NOTE: The models trained as part of this work are available on HuggingFace and can be queried live on this HuggingFace Space
Goal of this project is to train LLMs (Large Language Models) based on the transformer architecture to behave as a "sarcasm defuser". The sarcasm defuser's job is to take as input a sarcastic comment and convert it into a "neutral" comment with the same intended meaning but stripped of all sarcastic connotation.
- ORIGINAL COMMENT: "Because everything wrong with the community is the fault of liberals."
- NEUTRAL COMMENT: "It seems like many problems within the community are attributed to liberals."
The original comment decries sarcastically how many problems in the community are ascribed to liberals. The neutral comment conveys the same meaning but with a less sarcastic tone.
In the rest of this document, we will:
- Briefly introduce the transformer architecture and LLMs.
- Describe the dataset that we use to train our LLMs to defuse sarcasm.
- Describe the methodology we use to quantitatively evaluate the trained model
- Describe how the models were trained and evaluated
- Present the results of the evaluation
- Finally, we'll outline the shortcomings of this analysis as well as possible ideas for future work
The Transformer architecture is the engine behind almost every modern AI you interact with today, from ChatGPT to Gemini.
It was introduced by Google researchers in the seminal 2017 paper: "Attention is All You Need."
Before Transformers, AI processed text like a human reads a book: one word at a time, from left to right (using RNNs or LSTMs). This was slow and "forgot" the beginning of long sentences. Transformers changed the game for two reasons:
Parallelization: They process entire sentences all at once rather than sequentially. This makes them incredibly fast to train on massive datasets using GPUs.
Attention: This mechanism allows the model to weigh the importance of every word in a sentence simultaneously. It "pays attention" to the relevant context, no matter how far apart the words are. If you're interested in the details of how the transformer's architecture works, we defer you to the original "Attention is All You Need." paper.
The original Transformer consists of two main parts:
-
The Encoder: Reads and understands the input text, turning it into a mathematical representation.
-
The Decoder: Uses that representation to generate an output (like a translation or a response).
Depending on which of the 2 above components are present, a transformer could be encoder-only, decoder-only, or encoder-decoder:
| Feature | Encoder-Only | Decoder-Only | Encoder-Decoder |
|---|---|---|---|
| Core Strength | Understanding & Context | Generation & Prediction | Translation & Transformation |
| Mechanism | Reads the whole sequence at once (Bi-directional). | Predicts the next word based on previous ones (Auto-regressive). | Maps an input sequence to a new output sequence. |
| Common Use Case | Sentiment analysis, NER, Classification. | Chatbots, Storytelling, Code generation. | Translation, Summarization, Speech-to-text. |
| Popular Models | BERT, RoBERTa, ELECTRA | GPT-4, Llama 3, Gemini, Claude | T5, BART, Original Transformer |
In this project, we're gonna work with both GPT (decoder-only) and BART (encoder-decoder) to try to turn them into effective sarcasm defusers
In order to train our defuser, we need examples of sarcastic comments together with their "neutral" versions.
Our sarcastic comments will come from the Sarcasm on Reddit dataset. The dataset contains around 1M sarcastic comments from the Reddit online forum, which have been marked as sarcastic with the \s tag by the very authors.
The sarcastic comments alone aren't enough to train our sarcasm defuser, because we also need their neutral or "defused" versions i.e. for each sarcastic comment, we would like to have a comment with the same meaning but with as much of the sarcasm removed as possible. Unfortunately, the Sarcasm on Reddit dataset doesn't help us here.
In order to create the neutral versions of the sarcastic comments, we resort to an LLM, namely Google's Gemma 3 with 12B parameter. We chose this model because it can be queried with a decent request rate via the free plan of Google's GenAI API. The idea is to ask Gemma to create a neutral version of some of our sarcastic comments so that we have paris of sarcastic/neutral comment we can use to train our models.
