TL;DR
This paper investigates "common sense bias" in modern Semantic Role Labeling (SRL) systems, particularly those based on large language models like BERT. They demonstrate that while these systems excel with typical sentences, they falter when presented with semantically unusual but grammatically valid constructions, revealing a reliance on learned "common sense" associations rather than pure linguistic understanding. The authors introduce a new challenge dataset, COMTE, to specifically probe this bias.
What Problem Does This Paper Address?
Semantic Role Labeling (SRL) aims to identify the predicate-argument structure of sentences, essentially answering "who did what to whom." Modern SRL systems, powered by models like BERT, have achieved impressive performance gains. However, like many NLP systems, they can inherit biases from their training data. This paper focuses on a specific type of bias: "common sense bias." This isn't about demographic biases, but rather the tendency of models to perform poorly on sentences that express unlikely states of affairs, even if they are grammatically correct and semantically interpretable. The authors argue that this bias stems from language models encoding conventional associations that often align with human "common sense," leading them to err when confronted with counterfactual or unusual scenarios. This is problematic because real-world language use isn't always conventional, and robust NLP systems should handle such variations.
Methods
The researchers evaluate three SRL parsers:
- Björkelund et al. (2009): A supervised, log-linear parser using traditional features and a syntactic parser.
- Stanovsky et al. (2018): A supervised, deep, linear-time parser based on GloVe embeddings and LSTMs.
- Shi and Lin (2019): A self-supervised (and supervised) deep, linear-time parser based on BERT.
Their core methodology involves creating simple, three-word transitive sentences (Noun-Verb-Noun) where the verb strongly prefers animate subjects and objects. They then systematically vary the noun arguments across different categories:
- Person names: Align with common sense expectations of animacy.
- Country names: Can sometimes be personified.
- Abstract nouns: Lack animacy.
- Random plural common nouns: Lack specific animacy expectations.
- Random strings: Completely out-of-vocabulary and semantically meaningless.
By comparing the performance of the SRL systems across these argument types, they aim to quantify the extent to which each parser is sensitive to common sense bias.
Datasets
- CoNLL 2005 benchmark: Used to report baseline performance for the three SRL systems (Table 1).
- NAMES library, WorldMap, YourDictionary, Princeton WordNet: Sources for generating arguments (person names, country names, abstract nouns, common nouns) for their experimental sentences.
- COMTE (Challenge of Common Sense Bias): A newly introduced dataset of 1,000 simple, three-word sentences. Each sentence has a gold standard analysis where the second word is the predicate, and the first and last words are its A0 and A1 arguments, respectively. The predicates are chosen from six verbs that strongly select for animate subjects/objects. The dataset is specifically curated to include sentences that both the Shi and Lin (2019) and Stanovsky et al. (2018) parsers failed to parse correctly, making it a probe for common sense bias.
Key Insights and Ideas
- Modern SRL systems exhibit "common sense bias": The BERT-based parser (Shi and Lin, 2019) shows a significant drop in performance when arguments are abstract or random, compared to when they are person names. This suggests it relies heavily on learned associations rather than purely linguistic cues.
- Older systems are more robust: The older, non-embedding-based parser (Björkelund et al., 2009) shows remarkably consistent performance across all argument types, indicating a more robust handling of unusual sentence structures.
- "Common sense" is encoded in language models: The paper highlights that what we perceive as "common sense" is often deeply embedded in the statistical patterns of large text corpora used to train models like BERT.
- Probing bias requires carefully constructed data: Standard benchmarks, often derived from news articles, may not expose this type of bias because they tend to contain conventional language. New datasets like COMTE are needed to specifically test for it.
Key Results
- The BERT-based parser (Shi and Lin, 2019) performed significantly worse on sentences with abstract nouns (13.3% error) and random nouns (31.0% error) compared to person names (7.7% error).
- The older log-linear parser (Björkelund et al., 2009) showed near-uniform performance across all argument types, with errors ranging from 0.000% to 0.010%.
- The LSTM-based parser (Stanovsky et al., 2018) showed surprising sensitivity to country names and even improved with random strings, a behavior not fully explained by the authors.
- The COMTE dataset consists of 1,000 sentences designed to be trivially simple but difficult for modern parsers due to their unusual semantic content.
Limitations and Future Work
- Limited scope of verbs: The experiments primarily focus on a small set of six verbs that strongly prefer animate subjects and objects.
- Synthetic data: The COMTE dataset is synthetically generated, which may not perfectly reflect the nuances of naturally occurring unusual sentences.
- Unexplained behavior: The authors note they found no clear explanation for the Stanovsky et al. (2018) parser's peculiar performance with country names and random strings.
Future work could involve expanding the COMTE dataset to include a wider variety of verbs and sentence structures, and investigating methods to mitigate common sense bias in SRL systems, perhaps by explicitly training them to handle counterfactual or semantically unusual inputs.
Summary updated