ChatGPT Evaluation for NLP

ChatGPT Evaluation for NLP
A Meta Survey
Xiachong Feng
Update: 2023.3.13
1

Agenda
● Introduction
● Social Media
● NLP Tasks
○ Summarization
○ Natural Language Generation Evaluation
○ Information Extraction
○ Machine Translation
○ Data Augmentation
● ChatGPT Failures
● Conclusion
P 2

Introduction
3
● Big-LLMs
● Instruction-tuning
● ChatGPT

Introduction: Instruction-Tuning
[FLAN] Finetuned Language Models Are Zero-Shot Learners
[T0] Multitask Prompted Training Enables Zero-Shot Task Generalization
[CoT] Chain of Thought Prompting Elicits Reasoning in Large Language Models
[InstructGPT] Training language models to follow instructions with human feedback
[SUPER-NATURALINSTRUCTIONS] Super-NaturalInstructions: Generalization via
Declarative Instructions on 1600+ NLP Tasks
[Zero-shot CoT] Large Language Models are Zero-Shot Reasoners
[FLAN-T5] Scaling Instruction-Finetuned Language Models
[mT0] Crosslingual Generalization through Multitask Finetuning
[🔥🔥🔥ChatGPT] Introducing ChatGPT
[LLaMA] LLaMA: Open and Efficient Foundation Language Models
2021.09.03
2021.10.15
2022.01.28
2022.03.04
2022.04.16
2022.05.24
2022.10.20
2022.11.30
2023.02.27
P 5

https://yaofu.notion.site/How-does-GPT-Obtain-its-Ability-Traci
ng-Emergent-Abilities-of-Language-Models-to-their-Sources-b
9a57ac0fcf74f30a1ab9e3e36fa1dc1
Introduction: ChatGPT
P 6

Social Media: Twitter
7
● ChatGPT: A Meta-Analysis after 2.5 Months 2023-2-20

ChatGPT: A Meta-Analysis after 2.5 Months
Social Media Analysis
● Aim: acquire insights into public opinion and sentiment on ChatGPT and understand public
attitudes toward different topics related to ChatGPT.
● The dataset contains tweets across 61 languages. Over 68% of them are in English, other
major languages are Japanese (6.4%), Spanish (5.3%), French (5.0%), and German (3.3%).
#ChatGPT English
Tweets
Sentiment
Analysis
there is a relatively large proportion of positive sentiment, with
100k instances, and a smaller but still notable number of
tweets of negative sentiments, with 60k instances. P 8

Observe an overall downward trend of
sentiment (black solid line) during the
course of ChatGPT’s first 2.5 months:
an initial rise in average sentiment was
followed by a decrease from January
2023 onwards.
Time
Sentiment
Score
P 9

Observe an overall downward trend of
sentiment (black solid line) during the
course of ChatGPT’s first 2.5 months:
an initial rise in average sentiment was
followed by a decrease from January
2023 onwards.
Overall tweets in English have a more
positive perception of ChatGPT. This
suggests that ChatGPT may be better
in English, which constituted the
majority of its training data; but see
also our topic-based analysis below.
Time
Sentiment
Score
P 10

While the percentage of negative
tweets is stable over time, the
percentage of positive tweets
decreases and there is a clear increase
in tweets with the neutral sentiment.
This may indicate that the public view
of ChatGPT is becoming more rational
after an initial hype of this new
“seemingly omnipotent” bot
the
count
of
tweets
per
week
Time
P 11

During the course of 2.5 months after
ChatGPT’s debut, OpenAI announced 5
new releases claiming various updates.
Our data covers the period of the first
three releases on the 15th of
December 2022, the 9th of January,
and the 3rd of January in 2023.
the
count
of
tweets
per
week
Time
P 12

Sentiment across Language
Tweets in English have the most positive
view of ChatGPT
The sentiment of English, German, and
French tweets are trending downward while
Spanish and Japanese tweets start from a
low point and trend upwards.
P 13

Sentiment across Topic
5 major classes, which cover 86.3% of
tweets in our dataset: science & technology
(38.6%), learning & educational (15.2%), news
& social concern (13.0%), diaries & daily life
(10.2%), and business & entrepreneurs
(9.3%).
The share of science & technology topic
ranks the highest in all of the 5 languages.
P 14