After playing around with a few prompts, we identify the following to be quite effective:
given this sarcastic comment: <SARCASTIC_COMMENT>, which is a response to this other comment: <CONTEXT>,
remove all the sarcasm from it while keeping the original meaning. Don't output anything else, and don't try to describe the comment in the third person
here, <SARCASTIC_COMMENT> is the original sarcastic comment we want Gemma to neutralize while <CONTEXT> is the comment that was preceding the sarcastic comment in the Reddit thread. This comment is available from the Kaggle dataset and we believe it gives Gemma additional information to do a better job as neutralizing the sarcastic comment while keeping the same meaning. We sample a bit less than 5000 comments from the dataset and query Gemma via the API using the prompt to obtain the neutralized versions. Here's a couple of examples:
| orignal sarcastic comment | neutral comment generated by Gemma |
|---|---|
| B/c the DC narrative reflects the ideals of the majority of Americans amirite? | Considering the DC narrative doesn't necessarily align with the views of most Americans, do you agree? |
| He meant "tossback" since he can't throw far enough, | He meant "throwback" because he doesn't throw very far. |
| Rich Piana's Real Food helped me to get bigger by the day | Rich Piana's Real Food helped me gain size. |
Of this ~5000 comments, we sample 90% to be our train set and 10% to be our test set, resulting in a training and test set of 4448 and 495 comments, respectively.
Once we train a defuser model on the pairs of sarcastic/neutral comments, how can we quantitavely assess its ability at "defusing" sarcasm? Remember, we'd like the model to generate comments that are (1) non-sarcastic and (2) have the same meaning as the original one.
We would like therefore to assign to each comment generated by our defuser:
- a semantic textual similarity score between the defused comment and the sarcastic one
- a sarcasm score
To compute 1 and 2, we will use some similarity and sarcasm models available via Hugging Face and Sentence Transformers.
Here's a summary of the models we tested:
| Model | Description |
|---|---|
| helinivan/english-sarcasm-detector | English Sarcasm Detector is a text classification model built to detect sarcasm from news article titles. It is fine-tuned on bert-base-uncased |
| helinivan/multilingual-sarcasm-detector | Multilingual Sarcasm Detector is a text classification model built to detect sarcasm from news article titles. It is fine-tuned on bert-base-multilingual-uncased |
| pn89348/sarcasm_model | Fine-tuned version of distilbert/distilbert-base-uncased |
| Model | Description |
|---|---|
| all-MiniLM-L12-v2 | All-round model tuned for many use-cases. Trained on a large and diverse dataset of over 1 billion training pairs. |
| all-MiniLM-L6-v2 | All-round model tuned for many use-cases. Trained on a large and diverse dataset of over 1 billion training pairs. |
| all-distilroberta-v1 | All-round model tuned for many use-cases. Trained on a large and diverse dataset of over 1 billion training pairs. |
| all-mpnet-base-v2 | All-round model tuned for many use-cases. Trained on a large and diverse dataset of over 1 billion training pairs. |
| multi-qa-MiniLM-L6-cos-v1 | This model was tuned for semantic search: Given a query/question, it can find relevant passages. It was trained on a large and diverse set of (question, answer) pairs. |
| multi-qa-distilbert-cos-v1 | This model was tuned for semantic search: Given a query/question, it can find relevant passages. It was trained on a large and diverse set of (question, answer) pairs. |
| multi-qa-mpnet-base-dot-v1 | This model was tuned for semantic search: Given a query/question, it can find relevant passages. It was trained on a large and diverse set of (question, answer) pairs. |
| paraphrase-MiniLM-L3-v2 | This is a sentence-transformers model: It maps sentences & paragraphs to a 384 dimensional dense vector space and can be used for tasks like clustering or semantic search. |
| paraphrase-albert-small-v2 | This is a sentence-transformers model: It maps sentences & paragraphs to a 768 dimensional dense vector space and can be used for tasks like clustering or semantic search. |
| paraphrase-multilingual-mpnet-base-v2 | This is a sentence-transformers model: It maps sentences & paragraphs to a 768 dimensional dense vector space and can be used for tasks like clustering or semantic search. |
These models can be instantiated and queried using the Sentence Transformers library. They assign each pair of sentences a semantic similarity score. Different scoring metrics can be used. We choose to use cosine similarty because it's bounded in [-1, 1] making it easier to compare across different models.
Out of the models listed above, we want to pick the 2 best ones (one for sarcasm, one for semantic similarity) that we will use to evaluate our trained defusing models.
