diff --git a/website/package-lock.json b/website/package-lock.json
index f6d95f21..c1aca5cc 100644
--- a/website/package-lock.json
+++ b/website/package-lock.json
@@ -141,6 +141,16 @@
"markdown-it": "bin/markdown-it.js"
}
},
+ "node_modules/@babel/helper-string-parser": {
+ "version": "7.18.10",
+ "resolved": "https://registry.npmjs.org/@babel/helper-string-parser/-/helper-string-parser-7.18.10.tgz",
+ "integrity": "sha512-XtIfWmeNY3i4t7t4D2t02q50HvqHybPqW2ki1kosnvWCwuCMeo81Jf0gwr85jy/neUdg5XDdeFE/80DXiO+njw==",
+ "dev": true,
+ "peer": true,
+ "engines": {
+ "node": ">=6.9.0"
+ }
+ },
"node_modules/@babel/helper-validator-identifier": {
"version": "7.18.6",
"resolved": "https://registry.npmjs.org/@babel/helper-validator-identifier/-/helper-validator-identifier-7.18.6.tgz",
@@ -152,9 +162,9 @@
}
},
"node_modules/@babel/parser": {
- "version": "7.18.9",
- "resolved": "https://registry.npmjs.org/@babel/parser/-/parser-7.18.9.tgz",
- "integrity": "sha512-9uJveS9eY9DJ0t64YbIBZICtJy8a5QrDEVdiLCG97fVLpDTpGX7t8mMSb6OWw6Lrnjqj4O8zwjELX3dhoMgiBg==",
+ "version": "7.18.10",
+ "resolved": "https://registry.npmjs.org/@babel/parser/-/parser-7.18.10.tgz",
+ "integrity": "sha512-TYk3OA0HKL6qNryUayb5UUEhM/rkOQozIBEA5ITXh5DWrSp0TlUQXMyZmnWxG/DizSWBeeQ0Zbc5z8UGaaqoeg==",
"dev": true,
"peer": true,
"bin": {
@@ -165,12 +175,13 @@
}
},
"node_modules/@babel/types": {
- "version": "7.18.9",
- "resolved": "https://registry.npmjs.org/@babel/types/-/types-7.18.9.tgz",
- "integrity": "sha512-WwMLAg2MvJmt/rKEVQBBhIVffMmnilX4oe0sRe7iPOHIGsqpruFHHdrfj4O1CMMtgMtCU4oPafZjDPCRgO57Wg==",
+ "version": "7.18.10",
+ "resolved": "https://registry.npmjs.org/@babel/types/-/types-7.18.10.tgz",
+ "integrity": "sha512-MJvnbEiiNkpjo+LknnmRrqbY1GPUUggjv+wQVjetM/AONoupqRALB7I6jGqNUAZsKcRIEu2J6FRFvsczljjsaQ==",
"dev": true,
"peer": true,
"dependencies": {
+ "@babel/helper-string-parser": "^7.18.10",
"@babel/helper-validator-identifier": "^7.18.6",
"to-fast-properties": "^2.0.0"
},
@@ -303,9 +314,9 @@
"peer": true
},
"node_modules/@types/node": {
- "version": "18.6.2",
- "resolved": "https://registry.npmjs.org/@types/node/-/node-18.6.2.tgz",
- "integrity": "sha512-KcfkBq9H4PI6Vpu5B/KoPeuVDAbmi+2mDBqGPGUgoL7yXQtcWGu2vJWmmRkneWK3Rh0nIAX192Aa87AqKHYChQ==",
+ "version": "18.6.3",
+ "resolved": "https://registry.npmjs.org/@types/node/-/node-18.6.3.tgz",
+ "integrity": "sha512-6qKpDtoaYLM+5+AFChLhHermMQxc3TOEFIDzrZLPRGHPrLEwqFkkT5Kx3ju05g6X7uDPazz3jHbKPX0KzCjntg==",
"dev": true,
"peer": true
},
@@ -2162,9 +2173,9 @@
"peer": true
},
"node_modules/is-core-module": {
- "version": "2.9.0",
- "resolved": "https://registry.npmjs.org/is-core-module/-/is-core-module-2.9.0.tgz",
- "integrity": "sha512-+5FPy5PnwmO3lvfMb0AsoPaBG+5KHUI0wYFXOtYPnVVVspTFUuMZNfNaNVRt3FZadstu2c8x23vykRW/NBoU6A==",
+ "version": "2.10.0",
+ "resolved": "https://registry.npmjs.org/is-core-module/-/is-core-module-2.10.0.tgz",
+ "integrity": "sha512-Erxj2n/LDAZ7H8WNJXd9tw38GYM3dv8rk8Zcs+jJuxYTW7sozH+SS8NtrSjVL1/vpLvWi1hxy96IzjJ3EHTJJg==",
"dev": true,
"peer": true,
"dependencies": {
@@ -4804,9 +4815,9 @@
}
},
"node_modules/yargs-parser": {
- "version": "21.0.1",
