Implementation:Kserve Kserve AnchorText Explainer

Knowledge Sources	Kserve_Kserve
Domains	Explainability, Natural Language Processing
Last Updated	2026-02-13 00:00 GMT

Overview

Concrete tool for generating anchor-based explanations for text classification models using the Alibi library and spaCy NLP provided by the KServe sample code.

Description

AnchorText extends ExplainerWrapper and wraps the Alibi AnchorText explainer. During construction, it receives a prediction function and an optional pre-built Alibi AnchorText explainer instance. If no explainer is provided, it lazily loads a spaCy language model (defaulting to en_core_web_md) and creates the AnchorText explainer on the first explain() call using an "unknown" sampling strategy. The explain() method extracts the first text input from the batch, checks whether the predictor returns class labels or probabilities (applying ArgmaxTransformer if needed), and runs the anchor text explanation algorithm. The result identifies which words in the text are sufficient for the classification decision.

Usage

Use this class as part of the Alibi explainer server pipeline when you need to generate word-level explanations for text classification models, identifying which words drive the model's prediction.

Code Reference

Source Location

• Repository: Kserve_Kserve
• File: docs/samples/explanation/alibi/alibiexplainer/alibiexplainer/anchor_text.py
• Lines: 1-62

Signature

class AnchorText(ExplainerWrapper):
    def __init__(
        self,
        predict_fn: Callable,
        explainer: Optional[alibi.explainers.AnchorText],
        spacy_language_model: str = "en_core_web_md",
        **kwargs,
    ):
        ...

    def explain(self, inputs: List, headers: Dict[str, str] = None) -> Explanation:
        ...

Import

from alibiexplainer.anchor_text import AnchorText

I/O Contract

Inputs

Constructor

Name	Type	Required	Description
predict_fn	Callable	Yes	Prediction function that accepts text input and returns predictions
explainer	Optional[alibi.explainers.AnchorText]	No	A pre-built Alibi AnchorText explainer instance; if None, one is created lazily
spacy_language_model	str	No	Name of the spaCy language model to load (default: "en_core_web_md")
**kwargs	dict	No	Additional keyword arguments passed to the Alibi explain call

explain()

Name	Type	Required	Description
inputs	List	Yes	List of text strings to explain (first text is used)
headers	Dict[str, str]	No	Optional HTTP headers

Outputs

explain()

Name	Type	Description
anchor_exp	Explanation	Alibi Explanation object containing anchor words and metadata

Usage Examples

Basic Usage

from alibiexplainer.anchor_text import AnchorText
import numpy as np

# Define a text prediction function
def my_predict_fn(texts):
    # Returns class probabilities for each text
    return np.array([[0.1, 0.9]])

wrapper = AnchorText(
    predict_fn=my_predict_fn,
    explainer=None,
    spacy_language_model="en_core_web_md"
)

# Generate an explanation for a text input
texts = ["This movie was absolutely fantastic and wonderful"]
explanation = wrapper.explain(texts)
print(explanation.anchor)  # e.g., ["fantastic", "wonderful"]

Related Pages

• Environment:Kserve_Kserve_Python_Runtime