TL;DR
This paper investigates the quality of Natural Language Inference (NLI) annotations, revealing that even seemingly straightforward datasets can suffer from annotation inconsistencies due to linguistic phenomena like coreference and loose definitions. The authors propose that incorporating annotator justifications and analyzing specific linguistic features is crucial for building more reliable NLI datasets and developing truly explainable AI systems.
What Problem Does This Paper Address?
The field of Natural Language Inference (NLI), which aims to determine the semantic relationship (entailment, contradiction, or neutral) between two sentences, has seen a surge in large-scale datasets and impressive model performance. However, recent work has highlighted that these high performance metrics can be misleading due to "annotation artifacts" – subtle biases or inconsistencies in the data that models learn to exploit rather than truly understanding language. This paper addresses the critical question of NLI annotation quality, specifically focusing on how linguistic phenomena and the annotation process itself can lead to confusion and unreliable data, even in datasets designed to be simpler. They aim to identify why humans struggle with certain NLI tasks and how this can be improved for both human annotators and automatic systems.
Methods
The researchers conducted an annotation experiment using a subset of the SICK corpus. They provided their annotators (graduate students) with clearer, more detailed guidelines than those used for the original SICK corpus. Crucially, they also asked annotators to provide justifications for their NLI decisions and a "computational feasibility" score (estimating how likely an automatic system would get it right).
Following this, they performed a second, "meta-annotation" experiment where they themselves categorized pairs based on specific linguistic phenomena identified as potentially problematic: coreference difficulty, directionality of the inference, looseness/vagueness of definitions, atomicity of sentences, negation, and quantification. They then statistically analyzed the inter-annotator agreement (IAA) and computational feasibility scores in relation to these phenomena.
Datasets
The primary dataset used is a subset of the SICK (Sentences Involving Compositional Knowledge) corpus. The authors re-annotated a portion of SICK with improved guidelines. They also reference other NLI datasets like SNLI and Multi-NLI as context for the scale of data creation in the field.
Key Insights and Ideas
- "Human-level" is not always perfect: Even with clear guidelines, human annotators can disagree, suggesting that the ideal of perfect agreement might be unattainable for certain linguistic phenomena.
- Justifications are gold: Asking annotators why they made a decision is invaluable. It reveals ambiguities in guidelines, highlights difficult linguistic phenomena, and provides insights that simple agreement scores miss.
- Not all linguistic challenges are equal: Phenomena that are considered difficult for machines (like coreference) are also difficult for humans, but some phenomena that are easy for humans (like quantifiers) might be challenging for machines, and vice-versa.
- Directionality matters: The order of sentences in an inference pair can significantly impact agreement, suggesting that NLI datasets should consider this asymmetry.
- "Loose definitions" are a major source of confusion: Vague or subjective concepts lead to lower agreement, indicating a need for more precise definitions or handling strategies.
Key Results
- The re-annotation experiment achieved an overall inter-annotator agreement (IAA) of 73.25% with Cohen's κ of 0.64, which they deemed comparable to the calibration set and indicative of reasonably good guidelines.
- However, their annotators differed from the original SICK annotations in 17% of cases, raising questions about the original corpus quality.
- Statistical analysis revealed that hard coreference and loose definitions significantly decrease inter-annotator agreement.
- Directionality also negatively impacts agreement, with pairs in the "harder" direction showing lower IAA.
- Counter-intuitively, negation was found to increase IAA, likely because the dataset contained clear-cut examples of negation.
- For computational feasibility (CF) scores, hard coreference led to lower CF scores (meaning annotators thought it was harder for machines), while negation led to higher CF scores (easier for machines).
- They found no significant correlation between their annotators' perception of machine difficulty (CF scores) and actual machine performance on a subset of the data.
Limitations and Future Work
The authors acknowledge that their experiment was conducted on a small subset of SICK. They suggest that future work should:
- Trace and quantify similar trends in other large NLI datasets like SNLI.
- Further investigate the "loose definitions" category, potentially drawing on prior work.
- Develop improved guidelines that explicitly address phenomena like coreference and directionality.
- Explore methods for creating NLI corpora that account for inherent human variability and develop systems that can handle these complexities, potentially by incorporating justifications as features.
- Investigate how to integrate justifications into "static" learning systems for more explainable AI.
Summary updated