- "resolved": "https://registry.npmjs.org/yargs-parser/-/yargs-parser-21.0.1.tgz",
- "integrity": "sha512-9BK1jFpLzJROCI5TzwZL/TU4gqjK5xiHV/RfWLOahrjAko/e4DJkRDZQXfvqAsiZzzYhgAzbgz6lg48jcm4GLg==",
+ "version": "21.1.0",
+ "resolved": "https://registry.npmjs.org/yargs-parser/-/yargs-parser-21.1.0.tgz",
+ "integrity": "sha512-xzm2t63xTV/f7+bGMSRzLhUNk1ajv/tDoaD5OeGyC3cFo2fl7My9Z4hS3q2VdQ7JaLvTxErO8Jp5pRIFGMD/zg==",
"dev": true,
"peer": true,
"engines": {
@@ -4915,6 +4926,13 @@
"normalize-path": "^3.0.0"
}
},
+ "@babel/helper-string-parser": {
+ "version": "7.18.10",
+ "resolved": "https://registry.npmjs.org/@babel/helper-string-parser/-/helper-string-parser-7.18.10.tgz",
+ "integrity": "sha512-XtIfWmeNY3i4t7t4D2t02q50HvqHybPqW2ki1kosnvWCwuCMeo81Jf0gwr85jy/neUdg5XDdeFE/80DXiO+njw==",
+ "dev": true,
+ "peer": true
+ },
"@babel/helper-validator-identifier": {
"version": "7.18.6",
"resolved": "https://registry.npmjs.org/@babel/helper-validator-identifier/-/helper-validator-identifier-7.18.6.tgz",
@@ -4923,19 +4941,20 @@
"peer": true
},
"@babel/parser": {
- "version": "7.18.9",
- "resolved": "https://registry.npmjs.org/@babel/parser/-/parser-7.18.9.tgz",
- "integrity": "sha512-9uJveS9eY9DJ0t64YbIBZICtJy8a5QrDEVdiLCG97fVLpDTpGX7t8mMSb6OWw6Lrnjqj4O8zwjELX3dhoMgiBg==",
+ "version": "7.18.10",
+ "resolved": "https://registry.npmjs.org/@babel/parser/-/parser-7.18.10.tgz",
+ "integrity": "sha512-TYk3OA0HKL6qNryUayb5UUEhM/rkOQozIBEA5ITXh5DWrSp0TlUQXMyZmnWxG/DizSWBeeQ0Zbc5z8UGaaqoeg==",
"dev": true,
"peer": true
},
"@babel/types": {
- "version": "7.18.9",
- "resolved": "https://registry.npmjs.org/@babel/types/-/types-7.18.9.tgz",
- "integrity": "sha512-WwMLAg2MvJmt/rKEVQBBhIVffMmnilX4oe0sRe7iPOHIGsqpruFHHdrfj4O1CMMtgMtCU4oPafZjDPCRgO57Wg==",
+ "version": "7.18.10",
+ "resolved": "https://registry.npmjs.org/@babel/types/-/types-7.18.10.tgz",
+ "integrity": "sha512-MJvnbEiiNkpjo+LknnmRrqbY1GPUUggjv+wQVjetM/AONoupqRALB7I6jGqNUAZsKcRIEu2J6FRFvsczljjsaQ==",
"dev": true,
"peer": true,
"requires": {
+ "@babel/helper-string-parser": "^7.18.10",
"@babel/helper-validator-identifier": "^7.18.6",
"to-fast-properties": "^2.0.0"
}
@@ -5043,9 +5062,9 @@
"peer": true
},
"@types/node": {
- "version": "18.6.2",
- "resolved": "https://registry.npmjs.org/@types/node/-/node-18.6.2.tgz",
- "integrity": "sha512-KcfkBq9H4PI6Vpu5B/KoPeuVDAbmi+2mDBqGPGUgoL7yXQtcWGu2vJWmmRkneWK3Rh0nIAX192Aa87AqKHYChQ==",
+ "version": "18.6.3",
+ "resolved": "https://registry.npmjs.org/@types/node/-/node-18.6.3.tgz",
+ "integrity": "sha512-6qKpDtoaYLM+5+AFChLhHermMQxc3TOEFIDzrZLPRGHPrLEwqFkkT5Kx3ju05g6X7uDPazz3jHbKPX0KzCjntg==",
"dev": true,
"peer": true
},
@@ -6537,9 +6556,9 @@
"peer": true
},
"is-core-module": {
- "version": "2.9.0",
- "resolved": "https://registry.npmjs.org/is-core-module/-/is-core-module-2.9.0.tgz",
- "integrity": "sha512-+5FPy5PnwmO3lvfMb0AsoPaBG+5KHUI0wYFXOtYPnVVVspTFUuMZNfNaNVRt3FZadstu2c8x23vykRW/NBoU6A==",
+ "version": "2.10.0",
+ "resolved": "https://registry.npmjs.org/is-core-module/-/is-core-module-2.10.0.tgz",
+ "integrity": "sha512-Erxj2n/LDAZ7H8WNJXd9tw38GYM3dv8rk8Zcs+jJuxYTW7sozH+SS8NtrSjVL1/vpLvWi1hxy96IzjJ3EHTJJg==",
