1

Consider the following example sentence:

"For eight weeks, the police say he was tortured by X. "

Many native speakers would attribute "eight weeks" to "torture". The question is now how do you inform computers about this?

Dependency parse doesn't work here (it connects the temporal mention to "say"): https://demo.allennlp.org/dependency-parsing/NzMwNzE3

Similar observation with respect to semantic roles: https://demo.allennlp.org/semantic-role-labeling/NzMwNzE5

Any thoughts or suggestions are appreciated.

  • The problem with your sentence is that it is not correct English. You could write: "The police say he was tortured by X for eight weeks" or "For eight weeks the police have been saying that he was tortured by X" (which obviously has a different meaning). – fdb May 6 '19 at 17:05
  • 1
    Sure, that's a good point. But any native speaker would understand my sentence. – Daniel May 6 '19 at 19:53
  • If the speaker intended the duration to apply to 'say', they wouldn't have used habitual tense. – amI May 7 '19 at 6:07
2

I see no remedy short of a human-oriented semantic analysis which can zero in on what would typically be important to a person in various situations. If I modify your example to say

"For eight weeks, the police wonder whether he was tortured by X."

now the ordinary interpretation has the duration adverb going with "wonder" rather than with "torture".

| improve this answer | |
  • 1
    Smart point. Looks like it is a well-known problem. A colleague pointed me to "pp-attachment" problem. – Daniel May 7 '19 at 3:00
2

This is a hard problem... many vendors do a poor job of attaching prepositions (including Spacy and Allen AI), but they're all getting better. We use statistics tables to fix the attachment issues (as a post-processor). We've cached the stats of all words in the dictionary for verb-to-pp and noun-to-pp combinations.

See: https://s3.amazonaws.com/wordnetparaphrase/verb-prep-stats.html

This simple trick cleans up most of the issue. Unfortunately, in your example "say" and "torture" have attachment rates that are very similar. The result can be further enhanced if you perform PP Disambiguation. There are other clues, such as PP's very rarely attach to the verb that precedes an infinitive marker.

This isn't an intractable problem, it just requires some work on the part of the vendor/provider.

| improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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