Source-linked AI summary
Performance Impact Caused by Hidden Bias of Training Data for Recognizing Textual Entailment
Masatoshi Tsuchiya
TL;DR
Training-data quality is a central concern for learning-based RTE systems, but corpus-level hidden bias requires a way to measure it. The paper proposes a two-phase statistical testing method and finds that SNLI hypothesis sentences contain label-predictive bias, with consequences for NN model performance.
Problem
Training-data quality is a crucial problem for learning-centered RTE approaches, motivating evaluation of hidden bias in large RTE corpora.
Method
The paper introduces a two-phase method that formulates label predictability from hypothesis sentences as a null hypothesis and tests it with a multinomial Naive Bayes model.
Results
SNLI’s hypothesis-only label predictor achieves 63.3% accuracy versus 34.3% for the label-distribution baseline, indicating hidden bias in SNLI.
Takeaways & Limitations
The hidden bias makes a large portion of NN models’ high RTE accuracy benefit from hypothesis sentences and risks making models work differently than intended.
Takeaways & Limitations
The paper’s stated assumption is that TE labels should not be determinable from hypothesis sentences without premise context.
Abstract
from arXiv · showhide
The quality of training data is one of the crucial problems when a learning-centered approach is employed. This paper proposes a new method to investigate the quality of a large corpus designed for the recognizing textual entailment (RTE) task. The proposed method, which is inspired by a statistical hypothesis test, consists of two phases: the first phase is to introduce the predictability of textual entailment labels as a null hypothesis which is extremely unacceptable if a target corpus has no hidden bias, and the second phase is to test the null hypothesis using a Naive Bayes model. The experimental result of the Stanford Natural Language Inference (SNLI) corpus does not reject the null hypothesis. Therefore, it indicates that the SNLI corpus has a hidden bias which allows prediction of textual entailment labels from hypothesis sentences even if no context information is given by a premise sentence. This paper also presents the performance impact of NN models for RTE caused by this hidden bias.
1. Introduction
The paper addresses training-data quality in RTE by proposing an empirical method for detecting hidden corpus bias. It applies the method to SNLI and SICK and examines how such bias can affect neural RTE models.
- Motivation: Training-data quality is a crucial concern because systematic annotation patterns can mislead learning-based models, including neural networks.The paper motivates both methods for improving training-data quality and metrics for evaluating its reliability.
- Finding: SNLI hypothesis sentences permit prediction of textual-entailment labels without premise context, contrary to the RTE task definition.In RTE, the label is intended to depend on the relationship between premise and hypothesis sentences.
- Contributions: The proposed method empirically reveals hidden bias in large RTE corpora by testing whether textual-entailment labels are predictable from hypothesis sentences.The method is presented as a new contribution for investigating corpus quality.
- Contributions: The method is applied to the SNLI and SICK corpora, revealing hidden bias in SNLI.The paper contrasts the two corpora through its empirical evaluation.
- Impact: The hidden bias creates a risk that an NN model for RTE works as an entirely different model from what its constructor expects.The paper identifies this risk as a separate contribution concerning the performance impact of biased training data.
2. Proposed Method
The proposed method tests whether textual-entailment labels can be predicted from hypothesis sentences without premise context, using a two-phase statistical procedure. It compares a Naive Bayes label-prediction model with a label-frequency baseline to assess hidden bias.
- 2.1. Predictability of TE Labels without Premise Sentences: The method derives a null hypothesis stating that a hypothesis sentence’s TE label is determinable without its premise sentence.This hypothesis is treated as extremely unacceptable when a corpus has no hidden bias.
- 2.1. Predictability of TE Labels without Premise Sentences: The first phase defines the null hypothesis from the RTE task’s premise–hypothesis label structure.RTE partitions relationships into entailment, neutral, and contradiction, which normally require premise context.
- 2.2. TE Label Prediction Model: The second phase tests the null hypothesis using two models: a multinomial Naive Bayes predictor and a label-frequency baseline.The predictor uses hypothesis sentences without premise context, while the baseline uses only the corpus distribution P(y).