"dev": true,
"peer": true,
"requires": {
@@ -8633,9 +8652,9 @@
}
},
"yargs-parser": {
- "version": "21.0.1",
- "resolved": "https://registry.npmjs.org/yargs-parser/-/yargs-parser-21.0.1.tgz",
- "integrity": "sha512-9BK1jFpLzJROCI5TzwZL/TU4gqjK5xiHV/RfWLOahrjAko/e4DJkRDZQXfvqAsiZzzYhgAzbgz6lg48jcm4GLg==",
+ "version": "21.1.0",
+ "resolved": "https://registry.npmjs.org/yargs-parser/-/yargs-parser-21.1.0.tgz",
+ "integrity": "sha512-xzm2t63xTV/f7+bGMSRzLhUNk1ajv/tDoaD5OeGyC3cFo2fl7My9Z4hS3q2VdQ7JaLvTxErO8Jp5pRIFGMD/zg==",
"dev": true,
"peer": true
}
diff --git a/website/package.json b/website/package.json
index c381b055..d735a438 100644
--- a/website/package.json
+++ b/website/package.json
@@ -1,7 +1,7 @@
{
"name": "lit-website",
"version": "1.0.0",
- "description": "Code for the Languange Interpretability Tool website, which is hosted as a github page.",
+ "description": "Code for the Language Interpretability Tool website, which is hosted as a github page.",
"main": "",
"scripts": {},
"author": "People + AI Research",
diff --git a/website/src/assets/images/text-salience-image-1.png b/website/src/assets/images/text-salience-image-1.png
new file mode 100644
index 00000000..30f695a0
Binary files /dev/null and b/website/src/assets/images/text-salience-image-1.png differ
diff --git a/website/src/assets/images/text-salience-image-10.png b/website/src/assets/images/text-salience-image-10.png
new file mode 100644
index 00000000..041aa5a6
Binary files /dev/null and b/website/src/assets/images/text-salience-image-10.png differ
diff --git a/website/src/assets/images/text-salience-image-11.png b/website/src/assets/images/text-salience-image-11.png
new file mode 100644
index 00000000..5a112a09
Binary files /dev/null and b/website/src/assets/images/text-salience-image-11.png differ
diff --git a/website/src/assets/images/text-salience-image-12.png b/website/src/assets/images/text-salience-image-12.png
new file mode 100644
index 00000000..3c1471a0
Binary files /dev/null and b/website/src/assets/images/text-salience-image-12.png differ
diff --git a/website/src/assets/images/text-salience-image-13.png b/website/src/assets/images/text-salience-image-13.png
new file mode 100644
index 00000000..c703eda0
Binary files /dev/null and b/website/src/assets/images/text-salience-image-13.png differ
diff --git a/website/src/assets/images/text-salience-image-14.png b/website/src/assets/images/text-salience-image-14.png
new file mode 100644
index 00000000..0298bbfd
Binary files /dev/null and b/website/src/assets/images/text-salience-image-14.png differ
diff --git a/website/src/assets/images/text-salience-image-15.png b/website/src/assets/images/text-salience-image-15.png
new file mode 100644
index 00000000..8157c9fc
Binary files /dev/null and b/website/src/assets/images/text-salience-image-15.png differ
diff --git a/website/src/assets/images/text-salience-image-16.png b/website/src/assets/images/text-salience-image-16.png
new file mode 100644
index 00000000..f064bd70
Binary files /dev/null and b/website/src/assets/images/text-salience-image-16.png differ
