Understanding the nuances of human action in Natural Language Processing (NLP) requires more than just identifying verbs. It involves determining volitionality—whether an action was intentional—and subject animacy—whether the actor is a sentient being capable of intent. For instance, "taking a shower" is a volitional act, while "taking criticism" is generally non-volitional. Similarly, in the sentence "The white motorcycle chased the suspect," a human understands that the "motorcycle" refers to an animate entity (a police officer), whereas a basic AI might only see a machine.
Researchers from Kyoto University, in a study presented at AAAI-22, proposed a minimally-supervised framework to classify these properties jointly, overcoming the need for massive, expensive human-labeled datasets.
The Challenge: Context and Data Scarcity
Classifying volitionality and animacy is difficult because these properties are highly context-dependent. There is no static, exhaustive list of "intentional verbs." Furthermore, building a comprehensive dataset to cover every linguistic edge case is prohibitively expensive. The challenge is to create a system that understands the deep context of an event without requiring thousands of hours of manual data entry.
Methodology: Joint Learning with Regularization
The proposed framework uses a heuristic approach to label a massive raw corpus using a small "seed" lexicon of indicating words. To prevent the model from simply memorizing these seed words—a form of "cheating" that ignores broader context—the authors implemented several regularization techniques.
The core techniques include:
- Word Removal (WR): Training the model on versions of sentences where the indicating words are deleted.
- Explanation Regularization (SOC): A technique that penalizes the model if its prediction relies too heavily on a single "trigger" word.
- Adversarial Domain Adaptation (ADA): Forcing the model to find features common to both the labeled (with triggers) and unlabeled (without triggers) data.
- Consistency Loss: Ensuring logical consistency between tasks (e.g., an inanimate object cannot perform a volitional act).
Datasets and Seed Lexicons
The model was trained on a raw corpus of 30 million documents from the CC-100 dataset in both Japanese and English. The "minimally-supervised" aspect comes from using a tiny set of indicating words to bootstrap the learning process.
For evaluation, the researchers created a "gold standard" dataset of 1,200 events per category, labeled by crowd workers on Yahoo! Crowdsourcing (Japan) and Amazon Mechanical Turk (US).
Key Insights and Results
The joint learning approach proved highly effective, with the model successfully learning to infer semantics from context rather than just keywords. Key findings include:
- Synergy through Joint Learning: Learning volitionality and animacy together improved performance for both tasks. The logical link between "being alive" and "having intent" provided a powerful constraint that helped the model learn faster.
- Language-Specific Benefits: In Japanese, a "pro-drop" language where subjects are often omitted, the Word Removal technique was particularly effective. It forced the model to infer animacy directly from the verb alone.
- Qualitative Success: The model correctly distinguished between identical verbs in different contexts, such as the Japanese verb abiru used for both "taking a shower" (volitional) and "getting criticism" (non-volitional).
Limitations and Future Work
Despite its success, the model still faces challenges with "stative" verbs (e.g., "to stay" or "to exist") where intent is extremely subtle. Future research may look beyond single clauses to discourse context—the preceding and following sentences—to further resolve ambiguity. Additionally, the researchers noted that human agreement on these labels can be low, highlighting the inherent complexity of these semantic properties.
Paper Citation: Hirokazu Kiyomaru and Sadao Kurohashi.
- Minimally-Supervised Joint Learning of Event Volitionality and Subject Animacy Classification. In Proceedings of the Thirty-Sixth AAAI Conference on Artificial Intelligence (AAAI-22).
Summary updated