IBM Watson Knowledge Studio – Annotating negative / negated mentions

When annotating mentions using Watson Knowledge Studio, one often has examples such as:

“I received no feedback from in response to …” or “I have never received any feedback”.

If I were to annotate the mention “feedback” in the above, it is an “negative” example, i.e. it refers to something that did not happen. There are two possibilities when creating a custom entity type system:

(a) Include the negator in the mention, i.e. “no feedback” is the mention. This clearly does not work in the second example, since there is no negator before the work feedback.

(b) Do not include the negator in the mention, but add an attribute to the mention using the mention class NEG (https://www.ibm.com/watson/developercloud/doc/wks/wks_t_ts_intro.shtml)

Clearly option (b) is the more general approach. However, once a model is trained one needs to be able to extract entities unseen examples. For this you have to use the Natural Language Understanding API (https://www.ibm.com/watson/developercloud/doc/natural-language-understanding/#entities).

When one uses this API, there doesn’t seem to be a way to extract the mention attributes, i.e. when I do entity extraction, how do I see that a mention is negated or not, since using approach (b) you do not include the negator as part of the mention?
,When annotating mentions using Watson Knowledge Studio, one often has examples such as:

“I received no feedback from in response to …” or “I have never received any feedback”.

If I were to annotate the mention “feedback” in the above, it is an “negative” example, i.e. it refers to something that did not happen. There are two possibilities when creating a custom entity type system:

(a) Include the negator in the mention, i.e. “no feedback” is the mention. This clearly does not work in the second example, since there is no negator before the work feedback.

(b) Do not include the negator in the mention, but add an attribute to the mention using the mention class NEG (https://www.ibm.com/watson/developercloud/doc/wks/wks_t_ts_intro.shtml)

Clearly option (b) is the more general approach. However, once a model is trained one needs to be able to extract entities unseen examples. For this you have to use the Natural Language Understanding API (https://www.ibm.com/watson/developercloud/doc/natural-language-understanding/#entities).

When one uses this API, there doesn’t seem to be a way to extract the mention attributes, i.e. when I do entity extraction, how do I see that a mention is negated or not, since using approach (b) you do not include the negator as part of the mention?

Related:

Leave a Reply