diff --git a/website/src/assets/images/text-salience-image-2.png b/website/src/assets/images/text-salience-image-2.png
new file mode 100644
index 00000000..6e0bd216
Binary files /dev/null and b/website/src/assets/images/text-salience-image-2.png differ
diff --git a/website/src/assets/images/text-salience-image-3.png b/website/src/assets/images/text-salience-image-3.png
new file mode 100644
index 00000000..2fcb5edb
Binary files /dev/null and b/website/src/assets/images/text-salience-image-3.png differ
diff --git a/website/src/assets/images/text-salience-image-4.png b/website/src/assets/images/text-salience-image-4.png
new file mode 100644
index 00000000..e8d93931
Binary files /dev/null and b/website/src/assets/images/text-salience-image-4.png differ
diff --git a/website/src/assets/images/text-salience-image-5.png b/website/src/assets/images/text-salience-image-5.png
new file mode 100644
index 00000000..e286d4b3
Binary files /dev/null and b/website/src/assets/images/text-salience-image-5.png differ
diff --git a/website/src/assets/images/text-salience-image-6.png b/website/src/assets/images/text-salience-image-6.png
new file mode 100644
index 00000000..2b24d681
Binary files /dev/null and b/website/src/assets/images/text-salience-image-6.png differ
diff --git a/website/src/assets/images/text-salience-image-7.png b/website/src/assets/images/text-salience-image-7.png
new file mode 100644
index 00000000..7a1d4c21
Binary files /dev/null and b/website/src/assets/images/text-salience-image-7.png differ
diff --git a/website/src/assets/images/text-salience-image-8.png b/website/src/assets/images/text-salience-image-8.png
new file mode 100644
index 00000000..3143e4ed
Binary files /dev/null and b/website/src/assets/images/text-salience-image-8.png differ
diff --git a/website/src/assets/images/text-salience-image-9.png b/website/src/assets/images/text-salience-image-9.png
new file mode 100644
index 00000000..54a5e3f6
Binary files /dev/null and b/website/src/assets/images/text-salience-image-9.png differ
diff --git a/website/src/tutorials.md b/website/src/tutorials.md
index a1c9b713..c6bd8063 100644
--- a/website/src/tutorials.md
+++ b/website/src/tutorials.md
@@ -25,6 +25,9 @@ c-copy: "Get familiar with the interface of the Language Interpretability Tool."
## Conducting analysis in LIT
+{% include partials/tutorial-link-element c-title: "Salience Maps for Text", link: "/tutorials/text-salience",
+c-copy: "Learn how to use salience maps for text data in LIT." %}
+
{% include partials/tutorial-link-element c-title: "Tabular Feature Attribution", link: "/tutorials/tab-feat-attr",
c-copy: "Learn how to use the Kernel SHAP based Tabular Feature Attribution module in LIT." %}
diff --git a/website/src/tutorials/text-salience.md b/website/src/tutorials/text-salience.md
new file mode 100644
index 00000000..02343734
--- /dev/null
+++ b/website/src/tutorials/text-salience.md
@@ -0,0 +1,337 @@
+---
+title: Salience Maps for Text
+layout: layouts/tutorial.liquid
+
+hero-image: /assets/images/sample-banner.png
+hero-title: "Salience Maps for Text"
+hero-copy: "Learn how to use salience maps for text data in LIT."
