가 드러나는 순간에 증거를 추가로 가져온다.

Isn't this just graph search?

There are certainly similarities.

The main difference is that the search objective is dynamically determined by the reasoning state. A conventional graph algorithm typically follows a predefined objective or traversal rule. Here, the agent interprets the question, evaluates intermediate evidence, and decides which operation should be performed next.

So graph search is one component of the system, but the agent decides when and why that search should occur.

핵심 — 그래프 탐색은 부품이고, 언제·왜 탐색할지를 결정하는 것은 에이전트다.

How do you prevent hallucination?

No, I would not claim that using a knowledge graph eliminates hallucination.

The graph can provide explicit evidence, but the language model may still interpret that evidence incorrectly. What we can do is constrain the answer generation process to retrieved evidence and verify whether the final claims are supported by that evidence.

So I would say that the goal is to reduce unsupported reasoning, rather than claim that hallucination has been completely eliminated.

핵심 — 환각의 '제거'가 아니라 '근거 없는 추론의 축소'가 목표라고 답한다. 주장의 강도를 조절하는 모범 사례다.

What happens when two facts conflict?

That's a very important issue. There are several ways to approach it.

The agent could compare provenance, timestamps, evidence strength, or source reliability. However, conflict resolution is itself a complex reasoning problem.

So in the current formulation, I would treat this as an important extension rather than a completely solved problem.

핵심 — 미해결 문제는 미해결이라고 말한다. '중요한 확장 과제'라는 표현으로 정직함과 전문성을 동시에 지킨다.

What is the main computational bottleneck?

The main bottleneck is likely to be iterative graph exploration. Each additional reasoning step may require another graph query and another model inference.

There is therefore a trade-off between reasoning depth and computational efficiency.

One possible direction is to use lightweight graph operations for simple decisions and invoke the language model only when semantic reasoning is necessary.

핵심 — 병목을 인정하고, trade-off의 언어로 답한 뒤, 완화 방향을 하나 제시한다.

Could a small language model work?

Potentially, yes.

One interesting hypothesis is that a structured graph could reduce the amount of knowledge the language model has to recover from its parameters. If the necessary evidence is explicitly available, a smaller model may be able to focus more on planning and reasoning.

But I would want to verify that experimentally before making a strong conclusion.

핵심 — 가설은 가설이라고 명시한다. "I would want to verify that experimentally"는 연구자의 방패이자 품격이다.

What would be your next step?

I see three immediate directions.

First, I would evaluate the approach on more complex multi-hop reasoning tasks. Second, I would study how qualifier completeness affects performance. And third, I would investigate adaptive stopping mechanisms so that the agent can balance reasoning quality and inference cost.

Ultimately, I think the most interesting question is how much reasoning should happen in the language model and how much should be delegated to the structured knowledge system.

핵심 — 향후 과제 질문에는 구체적 방향 셋과 큰 질문 하나로 답한다.

Part X · Seminar-Specific Frames

이 발표에서 먼저 자동화할 26개 표현

세미나 대본 전체에서 뼈대 역할을 하는 표현들이다. 이 26개가 자동화되면 대본의 절반은 이미 입에 붙은 셈이다.

01 Let me start with a simple question.
02 The key challenge is that...
03 The problem, however, is that...
04 Why does this matter?
05 This brings us to...
06 The key thing to notice here is...
07 Our key idea is simple.
08 Instead of A, we do B.
09 You can think of this as...
10 At a high level, the system consists of...
11 Let me walk you through how this works.
12 The important thing to remember is...
13 Notice what happens here.
14 In other words,...
15 The key takeaway from this figure is...
16 One possible explanation is that...
17 This suggests that...
18 However, there is a trade-off.
19 Taken together, these results suggest that...
20 There are, however, several important limitations.
21 That's a fair question.
22 The short answer is..., but...
23 There are two ways to look at this.
24 I would not make that claim.
25 I would want to verify that experimentally.
26 If you remember one thing from this talk, it should be this: ...

Part XI · Script to Speech

세미나 대본을 스피킹 능력으로 바꾸는 방법

대본을 갖는 것과 대본 없이 말하는 것 사이에는 세 번의 변환이 있다. 그대로 말하기, 다르게 말하기, 내 것으로 말하기.

