DBpedia Spotlight的confidence与support参数含义及配置建议问询

Great question—these two parameters are core to tuning DBpedia Spotlight for your entity recognition needs, so let’s break them down clearly:

Parameter Definitions #

• confidence: This is your confidence threshold for entity linking. Put simply, DBpedia Spotlight assigns a 0–1 score to how sure it is that a detected n-gram maps to a specific DBpedia entity. Only results with a score above your confidence value get returned.
  • Higher values = fewer false positives, but you might miss valid entities the model is less certain about. Lower values = more entities captured, but you’ll get more incorrect links mixed in.
• support: This refers to the number of in-links a DBpedia entity has (i.e., how many other DBpedia pages link to it). It’s a proxy for how "notable" or widely recognized the entity is within the DBpedia knowledge graph. Setting a support threshold filters out obscure, rarely-mentioned entities that don’t meet your significance bar.

Rule-of-Thumb Configurations #

The ideal values depend on your text type and goals (precision vs. recall), but here are standard starting points:

General Text (News, Blogs, General Web Content) #

• confidence: Start with 0.3–0.5.
  • Use 0.3 if you want to maximize recall (catch as many relevant entities as possible, even if some are slightly uncertain).
  • Use 0.5 if precision is your priority (minimize wrong links, even if you miss a few borderline valid ones). For strict precision, you can bump this up to 0.7, but expect a big drop in the number of entities returned.
• support: Start with 20–50.
  • 20 filters out most truly obscure entities while keeping mid-tier notable ones.
  • 50 is stricter, only returning entities widely referenced in DBpedia (great if you only care about major, well-known entities).

Domain-Specific Text #

• Academic/Technical Content:
  • confidence: 0.2–0.4 (domain-specific terms often have lower model confidence but are critical to your text).
  • support: 5–15 (many specialized entities have low DBpedia in-links but are highly significant in your field).
• Social Media/Short Text:
  • confidence: 0.4–0.6 (noisy, informal text needs higher confidence to cut down on false links).
  • support: 10–30 (filter out niche entities unlikely to be relevant in casual conversations).

Pro tip: Always test these values on a sample of your actual target text. Precision and recall are a tradeoff—tweak the parameters until you get the right balance for your use case!

内容的提问来源于stack exchange，提问作者J Cena