Sentiment across Topic
business & entrepreneurs has the lowest
proportion of negative tweets while the
topic news & social concern contains the
highest proportion of negative tweets.
P 15

Sentiment: Human Evaluation
We manually annotated and analyzed the sentiment expressed within 40 randomly selected
tweets.
● 20 random positive tweets from the period including the last two weeks of 2022 and the
first week of 2023, where the general sentiment reaches the peak
● 20 random negative tweets from the second week to the fourth week of 2023, where the
general sentiment declines
For the first period [Positive]:
ChatGPT’s ability to generate human-like and concise text.
For the second period [Negative]
Potential factual inaccuracies, the detectability of the model-generated text, ethical
concerns, biased output or the potential increase in misinformation, job loss
P 16

Summarization
17
● Cross-Lingual Summarization via ChatGPT 2023-2-28
● Exploring the Limits of ChatGPT for Query or Aspect-based Text
Summarization 2023-2-16

Cross-Lingual Summarization via ChatGPT
Cross-lingual Summarization: Prompts
Experimental Prompts for CLS
E2E Please summarize the following text in Chinese: [English Doc]
E2E+Interact + Please make the Chinese summary shorter
Trans-Sum
Please first translate the following text to Chinese and then
summarize the translated text in Chinese: [English Doc]
Trans-Sum+Interact + Please make the Chinese summary shorter
Sum-Trans
Please first summarize the following text and then translate the
summary to Chinese: [English Doc]
Sum-Trans+Interact + Please make the Chinese summary shorter
P 18

Cross-lingual Summarization: Examples
P 19

Cross-lingual Summarization: Datasets
We randomly sample 50 documents from the
test set of each CLS dataset for evaluation.
P 20

Cross-lingual Summarization: Results
Interactive prompt is very important
mBART-50 Is still very strong
(needs more carefully-designed human evaluation)
P 21

Exploring the Limits of ChatGPT for Query or Aspect-based Text Summarization
Query-based Summarization: Dataset
P 22

Query-based Summarization: Prompts
P 23
SQuALITY
Q: Query. Answer the question in around 200 words.
Article: story. specific question
Q: Query. Answer the question in around 450 words.
Article: story.
Your response is too short. Please answer it in around
450 words. general question
QMSum
Q: Query. Article: meeting
Q: Query. Article: golden meeting
meeting is the initial meeting, while golden meeting
is the provided golden spans
CovidET
Q: Summarize this article with respect to Aspect
within one short sentence. Article0. A: Answer0. Q:
Summarize this article with respect to Aspect within
one short sentence. Article. A:
NEWTS
Article. Summarize this article with respect to Aspect:

Query-based Summarization: Results
P 24
Fine-tuning methods are still strong.

Query-based Summarization: Results
P 25
ChatGPT is smart but not that smart.

NLG Evaluator
26
● Is ChatGPT a Good NLG Evaluator? A Preliminary Study 2023-3-7

Is ChatGPT a Good NLG Evaluator? A Preliminary Study
NLG Evaluator: Prompt
Is ChatGPT a good NLG evaluator?
I would give this news summary four stars for fluency.
The summary is well-written and captures the main points
of the news article P 27
ChatGPT

NLG Evaluator: Metrics
● Spearman correlation (Zar, 2005) assesses the monotonic relationships between two
variables;
● Pearman correlation (Mukaka, 2012) measures the linear relationships between two sets
of data;
● Kendall’s Tau (Kendall, 1938) evaluates the ordinal association between two measured
quantities.
The generated text of m-th model for the i-th condition text is denoted as g_i,m.
Conditioned text {c1, c2, ..., cn} and M NLG models
P 28

NLG Evaluator: Results (Summarization for Example)
SummEval collects 16 model-generated
summaries on the CNN/DM dataset and
annotates human judgments upon these
summaries covering aspects of coherence,
relevance, consistency and fluency.
ChatGPT achieves state-of-the-art
or competitive correlation with
golden human judgments.
P 29

Machine Translation
30
● Is ChatGPT A Good Translator? A Preliminary Study 2023-1-31
● Towards Making the Most of ChatGPT for Machine Translation 2023-3-11

Is ChatGPT A Good Translator? A Preliminary Study
Machine Translation: Prompts
P 31
Ask ChatGPT for providing prompts.

