3
$\begingroup$

In this paper, the authors say that they used IO schema instead of BIO in their dataset, which, if I am not wrong, means they just tag the corresponding Entity Type or "O" in case the word is not a Named Entity. What advantage does this method have? I would imagine that it just takes away valuable information from the model and makes it harder to detect entities that span multiple words

$\endgroup$

2 Answers 2

4
$\begingroup$

To my knowledge, there is no clear best among the different labelling schemes variants for NER: IO, BIO, BILO (L=last), BILOU (U=unique, for a unique word)... I might forget some.

In theory at least, the advantage of a simple scheme like IO is that the simplified set of labels may prevent the model from making mistakes cases where a word or type of word can appear in any position of an entity, for instance. A more complex scheme can be more accurate but is more likely to lead to overfitting, because the sample of cases for one class is smaller.

If possible, it's preferable to evaluate the different options and pick the best for specific target data.

$\endgroup$
0
$\begingroup$

With IO schema if there are 2 distinct entities together then the system won't be able to differentiate between them say for example

Ukraine I-LOC 
Russia I-LOC
war O
president O
Volodymyr I-PER
Zelenskyy I-PER
...

Here in this example the IO schema fails. As if we consider multiple spans as an entity then Ukraine Russia together would be a location, and if Ukraine and Russia is considered as single entities then Volodymyr and Zelenskyy would be a different Persons.

$\endgroup$
1
  • $\begingroup$ So, if it is BIO format (AKA IOB1), Russia should be B-LOC. Because according to en.wikipedia.org/wiki/…, "The B- prefix before a tag indicates that the tag is the beginning of a chunk that immediately follows another chunk without O tags between them. It is used only in that case: when a chunk comes after an O tag, the first token of the chunk takes the I- prefix" $\endgroup$
    – ARAT
    Sep 22, 2023 at 9:35

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.