0
$\begingroup$

In one of my classes, the lecturer explains the tagging scheme BIO (begin, inside, outside) which I knew of and I perceive as the standard. Then he introduced BIEWO which additionally has a tag for whole, so it's only one word and that entity is the whole thing; and end, which denotes an entity is the end of a tagged sequence.

Where is the difference for a ML algorithm? OWO should be equivalent (I mean actually equivalent) to OBO; it encodes the same information. Same goes for e.g., OBIIO vs. OBIEO. You can write simple rules to transform one into the other. Is my claim true? If no, when would I use which? What's the advantage of one over the other?

$\endgroup$

1 Answer 1

1
$\begingroup$

In your question, you’re thinking only about modeling the sequence itself. But your goal is to generate that sequence for a given piece of text.

The features of a single-word named entity may be different from those of a multi-word named entity. Using BIEWO lets you decouple these and model them with different feature weights. Conversely, it may actually be useful to share that information—in which case BIO tagging makes sense.

The choice of tag set is one of many you can make when defining your model. Because you only care about the spans at the end of the day—and either encoding captures these—be sure to evaluate the spans, not the particular tags applied to each word.


(Also, including this for better discoverability by search engines: BIEWO is more commonly called BIOES or BILOU.)

$\endgroup$
2
  • $\begingroup$ I still don't understand the difference in terms of information. I found this question stackoverflow.com/questions/17116446/…. The answers and discussion around it, and especially Vladislav's experiment point to that there is no significant difference. $\endgroup$ Jul 18, 2021 at 16:51
  • $\begingroup$ Vladislav admits, “My experiment was on one dataset only and may not be representative.” / There is no difference in the information in the sequences. There is a difference in the features that a model can use. For instance, a feature like (word is Northern and tag is B) would likely have higher weight than (word is Northern and tag is W). $\endgroup$ Jul 18, 2021 at 19:19

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.