Machine Translation: Baseline Results
P 32
● Chinese-to-English (Zh⇒En) translation task with the test set from Flores-101
● We randomly sample 50 sentences from each set for evaluation.
Still lags behind the baselines by
at least 5.0 BLEU points
Follow experiments are
based on this prompt.

Machine Translation: Multilingual Translation Results
P 33
● ChatGPT in multilingual
translation, including German
(De), English (En), Romanian (Ro),
and Chinese (Zh)
● The first three languages come
from the same family with Latin
scripts while the last is from
another family with Chinese
scripts
● We randomly sample 50
sentences from each set for
evaluation.
Relative to Google Translate
High-resource task Low-resource task
The huge resource difference of mono-lingual data between English
and Romanian limits the language modeling capability of Romanian,
which partially explains the poor performance on English⇒Romanian.
Romanian Language Modeling is poor!
Better knowledge transfer
within the same family than
between different families .

Machine Translation: Pivot Prompting
Ask ChatGPT to translate the source sentence into a high-resource pivot language (i.e., English
by default) first and then into the target language.
P 34
Prompt

Machine Translation: Translation Robustness
● To test the translation robustness, we adopt the test set of WMT19 Biomedical Translation Task (Bawden et
al., 2019, i.e., Bio) and the set2 and set3 of WMT20 Robustness Task.
● We randomly sample 50 sentences from each set for evaluation.
○ WMT19 Bio test set is composed of Medline abstracts, which require domain-specific knowledge to handle the
terminologies.
○ WMT20 Rob2 are comments from the social media website reddit.com that could contain various errors, including
spelling/typographical errors, word omission/insertion/repetition, grammatical errors, spoken languages, Internet slang,
and so on.
P 35
WMT20 Rob3 test set that contains a
crowdsourced speech recognition
corpus. It suggests that ChatGPT, which is
essentially an artificial intelligent chatting
machine, is capable of generating more
natural spoken languages than these
commercial translation systems.

Towards Making the Most of ChatGPT for Machine Translation
Weakness and Motivation
Previous: Adopt simple prompts and basic settings regardless of the significant influence of the
prompts’ quality.
This paper: In this report, we aim to further elicit the capability of ChatGPT by revisiting the
following three aspects and correspondingly propose two simple but effective prompts:
Task-Specific Prompts (TSP) and Domain-Specific Prompts (DSP).
P 36
Temperature
Decoding with higher temperatures
displays greater linguistic variety, a
diverse generation may impede its
translation quality.
Task Information
The task inconsistency (ChatGPT is a
conversational system) will limit its
translation ability to a certain
degree. In response to this problem,
we proposed Task-Specific
Prompts (TSP) to further emphasize
the task information to bridge the
task gap, i.e., conversation and
translation.
Domain Information
ChatGPT can incorporate additional
information, like human interactions,
through the input prompts. We
argue that such flexible interaction
may alleviate some classical MT
challenges, e.g., crossdomain
generalization. Therefore, propose
Domain-Specific Prompts (DSP) to
introduce the domain navigation
information to elicit ChatGPT’s
generalization ability across
different domains.

Machine Translation: Experimental Setting and Datasets
ChatGPT: gpt3.5-turbo-0301 models
Flores-200 for Multilingual MT
WMT19 for Cross-domain MT
We test all samples through OpenAI API.
P 37

Machine Translation: Effect of Temperature
P 38
ChatGPT’s performance largely depends on the
temperatures, especially in difficult languages.
Generally, setting a lower temperature can
result in higher performance.
The impact of temperature is relatively small
when translating to high-resource languages,
while for complex languages, e.g., Chinese, it
has a large degradation in performance

Machine Translation: Effect of Task Information
P 39
TSP: Prepend the sentence "You are a machine translation system." to
"Please provide the [TGT] translation for the following sentence:"
High-resource task
Low-resource task
Distant language
Non-English-Centric Language Pairs
When tackling non English-centric MT
language pairs, ChatGPT tends to
generate hallucinations.
Emphasizing the task information
in prompts can further improve
ChatGPT’s performance, especially
in complex tasks.