In order to do so, we compute:
- similarity between each pair of comments (original/neutral) in the training dataset
- sarcasm score of each original comments in the training dataset
- sarcasm score of each neutral comments in the training dataset
For all of the above, we compute median, mean, and quartiles. We also compute the ratio of the mean (median) sarcasm score of the neutral comment to the mean (median) of the sarcasm score of the original comments.
Here's the results for all models:
All boxplots show median (red bar) 1st and 3rd quartiles (edges of the box), whiskers computed from the inter-quartile range, and the mean as a green triangle. Both similarity and sarcasm scores of the original comments are not as high as we had expected.
However, for the purpose of this analysis, we select multi-qa-mpnet-base-dot-v1 as a semantic similarity model because it provides the highest similarity between comment pairs,
and we select helinivan/english-sarcasm-detector as the sarcasm scoring model because it provides the best separation between neutral and original comments, with the lowest ratio of median score of neutral generated comments to original ones.
Therefore, in all subsequent analysis, we will use these 2 models to measure semantic similarity and sarcasm, respectively. We will evaluate the goodness of our trained models based on how they perform compared to the baseline provided by gemma-3-12b-it. Our goal is for our models to produce neutral comments with a similarity score at least as high as the ones produced by Gemma, and with a sarcasm score at least as low.
We will train the following models:
| Model | Type | Parameters |
|---|---|---|
| gpt2-small | decoder-only | 0.1B |
| gpt2-medium | decoder-only | 0.4B |
| bart-base | encoder-decoder | 0.1B |
Using the Huggingface trainer library. The library is built on PyTorch and provides high level tools to batch, tokenize, and train our transformer models.
These models are pretrained on a large corpus of English-language documents. GPT2 is a casual model trained to predict the next work in a sentence, whereas BART is a denoising encoder-decoder (seq2seq) model trained to reconstruct the original input sentence after some of its text has been randomly corrupted.
The models process inputs by encoding them into sequences of numerical tokens, and convert the output token sequences into words. For all models, the vocabulary size (number of different tokens) is around 50257 tokens.
We will fine-tune all these models by training them on our train dataset with ~4500 sentence pairs for 10 epochs. We will feed batches of sarcastic sentences into the models and train them to produce the "defused" neutral sentence as output. We use cross-entropy loss as the loss function during training.
All training and evaluation was performed on machines rented via the VastAI platform.
However, some details of the training will be different depending on whether we're dealing with a decoder-only or with an encoder-decoder model.
Decoder-only models are causal: given an input sentence or "prompt" they are trained to generate the N words (or tokens) that should follow the prompt. In order to train them to defuse sarcasm, we construct the following prompt:
- Training Prompt:
SARCASTIC_COMMENT<|BOS|>NEUTRAL_COMMENT<|endoftext|>
Where <|BOS|> and <|endoftext|> are special tokens marking the beginning and the end of the neutral comment, respectively.
During the training, the input will be the tokenized version of the above prompt. The training label will be the same prompt (which gets shifted internally by one token to the right), as the model has to learn
how a neutral comment should causally "follow" a sarcastic one. We make sure that the model focuses all its computing power
on learning how to produce the neutral comment by computing the cross-entropy loss only on the NEUTRAL_COMMENT<|endoftext|> part. In other words, we don't penalize the model for not reconstructing the original SARCASTIC_COMMENT in its output. Including <|endoftext> is important: the model has to learn when to "shut up" and where the neutral comment should end. Otherwise, it might continue to "babble" forever (or until the maximum number of output tokens is reached) without ever producing an <|endoftext|> token.
Encoder-decoder models operate in a different way: they translate input sentences (which could use a different vocabulary) into output sentences. They are conceptually similar to translators, and seq2seq machine translations is actually one of their main use cases.
Training these models is more straightforward as we simply set the input to be SARCASTIC_COMMENT and the training label to be NEUTRAL_COMMENT.
After training, we evaluate the models by feeding the sarcastic sentences from our test dataset as input to the models and measuring similarity and sarcasm score of the output sentences using the multi-qa-mpnet-base-dot-v1 and helinivan/english-sarcasm-detector scoring models identified earlier.