+
+bc-anchor-category: "analysis"
+bc-category-title: "Analysis"
+bc-title: "Salience Maps for Text"
+
+time: "15 minutes"
+takeaways: "Learn how to use salience maps for text data in LIT."
+---
+
+## Tutorial : Salience Maps for Text
+
+{% include partials/link-out link: "../../demos/glue.html", text: "Explore this demo yourself." %}
+
+Or, run your own with [`examples/glue_demo.py`](https://github.com/PAIR-code/lit/blob/main/lit_nlp/examples/glue_demo.py)
+
+LIT enables users to analyze individual predictions for text input using
+salience maps, for which gradient-based and/or blackbox methods are available.
+In this tutorial, we will explore how to use salience maps to analyze a text
+classifier in the [Classification and Regression models demo](https://pair-code.github.io/lit/demos/glue.html)
+from the LIT website, and
+how these findings can support counterfactual analysis using LIT’s generators,
+such as Hotflip, to test hypotheses. The Salience Maps module can be found under
+the Explanations tab in the bottom half of this demo and it supports four
+different methods for the GLUE model under test (with other models it might
+support a different number of these methods) -
+[Grad L2 Norm](https://aclanthology.org/P18-1032/),
+[Grad · Input](https://arxiv.org/abs/1412.6815),
+[Integrated Gradients](https://arxiv.org/pdf/1703.01365.pdf) (IG)
+and [LIME](https://arxiv.org/pdf/1602.04938v3.pdf).
+
+### Heuristics : Which salience method for which task?
+
+Salience methods are imperfect. Research has shown that salience methods are
+often “[sensitive to factors that do not contribute to a model’s prediction](https://arxiv.org/abs/1711.00867)”;
+that people tend to [overly trust salience values or use methods they believe they know incorrectly](https://dl.acm.org/doi/10.1145/3313831.3376219);
+and that [model architecture may directly impact the utility](https://arxiv.org/pdf/2111.07367.pdf)
+of different salience methods.
+
+With those limitations in mind, the question remains as to which methods should
+be used and when. To offer some guidance, we have come up with the following
+decision aid that provides some ideas about which salience method(s) might be
+appropriate.
+
+{% include partials/inset-image image: '/assets/images/text-salience-image-1.png',
+ caption: 'TODO'%}
+
+If your model does not output gradients with its predictions (i.e., is a
+blackbox), [LIME](https://arxiv.org/pdf/1602.04938v3.pdf) is your only choice as
+it is currently the only black-box method LIT supports for text data.
+
+If your model does output gradients, then you can choose among three methods:
+[Grad L2 Norm](https://aclanthology.org/P18-1032/),
+[Grad · Input](https://arxiv.org/abs/1412.6815), and
+[Integrated Gradients](https://arxiv.org/pdf/1703.01365.pdf) (IG).
+Grad L2 Norm and Grad · Input are easy to use and fast to compute, but can suffer from gradient
+saturation. IG addresses the gradient saturation issue in the Grad methods
+(described in detail below), but requires that the model output both gradients
+and embeddings, is much more expensive to compute, and requires parameterization
+to optimize results.
+
+Remember that a good investigative process will check for commonalities and
+patterns across salience values from multiple salience methods. Further,
+salience methods should be an entry point for developing hypotheses about your
+model’s behavior, and for identifying subsets of examples and/or creating
+counterfactual examples that test those hypotheses.
+
+### Salience Maps for Text : Theoretical background and LIT overview
+
+All methods calculate salience, but there are subtle differences in their
+approaches towards calculating a salience score for each token. Grad L2 Norm
+only produces absolute salience scores while other methods like Grad · Input
+(and also Integrated Gradients and LIME) produce signed values, leading to an
+improved interpretation of whether a token has positive or negative influence on
+the prediction.
