{
"cells": [
{
"cell_type": "markdown",
"metadata": {},
"source": [
"# Positive vs. Negative Sentiment Classification\n",
"\n",
"In this notebook, we demonstrate how to interpret a sentiment classification model using SHAP. The goal is to understand how individual words in a movie review influence the model's prediction of positive or negative sentiment."
]
},
{
"cell_type": "code",
"execution_count": 1,
"metadata": {},
"outputs": [
{
"name": "stderr",
"output_type": "stream",
"text": [
"/Users/aribaa/Library/Python/3.9/lib/python/site-packages/urllib3/__init__.py:35: NotOpenSSLWarning: urllib3 v2 only supports OpenSSL 1.1.1+, currently the 'ssl' module is compiled with 'LibreSSL 2.8.3'. See: https://github.com/urllib3/urllib3/issues/3020\n",
" warnings.warn(\n",
"/Users/aribaa/Library/Python/3.9/lib/python/site-packages/tqdm/auto.py:21: TqdmWarning: IProgress not found. Please update jupyter and ipywidgets. See https://ipywidgets.readthedocs.io/en/stable/user_install.html\n",
" from .autonotebook import tqdm as notebook_tqdm\n"
]
}
],
"source": [
"import datasets\n",
"import numpy as np\n",
"import transformers\n",
"\n",
"import shap"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## Load the IMDB movie review dataset"
]
},
{
"cell_type": "code",
"execution_count": 2,
"metadata": {},
"outputs": [],
"source": [
"# Load IMDB dataset (movie reviews labeled as positive/negative)\n",
"dataset = datasets.load_dataset(\"imdb\", split=\"test\")\n",
"\n",
"# shorten the strings to fit into the pipeline model\n",
"short_data = [v[:500] for v in dataset[\"text\"][:20]]"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## Load and run a sentiment analysis pipeline"
]
},
{
"cell_type": "code",
"execution_count": 3,
"metadata": {},
"outputs": [
{
"name": "stderr",
"output_type": "stream",
"text": [
"No model was supplied, defaulted to distilbert/distilbert-base-uncased-finetuned-sst-2-english and revision 714eb0f (https://huggingface.co/distilbert/distilbert-base-uncased-finetuned-sst-2-english).\n",
"Using a pipeline without specifying a model name and revision in production is not recommended.\n",
"Device set to use mps:0\n",
"/Users/aribaa/Library/Python/3.9/lib/python/site-packages/transformers/pipelines/text_classification.py:111: UserWarning: `return_all_scores` is now deprecated, if want a similar functionality use `top_k=None` instead of `return_all_scores=True` or `top_k=1` instead of `return_all_scores=False`.\n",
" warnings.warn(\n"
]
},
{
"data": {
"text/plain": [
"[[{'label': 'NEGATIVE', 'score': 0.07581914216279984},\n",
" {'label': 'POSITIVE', 'score': 0.924180805683136}],\n",
" [{'label': 'NEGATIVE', 'score': 0.01834261603653431},\n",
" {'label': 'POSITIVE', 'score': 0.9816573858261108}]]"
]
},
"execution_count": 3,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"# Load pretrained sentiment analysis model from HuggingFace\n",
"# Note: Model will be downloaded on first run\n",
"classifier = transformers.pipeline(\"sentiment-analysis\", return_all_scores=True)\n",
"classifier(short_data[:2])"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## Explain the sentiment analysis pipeline"
]
},
{
"cell_type": "code",
"execution_count": 4,
"metadata": {},
"outputs": [],
"source": [
"# Create SHAP explainer to compute word-level importance\n",
"explainer = shap.Explainer(classifier)"
]
},
{
"cell_type": "code",
"execution_count": 5,
"metadata": {},
"outputs": [
{
"name": "stderr",
"output_type": "stream",
"text": [
"PartitionExplainer explainer: 3it [00:11, 11.25s/it] \n"
]
}
],
"source": [
"# explain the predictions of the pipeline on the first two samples\n",
"shap_values = explainer(short_data[:2])"
]
},
{
"cell_type": "code",
"execution_count": 6,
"metadata": {},
"outputs": [
{
"data": {
"text/html": [
"\n",
" \n",
" \n",
"
[0]
\n",
"
\n",
"
inputs
0.021 / 2
I
0.045 / 2
love sci
0.012 / 3
-fi and
0.016 / 9
am willing to put up with a lot.
-0.013 / 8
Sci-fi movies/TV are usually
-0.016 / 4
underfunded,
-0.016
under
-0.016
-
-0.013
appreciated
-0.015
and
-0.027 / 2
misunderstood.