There are quite a few parameters that control how these generative models generate tokens from an input, and we're going to test the following ones:
| Parameter | Description | Tested Values |
|---|---|---|
| max_new_tokens | Maximum number of tokens that the model will generate | 20, 50, 100 |
| greedy/sampling | Whether the model always chooses the next token having highest probability (greedy) or it samples from a distribution (sampling) | greedy,sampling |
| temperature | When sampling: smoothness or the probability distribution. The higher the temperature, the closer the distribution is to uniform | 1, 1.5, 0.5 |
We test the trained models with all possible combinations of these parameters (12 in total).
For all models identified above and the different generation parameters, we plot (1) the median similarity and sarcasm scores on a similarity/sarcasm plot and (2)
we identify the models and generation parameters that beat our Gemma baseline (i.e. similarity at least as high, sarcasm score at least as low).
In the evaluation we also include the "vanilla" base gpt2 and bart models (i.e. without our fine-tuning) to see how their performance as a sarcasm defuser was improved by our training.
The left-hand plot shows all models/configuration.
As expected, the "vanilla" base gpt2 models are unable to produce outputs with levels of similarity higher than the baseline. The sarcasm might be low, but without high semantic similarity to the original comment this is likely to be a result of the output being a random "babbling" which is typical for these base models. It's interesting to observe how the bart model, because it practically outputs the same sentences as the input's (it is trained to guess missing portions of the input sentences), achieves a similarity of 1.0 and a very high median sarcasm score (i.e. the same as the original sentence).
On the other hand, some of the configurations of our trained models are able to beat the baseline (higher similarity, lower sarcasm). The best of these configurations for each of the 3 models are shown on the right-hand plot.
Finally, for the 3 winning models/configurations, we plot boxplots of similarity/sarcasm score, as well as the neutral/original ratio of the median sarcasm scores:
All these 3 models achieve a very high reduction in sarcasm. However, bart-base-sarcasm-defuser stands out because it also achieves a very high similarity score, even though the sarcasm is a bit lower than the other 2 models.
It's interesting to note that there seem to be no considerable improvement between gpt2-sarcasm-defuser with 0.1B parameter to gpt2-medium-sarcasm-defuser with 0.4B parameters.
These are relative simple models trained on a relatively small dataset. We see the following areas for improvement:
- Increase the size of the training set by producing more sarcastic comments and their neutral version. In order to do so, we should even use more sophisticated models than
gemma(that will come at higher cost of course) - Use better methods to measure sarcasm and similarity. Especially the sarcasm score of the original comments is not very high to begin with, making the "sarcasm signal" low and probably unreliable. This suggests that the available models we used to measure sarcasm level aren't very good or at the very least need to be fine-tuned. More research on how to proper measure sarcasm would make the evaluation of our sarcasm defusing models more precise and give us more confidence in the results
- Catch subtle nuances. The sarcasm defuser works in some cases, but fails in many others. For example, the sarcastic comment:
Oh that's great, thank you very much!gets translated into:That's great, I appreciate you sharing that. It's clear that the above comment, when expressed sarcastically, means exactly the opposite of what it's saying, namely that the speaker isn't happy at all with the current situation. Our current defuser is only occasionally able to deal with this kind of comments and typically misses the nuance in the sarcastic expression. In order to improve on this, we could introduce a "meaning reversal" metric and measure if the meaning of the original comment has been "reversed" in the translated one. So to evaluate the translated comment we would use the following criteria:- Sarcasm score: has to be lower than the original comment's
- Semantic Similarity score: has to be high between original and sarcastic comment
- Meaning: has to be reversed from the original comment's
- Add context. Context is important to sarcasm and in fact we used the context (i.e. the comment preceding the sarcastic one) as input to
Gemmain order to produce neutral comments. For the sake of simplicity, we didn't use the context when training our models. Adding the context might considerably increase the performance of our models and maybe even solve some of the problems outlined at point (3) above
The trained models are available on Huggingface:
- maxmarcon/gpt2-sarcasm-defuser
- maxmarcon/gpt2-medium-sarcasm-defuser
- maxmarcon/bart-base-sarcasm-defuser
And they can be tested on this web application.
The code used to train and evaluate the models is available here.