401차 — 그대로 말하기

예를 들어 Slide 4를 선택한다. /마다 의미 단위로 끊고, 강조 지점은 key idea, one-shot operation, iterative reasoning process다.

Chunk Reading

Our ~~key idea~~ is simple. / Instead of treating knowledge retrieval / as a ~~one-shot operation~~, / we treat it / as an ~~iterative reasoning process~~.

412차 — 같은 논리를 다른 영어로 말하기

원문

Instead of treating knowledge retrieval as a one-shot operation, we treat it as an iterative reasoning process.

자기 영어 — The basic idea is quite simple. Instead of retrieving everything at once, the agent retrieves information step by step, depending on what it currently needs.

문장이 달라져도 논리와 의미가 유지되면 성공이다. 이 훈련의 목적은 재현이 아니라 재구성이다.

423차 — 자신의 연구로 Slot Substitution

Frame

~~Instead of A, we do B.~~

Instead of retrieving documents independently, we jointly reason over the graph.
Instead of ignoring qualifiers, we use them as reasoning constraints.
Instead of using a fixed retrieval depth, we let the agent decide when to continue.
Instead of generating the answer immediately, we verify the supporting evidence first.
Instead of treating provenance as metadata, we explicitly use it during reasoning.

이 단계에 이르면 sentence frame은 암기한 문장이 아니라 발표 중 즉석에서 꺼내 쓰는 언어 도구가 된다.

Part XII · Applied Formula

이 주제에 적용한 설득의 5문장

Part VI의 공식을 이 세미나 주제에 그대로 적용하면 다음과 같다. 이 다섯 문장을 한 호흡의 논리로 연결해 반복한다.

Claim

Our agent explicitly reasons over contextual qualifiers.

Evidence

As you can see, it checks temporal, provenance, and evidence constraints before producing the answer.

Interpretation

This suggests that the agent is reasoning about the conditions of a fact rather than the fact alone.

Qualification

However, this comes at the cost of additional graph exploration.

Takeaway

So the key takeaway is that qualifier-aware reasoning provides a more explicit reliability–efficiency trade-off.

Part XIII · Technical Background

기술적 배경과 검토 메모

세미나 대본의 기술적 서술이 어디까지 문헌에 근거하고 어디부터 교육용 추상화인지를 구분해 둔다. 발표 연습 자료라도 사실과 예시의 경계는 분명해야 한다.

43Hyper-Relational Knowledge Graph와 StarE

StarE 논문은 hyper-relational KG에서 main triple에 추가 key–value qualifier를 연결하여 사실을 명확히 하거나 유효 범위를 제한할 수 있는 표현을 다룬다. 또한 qualifier를 main triple의 의미적 역할과 구분하면서 함께 모델링하는 message-passing 기반 encoder를 제안한다.

따라서 이 문서의 다음 설명은 기술적으로 타당한 학습용 요약이다.

검증된 서술

A hyper-relational knowledge graph extends an ordinary relation with additional qualifier information.

다만 Drug A, Disease B, Study X 예시는 설명을 위한 가상 예시다.

SOURCEMessage Passing for Hyper-Relational Knowledge Graphs (StarE)

Mikhail Galkin et al., EMNLP 2020

arxiv.org/abs/2009.10847

44ReAct

ReAct는 language model이 reasoning과 task-specific action을 interleaved manner로 생성하도록 하여, reasoning이 action plan을 갱신하고 action이 외부 환경이나 knowledge source에서 추가 정보를 얻도록 하는 접근이다.

따라서 세미나 예시의 Plan → Act → Observe → Update → Verify 사고 흐름은 ReAct류 agent reasoning을 설명하기 위한 자연스러운 학습 프레임이다. 단, 이 정확한 5단어 시퀀스가 ReAct 논문의 공식 알고리즘 명칭이라는 뜻은 아니다. 스피킹 훈련을 위해 일반화한 설명 프레임이다.

SOURCEReAct: Synergizing Reasoning and Acting in Language Models

Shunyu Yao et al.

arxiv.org/abs/2210.03629

45KG-Agent

KG-Agent는 LLM, toolbox, KG-based executor, knowledge memory를 결합하고, agent가 도구를 선택하고 memory를 갱신하는 반복 과정을 통해 KG 위에서 복잡한 reasoning을 수행하도록 구성한다.