+
+**_LIT uses different color scales to represent signed and unsigned salience scores_**.
+Methods that produce unsigned salience values, such as Grad L2 Norm,
+use a purple scale where darker colors indicate greater salience, whereas the other methods
+use a red-to-green scale, with red denoting negative scores and green denoting
+positive.
+
+{% include partials/info-box title: 'Interpreting salience polarity',
+ text: "Salience is always relative to the model’s prediction of one class.
+ Intuitively, a positive influence score (attribution) for a token (or word,
+ depending on your method) in an example means that if this token was removed
+ we expect a drop in model confidence in the prediction of the class.
+ Similarly, removing a negative token would correspond to an increase in the
+ model's confidence in the prediction of this class."%}
+
+{% include partials/inset-image image: '/assets/images/text-salience-image-2.png',
+ caption: 'TODO'%}
+
+#### Token-Based Methods
+
+[Gradient saturation](https://towardsdatascience.com/the-vanishing-gradient-problem-69bf08b15484)
+is a potential problem for all of the Gradient based methods, such as [Grad L2 Norm](https://aclanthology.org/P18-1032/)
+and [Grad · Input](https://arxiv.org/abs/1412.6815), that we need to look out
+for. Essentially if the model learning saturates for a particular token, then
+its gradient goes to zero and appears to have zero salience. At the same time,
+some tokens actually have a zero salience score, because they do not affect the
+predictions. And there is no simple way to tell if a token that we are
+interested in is legitimately irrelevant or if we are just observing the effects
+of gradient saturation.
+
+The [integrated gradients](https://arxiv.org/pdf/1703.01365.pdf) method
+addresses the gradient saturation problem by enriching gradients with
+embeddings.
+[Tokens](https://jalammar.github.io/a-visual-guide-to-using-bert-for-the-first-time/)
+are the discrete building blocks of text sequences, but they can also be
+represented as vectors in a [continuous embedding space](https://colah.github.io/posts/2014-07-NLP-RNNs-Representations/).
+IG computes per-token salience as the average salience over a set of local
+gradients computed by interpolating between the token’s embedding vectors and a
+baseline (typically the zero vector). The tradeoff is that IG requires more
+effort to identify the right number of interpolation steps to be
+effective (configurable in LIT’s interface), with the number of steps
+correlating directly with runtime. It also requires more information,
+which the model may or may not be able to provide.
+
+{% include partials/inset-image image: '/assets/images/text-salience-image-3.png',
+ caption: 'TODO'%}
+
+#### Blackbox Methods
+
+Some models do not provide tokens or token-level gradients, effectively making
+them blackboxes. [LIME](https://arxiv.org/pdf/1602.04938v3.pdf) can be used with
+these models. LIME works by generating a set of perturbed inputs, generally, by
+dropping out or masking tokens, and training a local linear model to reconstruct
+the original model's predictions. The weights of this linear model are treated
+as the salience values.
+
+LIME has two limitations, compared to gradient-based methods:
+
+1. it can be slow as it requires many evaluations of the model, and
+2. it can be noisy on longer inputs where there are more tokens to ablate.
+
+We can increase **_the number of samples to be used for LIME_**
+within LIT to counter the potential noisiness, however
+this is at the cost of computation time.
+
+In the example below, we go from the default number of samples (256) up to 4096,
+increasing in powers of 2 (i.e., 256, 512, 1024, 2048, 4096). Note through these
+transitions, how the model becomes more sure of the salience scores it outputs
+(indicated by stronger colors for the key tokens). A particularly interesting
+observation for this example is how the model isn’t sure about the word
+“although” in this sentence, but later on learns that it has a negative
+connotation and assigns a non-negligible negative salience score to it.
+
+{% include partials/inset-image image: '/assets/images/text-salience-image-4.png',
+ caption: 'TODO'%}
+
+{% include partials/inset-image image: '/assets/images/text-salience-image-5.png',
+ caption: 'TODO'%}
+
+{% include partials/inset-image image: '/assets/images/text-salience-image-6.png',
+ caption: 'TODO'%}
+
+{% include partials/inset-image image: '/assets/images/text-salience-image-7.png',
+ caption: 'TODO'%}
+
+{% include partials/inset-image image: '/assets/images/text-salience-image-8.png',
+ caption: 'TODO'%}
+
+Another interesting difference between the gradient based methods and LIME lies
+in how they analyze the input. The gradient based methods use the model’s
+tokenizer, which splits up words into smaller constituents, whereas LIME
+splits the text into words at whitespaces.