Machine Translation: Effect of Domain Information
P 40
Introducing the correct domain information
consistently improves ChatGPT’s performance while
wrong domain information leads to significant
degradation in performance.

Machine Translation: In-context Learning
P 41
Few-shot ICL can further improve ChatGPT’s performance.

Machine Translation: CoT
P 42
The CoT prompt leads to word-by-word translation
behavior, which is the main reason for the significant
translation degradation.

Information Extraction
43
● Zero-Shot Information Extraction via Chatting with ChatGPT 2023-02-20
● Exploring the Feasibility of ChatGPT for Event Extraction 2023-03-07

Zero-Shot Information Extraction via Chatting with ChatGPT
Information Extraction: ChatIE
P 44
Find out the existing types of entities, relations, or
events in the sentence respectively in three tasks.
NER: each turn
aims to extract
the entities of
one type.
① Event Classification
②
Argument
extraction
Entity-Relation
Triple Extraction

Zero-Shot Information Extraction via Chatting with ChatGPT
Information Extraction: Results
P 45

Exploring the Feasibility of ChatGPT for Event Extraction
Event Extraction
P 46

Event Extraction: Results
P 47
● Dataset: ACE 2005 corpus
● We randomly select 20 samples from the raw test set to evaluate the efficacy of ChatGPT.
Our results show that ChatGPT has, on average, only
51.04% of the performance of a task-specific model
such as EEQA in long-tail and complex scenarios.
The model’s performance improved after
eliminating the negative sample.

Event Extraction: Usability of ChatGPT
Recruited four professional and well-educated
annotators (e.g. postgraduate student on NLP
research) to evaluate ChatGPT’s usability.
We randomly selected ten samples from the
ACE05 test set and provided each annotator with
five attempts to create a task prompt that would
enable ChatGPT to extract structured events from
the given text.
P 48
ChatGPT is not robust enough.
ChatGPT’s performance is sensitive
to different prompt styles

Data Augmentation
49
● ChatAug: Leveraging ChatGPT for Text Data Augmentation
2023-2-28

ChatAug: Leveraging ChatGPT for Text Data Augmentation
ChatAug Framework
P 50
ChatGPT is applied to rephrase each input sentence into six additional sentences
Acne
Cough

ChatGPT Failures
51
● A Categorical Archive of ChatGPT Failures 2023-3-6

A Categorical Archive of ChatGPT Failures
Failure 1: Reasoning
● Critical thinking, decision making, and problem solving are all crucial activities that rely
heavily on the fundamental aspect of human intelligence known as reasoning.
● Models like ChatGPT lack a “world model”, meaning they do not possess a complete
understanding of the physical and social world, or the capability to reason about the
connections between concepts and entities. They can only generate text based on the
patterns they have learned during training.
P 52

Failure 1.1: Spatial Reasoning
Spatial reasoning refers to the ability to understand and manipulate the relationships between
objects, people, and places in the physical space around us.
P 53
ChatGPT does possess some level of spatial understanding, as evidenced
by its ability to translate the relative positions of grid boxes into language.

Failure 1.2: Physical reasoning
Physical reasoning refers to the ability to understand and manipulate physical objects and their
interactions in the real world.
It involves the application of physical laws and concepts to predict and explain the behavior of
physical systems.
P 54
The trophy didn’t fit in the suitcase because it was too small.
What was too small?
✅ ChatGPT (Jan 30, 2023)
❌ Older version

Failure 1.3: Temporal reasoning
Temporal reasoning is the ability to reason about and make predictions about events and their
ordering in time.
It involves understanding the temporal relationships between events, the duration of events, and
the timing of events relative to each other.
P 55

Failure 1.4: Psychological reasoning
Psychological reasoning refers to the ability to understand and make predictions about human
behavior and mental processes (a.k.a Theory of Mind).
It involves the application of psychological theories, models, and concepts to explain and predict
human behavior and mental states.
P 56

Failure 2: Logic
Logic is a branch of mathematics and philosophy that studies the principles of reasoning. It deals
with the rules and methods for correct reasoning, such as syllogisms, induction, and deduction.
P 57
Mike’s mum had 4 kids; 3 of them are Luis, Drake, and Matilda. What is the name of the 4th kid?
ChatGPT
It is not possible to determine the name of the fourth child without more information