따라서 이 문서의 Planner / Graph Explorer / Evidence Memory / Verifier 구조는 KG-Agent를 그대로 복제한 것이 아니라, agentic KG reasoning을 영어로 설명하기 좋도록 일반화한 가상 아키텍처다.

SOURCEKG-Agent: An Efficient Autonomous Agent Framework for Complex Reasoning over Knowledge Graph

Jinhao Jiang et al.

arxiv.org/abs/2402.11163

46이전 답변을 통합하면서 명확히 한 사항

YouTube 추천 자료와 기술 논문 출처를 분리했다.
실제 논문 주장과 스피킹 연습용 가상 실험 서술을 분리했다.
Agentic Reasoning using Hyper-Relational Knowledge Graphs 세미나는 특정 단일 논문의 발표문이 아니라 학습 목적의 합성 예제임을 명확히 했다.
Hyper-relational qualifier가 사실의 맥락이나 유효 조건을 나타낼 수 있다는 설명은 StarE 논문의 정의와 부합하도록 다듬었다.
Agentic reasoning의 Plan → Act → Observe → Update → Verify는 스피킹 연습을 위한 교육적 추상화이며 특정 논문의 공식 모듈 이름으로 오해하지 않도록 구분했다.
Q&A에서는 과도한 주장 대신 may, could, suggest, in our setting, we would need to verify 같은 학술적 hedging을 적극 사용하도록 정리했다.

Part XIV · Reference URLs

참고 자료

YouTube — English Technical Speaking Models

Research Papers — Technical Background

STARE / HYPER-RELATIONAL KGMessage Passing for Hyper-Relational Knowledge Graphs

Mikhail Galkin, Priyansh Trivedi, Gaurav Maheshwari, Ricardo Usbeck, Jens Lehmann

arxiv.org/abs/2009.10847

REACTReAct: Synergizing Reasoning and Acting in Language Models

Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, Yuan Cao

arxiv.org/abs/2210.03629

KG-AGENTKG-Agent: An Efficient Autonomous Agent Framework for Complex Reasoning over Knowledge Graph

Jinhao Jiang, Kun Zhou, Wayne Xin Zhao, Yang Song, Chen Zhu, Hengshu Zhu, Ji-Rong Wen

arxiv.org/abs/2402.11163

Part XV · Final Checklist

최종 체크리스트

실제 영어 발표 직전에 아래 질문을 확인한다. 하나라도 자신이 없으면 그 지점이 그날의 연습 지점이다.

Opening

첫 30초 안에 문제나 질문을 제시했는가?
발표의 main message를 한 문장으로 말할 수 있는가?

Problem

The key challenge is that...로 문제를 한 문장에 설명할 수 있는가?
Why does this matter?에 20초 안에 답할 수 있는가?

Method

Our key idea is simple. 다음에 핵심 아이디어를 한 문장으로 말할 수 있는가?
Instead of A, we do B. 구조로 기존 방법과 차이를 설명할 수 있는가?
아키텍처를 60초 안에 설명할 수 있는가?

Results

그래프를 읽지 않고 The key takeaway from this figure is...로 결론부터 말하는가?
결과 뒤에 This suggests that...로 해석을 붙이는가?
실제 측정 결과와 가설·해석을 명확히 구분하는가?

Limitations

중요한 한계를 최소 2개 스스로 말하는가?
과도한 prove 대신 suggest, indicate, provide evidence를 사용하는가?

Q&A

질문을 못 들었을 때 자연스럽게 확인할 수 있는가?
모르는 질문에 과장하지 않고 답할 수 있는가?
비판적 질문에 That's a fair point.로 침착하게 시작할 수 있는가?

Final Takeaway

If you remember one thing from this talk, it should be this: 뒤에 한 문장으로 핵심 메시지를 말할 수 있는가?

Closing Principle

영어 기술발표 훈련의 최종 목표는 다음 변화다.

문장 암기 → 문장 변형 → 내 연구에 적용 → 슬라이드 없이 설명 → 예상하지 못한 Q&A에 답하기

Andrew Ng이나 Andrej Karpathy의 영어를 똑같이 복제하는 것이 목적은 아니다. 진짜 목표는 그들이 사용하는 설명의 논리 구조, 청중을 안내하는 표현, 결과를 해석하는 방식, 주장의 강도를 조절하는 방식을 자신의 발표 습관으로 만드는 것이다.