BERT for Natural Language Inference (MNLI)

Description

A BERT-base-uncased checkpoint fine-tuned for natural language inference on the MultiNLI (MNLI) corpus. Given a premise and a hypothesis, it predicts whether the hypothesis is entailed by the premise, neutral toward it, or contradicts it.

In Spark NLP, load it with SampleEntailmentMatrix to compute bidirectional entailment between pairs of texts - two texts are treated as equivalent when each entails the other. This is the textbook approach to deciding whether two pieces of text “mean the same thing,” independent of exact wording, and is useful anywhere you need to cluster or deduplicate text by meaning rather than surface form.

Download Copy S3 URI

How to use

from sparknlp.base import *
from sparknlp.annotator import *
from pyspark.ml import Pipeline

document_assembler = DocumentAssembler() \
    .setInputCol("completions") \
    .setOutputCol("document")

entailment = SampleEntailmentMatrix.pretrained("bert_base_uncased_mnli_entailment_onnx", "en") \
    .setInputCols(["document"]) \
    .setOutputCol("completions_with_entailment")

nlp_pipeline = Pipeline(stages=[
    document_assembler,
    entailment
])

data = spark.createDataFrame([[[
    "The capital of France is Paris.",
    "Paris is France's capital.",
    "The capital of France is London."
]]]).toDF("completions")

result = nlp_pipeline.fit(data).transform(data)
result.selectExpr("completions_with_entailment[0].metadata['entailment_matrix'] as entailment_matrix").show(truncate=False)

import com.johnsnowlabs.nlp.base._
import com.johnsnowlabs.nlp.annotator._
import org.apache.spark.ml.Pipeline

val documentAssembler = new DocumentAssembler()
  .setInputCol("completions")
  .setOutputCol("document")

val entailment = SampleEntailmentMatrix.pretrained("bert_base_uncased_mnli_entailment_onnx", "en")
  .setInputCols("document")
  .setOutputCol("completions_with_entailment")

val pipeline = new Pipeline().setStages(Array(
  documentAssembler,
  entailment
))

val data = Seq(Seq(
  "The capital of France is Paris.",
  "Paris is France's capital.",
  "The capital of France is London."
)).toDF("completions")

val result = pipeline.fit(data).transform(data)
result.selectExpr("completions_with_entailment[0].metadata['entailment_matrix'] as entailment_matrix").show(false)

Results

+----------------------------------------------------------------------------------------------------------------------------------------------+
|entailment_matrix                                                                                                                             |
+----------------------------------------------------------------------------------------------------------------------------------------------+
|[[1.0,0.9935469627380371,3.128613461740315E-4],[0.9925073981285095,1.0,5.720576737076044E-4],[4.627297748811543E-4,0.0011771831195801497,1.0]]|
+----------------------------------------------------------------------------------------------------------------------------------------------+

Model Information

Model Name:	bert_base_uncased_mnli_entailment_onnx
Compatibility:	Spark NLP 7.0.0+
License:	Open Source
Edition:	Official
Input Labels:	[completions]
Output Labels:	[completions_with_entailment]
Language:	en
Size:	409.4 MB
Case sensitive:	false

PREVIOUSMARS Token Importance