TL;DR
This paper introduces "inspiration sets" as a novel way to help human moderators quickly update rules for data programming pipelines used in e-commerce product moderation. The findings suggest that moderators can significantly improve classifier training data quality using these unsupervised inspiration sets, even with limited time, without needing pre-labeled examples.
What Problem Does This Paper Address?
E-commerce platforms rely heavily on text classifiers to identify and remove inappropriate products, a task that is constantly evolving due to new policies and emerging trends. Training these classifiers requires up-to-date labeled data, which is traditionally a time-consuming and labor-intensive process of manual item-by-item labeling. This paper addresses the challenge of enabling product moderators, who possess domain expertise but may not be machine learning experts, to rapidly and effectively steer the training data creation process for these classifiers. The goal is to achieve "fast control" over moderation classifiers, allowing them to adapt quickly to new threats, such as the rapid removal of offensive items during political events or the emergence of fake cures during a pandemic.
Methods
The core of the paper's approach is a data programming pipeline, building on the Snorkel framework. Data programming uses "labeling functions" (LFs) – essentially human-defined rules – to programmatically generate training data labels, bypassing the need for extensive manual labeling. The novelty lies in the introduction of "inspiration sets." These are carefully selected subsets of unlabeled data presented to moderators. The moderators then use their domain knowledge to refine or create new LFs based on these inspiration sets.
The paper explores three strategies for generating these inspiration sets:
- Set 1: Abstain-based strategy: Randomly samples data points that were not covered by any existing LF. This is an unsupervised method.
- Set 2: Disagreement-based strategy: Randomly samples data points where existing LFs produced conflicting labels. This is also unsupervised.
- Set 3: Classifier-based strategy: Samples data points where a trained classifier made the largest errors. This is a supervised method.
The study investigates whether unsupervised methods (Set 1 and Set 2) can be as effective as supervised ones (Set 3) in improving the quality of the training data for moderation classifiers.
Datasets
The study was conducted in collaboration with professional moderators at bol.com, a large European e-commerce company. The data used for training and evaluation consisted of product listings from six specific categories deemed problematic: fur, illegal wildlife related, magnetic balls, weapon-grade knives, smoking-drug-related, and single-use plastic. The paper provides statistics on the number of data points in the training, development, and test sets for these categories. The "ground truth" for evaluating the quality of the generated labels was created by domain experts (the moderators themselves).
Key Insights and Ideas
- Domain experts can drive data programming without pre-labeled data: Professional moderators, armed with their domain knowledge, can define initial labeling functions that achieve a surprisingly good starting point for classifier training, without needing to manually label any data beforehand.
- Unsupervised inspiration sets are effective: Moderators can significantly improve the quality of the training data using inspiration sets that are generated without relying on ground truth labels. This is a crucial finding for real-world scenarios where obtaining ground truth is difficult or expensive.
- Fast control is achievable: The proposed approach allows for rapid iteration and improvement of moderation classifiers, enabling platforms to react quickly to evolving policy violations.
- Moderators tend to add and modify, not delete rules: Interestingly, moderators primarily updated existing rules or added new ones, rarely deleting rules, even if they didn't immediately improve performance. This suggests a cautious approach to rule modification.
Key Results
The study reports that using inspiration sets, even unsupervised ones, led to improvements in data quality, as measured by the F2 score (which balances precision and recall, with a focus on recall). For most categories, a 30-minute session with an inspiration set (Set 1 or Set 2) improved the F2 score. For instance, the "magnetic balls" category saw a significant boost in F2 score with Set 1. The "knives" category, starting with very low quality data, showed some improvement with the supervised Set 3. The paper also notes that while inspiration sets improve coverage, they don't typically achieve full coverage of the data.
Limitations and Future Work
The authors acknowledge that the study was conducted with a relatively high level of abstraction due to the sensitive nature of the work. They couldn't reveal exact statistics of inappropriate items or the specific rules used. A limitation is that inspiration sets were drawn once and not updated iteratively within a single session.
Future work could explore:
- Understanding the interaction between rules and why moderators are hesitant to discard them.
- Investigating the ability of moderators to predict rule interactions.
- Developing more sophisticated methods for selecting inspiration sets.
- Exploring the impact of continuous moderation on moderator expertise and rule updating.
Summary updated