- 2.2. TE Label Prediction Model: The Naive Bayes model uses all word unigrams in hypothesis sentences as features for predicting TE labels.The model estimates label probabilities from hypothesis features and selects the most probable label.
- 2.2. TE Label Prediction Model: A statistically significant difference between the predictor and baseline means the null hypothesis is not rejected and indicates hidden corpus bias.The baseline assigns the most frequent TE label when no premise or hypothesis information is available.
3. Experiment
The experiment evaluates hypothesis-only TE-label prediction on SNLI and SICK. SNLI supports substantially better prediction than its label-frequency baseline, whereas SICK performs near baseline, indicating hidden bias in SNLI hypotheses.
- 3. Experiment: 63.3% accuracy was achieved on SNLI by the hypothesis-only TE label prediction model, versus 34.3% for the SNLI label-distribution baseline.The sign test found a statistically significant difference, with p = 5.7e−202.
- 3. Experiment: 56.7% performance was achieved on SICK, close to its baseline, with no statistically significant difference between the models.The sign test reported p = 0.65 for the SICK comparison.
- 3. Experiment: SNLI’s prediction model attempts to assign appropriate TE labels to individual hypotheses, while SICK’s model mostly assigns the major label ‘neutral’.The contrasting confusion-matrix behavior supports rejection of the null hypothesis for SICK but not for SNLI.
- 3. Experiment: The SNLI corpus contains hidden bias that permits TE-label prediction from hypothesis sentences without premise sentences.This conclusion follows from the SNLI model’s performance and the associated statistical test.
4. Discussion
The discussion shows that hidden bias substantially affects SNLI-trained RTE models: high overall performance is concentrated in empirically easy examples, while performance drops sharply on hard examples and remains above chance even without premise information. The authors therefore warn that these models may behave as hypothesis-label predictors rather than RTE systems.
- Empirical classification: 63.3% of SNLI test pairs were classified as empirical easy examples because their TE labels were correctly predicted from hypotheses alone.The remaining test pairs formed the empirical hard set.
- NN models: The parallel and sequential LSTM models were evaluated as distinct RTE architectures, with the sequential model connecting two LSTMs directly.The sequential model used no attention in this experiment, while the parallel model concatenated premise and hypothesis sentence vectors.
- Performance impact: Both NN models achieved high accuracy on the whole and empirical easy test sets but showed drastic performance drops on the empirical hard test set.The reported drops indicate that much of their high accuracy benefits from the empirically easy subset.
- Performance impact: When premise words were replaced with unknown symbols, both models still performed above chance on the empirical easy set.Removing premise information preserved performance above the chance ratio, linking the easy-set behavior to TE-label prediction from hypotheses.
- Interpretation: The models worked as RTE models for hard examples but as TE-label prediction models for easy examples, contrary to their constructors’ expectations.This behavior is presented as the practical consequence of hidden bias in the SNLI training data.
- Corpus comparison: SNLI and SICK were highly similar in sentence domains and descriptive statistics, although their TE-label distributions differed.Both corpora originated from Flickr30k scene descriptions, while SNLI labels were balanced and SICK labels were not.
5. Conclusion
The paper introduces an empirical method for examining large RTE corpora and finds hidden bias in SNLI that enables TE-label prediction from hypotheses without premises. It further reports that this bias substantially affects NN-model performance and can make models behave differently from their intended RTE function.
- Conclusion: The method investigates large RTE-corpus quality by testing whether TE labels are predictable from hypothesis sentences without premise context.It uses a two-phase procedure based on a null hypothesis and a Naive Bayes model.
- Conclusion: SNLI contains hidden bias that allows TE-label prediction from hypothesis sentences without premise information.The experimental result does not reject the proposed null hypothesis for SNLI.
- Conclusion: A large portion of NN models’ high RTE accuracy benefits from this hidden bias.The paper reports that SNLI-trained models can function as TE-label prediction models rather than RTE models on biased test pairs.