-0.019 / 2
I tried
-0.011 / 3
to like this
-0.004
,
0.007 / 2
I really
0.006
did
-0.003
,
0.055
but
0.044 / 2
it is
0.058 / 2
to good
0.034
TV
0.03
sci
0.039 / 2
-fi
0.035 / 2
as Babylon
0.033 / 2
5 is
0.038 / 3
to Star Trek
0.015 / 4
(the original)
0.005
.
-0.002 / 5
Silly prosthetics
0.002
,
-0.022 / 4
cheap cardboard sets,
-0.016 / 4
stilted dialogues,
-0.007 / 3
CG that
-0.009 / 2
doesn'
-0.009
t
-0.014 / 3
match the background
-0.006
,
-0.014
and
-0.028 / 4
painfully one-dimensional
-0.001
characters
-0.008
cannot
-0.01 / 3
be overcome with
-0.013 / 8
a 'sci-fi' setting.
-0.0
(
0.02
I
0.016
'
0.03 / 2
m sure
-0.014 / 2
there are
-0.002 / 2
those of
-0.004
you
-0.0
out
-0.037
there
-0.049
who
0.006
think
-0.006
Babylon
-0.005
5
-0.028
is
0.268
good
0.059
sci
0.014
-
0.039
fi
-0.004
\n",
" \n",
" \n",
"
[1]
\n",
"
\n",
"
inputs
0.017
0.164
Worth
0.043
the
0.075
entertainment
-0.002 / 4
value of a rental
0.007
,
0.011
especially
-0.0
if
0.031
you
0.019
like
0.036 / 2
action movies
0.036
.
0.009 / 2
This one
0.015 / 2
features the
0.001 / 3
usual car chases
0.004
,
0.026 / 5
fights with the great Van
0.029 / 5
Damme kick style,
0.001 / 4
shooting battles with the
-0.031 / 4
40 shell load shotgun
-0.004
,
0.006
and
-0.015 / 2
even terrorist
-0.015 / 2
style bombs
0.004
.
0.003 / 2
All of
0.02
this
0.048
is
0.085
entertaining
0.019
and
0.018 / 2
competently
0.007
handled
-0.114
but
-0.012 / 2
there is
-0.111
nothing
-0.033
that
-0.054
really
-0.002
blows
0.002 / 3
you away if
0.009 / 2
you'
0.003 / 2
ve seen
-0.001 / 3
your share before
0.006
.
0.013 / 8
<br /><br />
0.003 / 2
The plot
0.047 / 2
is made
0.115
interesting
0.076
by
0.016 / 2
the inclusion
-0.023 / 3
of a rabbit
-0.014
,
0.009 / 4
which is clever but
-0.035
hardly
-0.018
profound
-0.016
.
-0.006 / 2
Many of
-0.005 / 3
the c
"
],
"text/plain": [
""
]
},
"metadata": {},
"output_type": "display_data"
}
],
"source": [
"# Visualize contribution of each word to the prediction for the \"POSITIVE\" class\n",
"shap.plots.text(shap_values[:, :, \"POSITIVE\"])"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## Wrap the pipeline manually\n",
"\n",
"SHAP explanations operate best in additive feature spaces. Since probabilities are not additive, we convert them into logits (log-odds). This ensures that SHAP values correctly sum up to the model output."
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"### Create a TransformersPipeline wrapper"
]
},
{
"cell_type": "code",
"execution_count": 7,
"metadata": {},
"outputs": [],
"source": [
"pmodel = shap.models.TransformersPipeline(classifier, rescale_to_logits=False)"
]
},
{
"cell_type": "code",
"execution_count": 8,
"metadata": {},
"outputs": [
{
"data": {
"text/plain": [
"array([[0.07581914, 0.92418081],\n",
" [0.01834262, 0.98165739]])"
]
},
"execution_count": 8,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"pmodel(short_data[:2])"
]
},
{
"cell_type": "code",
"execution_count": 9,
"metadata": {},
"outputs": [
{
"data": {
"text/plain": [
"array([[-2.50055699, 2.50055625],\n",
" [-3.98001525, 3.98001536]])"
]
},
"execution_count": 9,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"pmodel = shap.models.TransformersPipeline(classifier, rescale_to_logits=True)\n",
"pmodel(short_data[:2])"
]
},
{
"cell_type": "code",
"execution_count": 10,
"metadata": {},
"outputs": [
{
"name": "stderr",
"output_type": "stream",
"text": [
"PartitionExplainer explainer: 3it [00:17, 17.59s/it] \n"
]
},
{
"data": {
"text/html": [
"\n",
" \n",
" \n",
"