+Thus, LIME’s word-level results are often incomparable with the token-level
+results from other methods, as you can see in the salience maps below.
+
+{% include partials/inset-image image: '/assets/images/text-salience-image-9.png',
+ caption: 'TODO'%}
+
+### Single example use-case : What do we do with the output of the salience maps module
+
+Let’s take a concrete example and walkthrough how we might use the salience maps
+module and counterfactual generators to analyze the behavior of the `sst2-tiny`
+model on the classification task.
+
+First, let’s refer back to our heuristic for choosing appropriate methods.
+Because `sst2-tiny` does not have a LSTM architecture, we shouldn’t rely too much
+on Grad · Input. So, we are left with Grad L2 Norm, Integrated Gradients and
+LIME to base our decisions on.
+
+To gain some confidence in our heuristic, we look for examples where Grad ˙
+Input performs poorly compared to the other methods. There are quite a few in
+the dataset, for example the sentence below where Grad · Input predicts
+completely opposite salience scores to its counterparts.
+
+{% include partials/inset-image image: '/assets/images/text-salience-image-10.png',
+ caption: 'TODO'%}
+
+#### Use Case 1: Sexism Analysis with Counterfactuals
+
+Coming back to our use-case, we want to investigate if the model displays sexist
+behavior for a particular input sentence. We take a datapoint with a negative
+sentiment label, which talks about the performance of an actress in the movie.
+
+The key words/tokens (based on salience scores across the three chosen methods)
+in this sentence are “hampered”, “lifetime-channel”, “lead”, “actress”, “her”
+and “depth”. The only words out of this which are related to gender are
+“actress” and “her”. The words “actress” and “her” get a significant weight
+for both Grad L2 Norm and IG, and is assigned a positive score (IG scores are
+slightly stronger than Grad L2 Norm scores), indicating that the gender of the
+person is helping the model be sure of its predictions of this sentence being a
+negative review sentiment. However for LIME, the salience scores for these two
+words is a small negative number, indicating that the gender of the model is
+actually causing a small decrease in model confidence for the prediction of this
+being a negative review. Even with this small disparity between the token-based
+and blackbox methods in the gender related words in the sentence, it turns out
+that these are not the most important words. “Hampered”, “lifetime-channel” and
+“plot” are the dominating words/tokens for this particular example in helping
+the model make its decision. We still want to explore if reversing the gender
+might change this. Would it make the model give more or less importance to other
+tokens or the tokens we replaced? Would it change the model prediction
+confidence scores?
+
+To do this, we generate a counterfactual example using the Datapoint
+Editor which is located right beside the Data Table in the UI,
+changing "actress" with "actor" and "her" with "his" after selecting our
+datapoint of interest. An alternative to this approach is to use the Word
+replacer under the Counterfactuals tab in the bottom half of the LIT app to
+achieve the same task. If our model is predicting a negative sentiment
+due to sexist influences towards
+“actress” or “her”, then the hypothesis is that it should show opposite
+sentiments if we flip those key tokens.
+
+{% include partials/inset-image image: '/assets/images/text-salience-image-11.png',
+ caption: 'TODO'%}
+
+However, it turns out that there is very minimal (and hence negligible) change
+in the salience score values of any of the tokens. The model doesn’t change its
+prediction either. It still predicts this to be a negative review sentiment with
+approximately the same prediction confidence. This indicates that at least for
+this particular example, our model isn’t displaying sexist behavior and is
+actually making its prediction based on key tokens in the sentence which are not
+related to the gender of the actress/actor.
+
+#### Use Case 2: Pairwise Comparisons
+
+Let’s take another example. This time we consider the sentence “a sometimes
+tedious film” and generate three counterfactuals, first by replacing the two
+words “sometimes” and “tedious” with their respective antonyms one-by-one and
+then together to observe the changes in predictions and salience.