Failure 3: Math and Arithmetic
Arithmetic reasoning refers to the capability of utilizing mathematical concepts and logic to
solve arithmetic problems.
It requires logical thinking and the application of mathematical principles to find the right solution
to mathematical problems.
ChatGPT is limited in its capability to calculate mathematical expressions. Like most large
language models, it struggles with tasks such as multiplying large numbers, finding roots,
computing powers (especially with fractions), and adding or subtracting from irrational numbers
(e.g. pi or e)
P 58
Please refer to the paper for more examples

Failure 4: Factual Errors
Factual errors refer to inaccuracies in information or statements that are not in accordance with
reality or the truth.
Factual errors are often unintentional but can result in incorrect or misleading information.
● ChatGPT’s output lacks accuracy in regards to scientific facts.
● It sometimes lacks knowledge of basic facts, which can be quickly obtained through a
Google search.
P 59

Failure 5: Bias and Discrimination
Bias in a language model refers to the systematic inaccuracies or stereotypes in the generated
language output, which are influenced by the training data and reflect the societal and cultural
prejudices that exist in that data.
P 60

Failure 6: Wit and Humor
Humor is the quality of being amusing or comical, often expressed through words or actions that
entertain or make someone laugh.
P 61

Failure 7: Coding
P 62

Failure 8: Syntactic Structure, Spelling, and Grammar
Syntactic structure refers to the arrangement of words, phrases, and clauses in a sentence to
form a well-defined and meaningful structure according to the rules of a particular language.
P 63

Failure 9: Self Awareness
Self-awareness is the capacity to recognize oneself as an individual separate from others and to
have an understanding of one’s own thoughts, feelings, personality, and identity.
Self-awareness is considered an important aspect of consciousness and is closely related to
self-consciousness and introspection
P 64

Failure 10: Other Failures
1. ChatGPT’s difficulty in using idioms
2. ChatGPT lacks real emotions and thoughts
3. ChatGPT condenses the subject matter, but does not provide a distinctive perspective on
it.
4. ChatGPT tends to be excessively comprehensive and verbose
5. ChatGPT lacks human-like divergences and tends to be overly literal, leading to misses in
some cases
6. ChatGPT strives to maintain a neutral stance
7. ChatGPT’s responses tend to be formal in nature due to its programming to avoid informal
language.
8. If ChatGPT is informed that its answer is incorrect, it may respond by apologizing,
acknowledging its potential inaccuracies or confusion, correcting its answer, or maintaining
its original response.
P 65

Takeaways
● ChatGPT updates frequently, all the current conclusions may be changed after updates.
● Due to the lack of API early on, small tests are not enough to reveal the underlying pattern.
● Prompt engineering is of vital importance.
○ Different prompts or instructions can lead to contrasting results.
○ ChatGPT is highly sensitive to different prompt styles.
● ChatGPT is super strong, but we NLPers don't need to be afraid of it.
○ Full-shot fine-tuned models can still get better results.
P 67

References
[1] ChatGPT: A Meta-Analysis after 2.5 Months
[2] Exploring the Limits of ChatGPT for Query or Aspect-based Text Summarization
[3] Cross-Lingual Summarization via ChatGPT
[4] Is ChatGPT a Good NLG Evaluator? A Preliminary Study
[5] Is ChatGPT A Good Translator? A Preliminary Study
[6] Towards Making the Most of ChatGPT for Machine Translation
[7] Zero-Shot Information Extraction via Chatting with ChatGPT
[8] Exploring the Feasibility of ChatGPT for Event Extraction
[9] ChatAug: Leveraging ChatGPT for Text Data Augmentation
[10] A Categorical Archive of ChatGPT Failures
[11] [TBD] A Multitask, Multilingual, Multimodal Evaluation of ChatGPT on Reasoning, Hallucination, and
Interactivity
[12] [TBD] Is ChatGPT a General-Purpose Natural Language Processing Task Solver?
P 68

ChatGPT Evaluation for NLP

Recommended

More Related Content

What's hot (20)

Similar to ChatGPT Evaluation for NLP (20)

Recently uploaded (20)

ChatGPT Evaluation for NLP