+
+To create the counterfactuals, we can simply use the Datapoint Editor which is
+located right beside the Data Table in the UI. We can just select our data point
+of interest (data point 6), and then replace the words we are interested in with
+the respective substitutes. Then we assign a `label` to the newly created sentence
+and add it to our data. For this particular example, we are assigning 0 when
+"tedious" appears and 1 when "exciting" appears in the sentence.
+An alternative to this approach is to use the Word
+replacer under the Counterfactuals tab in the bottom half of the LIT app to
+achieve the same task.
+
+{% include partials/inset-image image: '/assets/images/text-salience-image-12.png',
+ caption: 'TODO'%}
+
+We can pin the original sentence in the data table and then cycle through the
+three available pairs by selecting each of the new sentences as our primary
+selection. This will give us a comparison-type output in the Salience Maps
+module between the pinned and the selected examples.
+
+{% include partials/inset-image image: '/assets/images/text-salience-image-13.png',
+ caption: 'TODO'%}
+
+When we replace “sometimes” with “often”, it gets a negative score of almost
+equal magnitude (reversing polarity) from LIME which makes sense, because
+“often” makes the next word in the sentence more impactful, linguistically. The model
+prediction doesn’t change either, and this new review is still classified as
+having a negative sentiment.
+
+{% include partials/inset-image image: '/assets/images/text-salience-image-14.png',
+ caption: 'TODO'%}
+
+On replacing “tedious” with “exciting”, the salience for “sometimes” changes
+from positive score to negative in the LIME output. In the IG output,
+“sometimes” changes from a strong positive score to a weak positive score. These
+changes are also justified because in this new sentence “sometimes” counters the
+positive effect of the word “exciting”. The main negative word in our original
+datapoint was “tedious” and by replacing this with a positive word “exciting”,
+the model’s classification of this new sentence also changes and the new
+sentence is classified as positive with a very high confidence score.
+
+{% include partials/inset-image image: '/assets/images/text-salience-image-15.png',
+ caption: 'TODO'%}
+
+And finally, when we replace both “sometimes tedious” with “often exciting”, we
+get strong positive scores from both LIME and IG, which is in line with the
+overall strong positive sentiment of the sentence. The model predicts this new
+sentence as positive sentiment, and the confidence score for this prediction is
+slightly higher than the previous sentence where instead of “often” we had used
+“sometimes”. This makes sense as well because “often” enhances the positive
+sentiment slightly more than using “sometimes” in a positive review.
+
+{% include partials/inset-image image: '/assets/images/text-salience-image-16.png',
+ caption: 'TODO'%}
+
+In this second example, we mostly based our observation on LIME and IG, because
+we could observe visual changes directly from the outputs of these methods. Grad
+L2 Norm outputs were comparatively inconclusive, highlighting the need to
+select appropriate methods and compare results between them. The model
+predictions were in
+line with our expected class labels and the confidence scores for predictions on
+the counterfactuals could be justified using salience scores assigned to the new
+tokens.
+
+#### Use Case 3: Quality Assurance
+
+A real life use case for the salience maps module can be in Quality Assurance.
+For example, if there is a failure in production (e.g., wrong results for a search
+query), we know the text input and the label the model predicted. We can use LIT
+Salience Maps to debug this failure and figure out which tokens were most
+influential in the prediction of the wrong label, and which alternative labels
+could have been predicted (i.e., is there one clear winner, or are there a few
+that are roughly the same?). Once we are done with debugging using LIT, we can
+make the necessary changes to the model or training data (eg. adding fail-safes
+or checks) to solve the production failure.
+
+### Conclusion
+
+Three gradient-based salience methods and one blackbox method are provided out
+of the box to LIT users who need to use these post-hoc interpretations to make
+sense of their language model’s predictions. This diverse array of built-in
+techniques can be used in combination with other LIT modules like
+counterfactuals to support robust exploration of a model's behavior, as
+illustrated in this tutorial. And as always, LIT strives to
+enable users to
+[add their own salience interpreters](https://github.com/PAIR-code/lit/wiki/api.md#interpretation-components)
+to allow for a wider variety
+of use cases beyond these default capabilities!