3 ms·

Is it feasible to program a machine learning system (such as Watson?) to run a straightforward legal case, with a natural language interface? The law is full o

Is it feasible to program a machine learning system (such as Watson?) to run a straightforward legal case, with a natural language interface? The law is full of context and interpretation details. One example my business constantly runs into is weather or not we have to register as a 'foreign business" when we do business in other states (my business is an engineering consulting business that does work in many states, sometimes requiring travel, sometimes not). All of the states we looked into have the same basic wording tin the foreign business statue, but they interpret it different - for some states, you only have to register is you maintain an office presence; for some states you have to register only if you spend a significant amount of time in the state, other states you have to register if the money originates in the state, even if you never set foot there. I'd love to see someone with ML or AI experience address the challenges involved in this. I would imagine it would be possible, but require a shitload of work, and even then may require some review.

wfn 14y ago
If the rules of interpretation (one might even say that this is the definition of 'context' in this sense) can be strictly defined (formalized), e.g. what you wrote: > for some states, you only have to register is you maintain an office presence; for some states you have to register only if you spend a significant amount of time in the state, [...] then this is doable, in principle. But the overall project and required workload would indeed be vast.. I wonder how much one might have to work to have some kind of a working PoC.. NLTK and their computational linguistics book / python tutorial might be a good start. e.g. http://nltk.org/book/ch10.html http://nltk.org/book/ch10.html - we can do things like dt = nltk.DiscourseTester(['A student dances', 'Every student is a person']) dt.add_sentence('No person dances', consistchk=True) - throws error 'Inconsistent discourse ..[list of predicates etc]' dt.retract_sentence('No person dances', verbose=True) - outputs Current sentences are s0: A student dances s1: Every student is a person But scaling this / making it work for something more than these simple linguistic tasks.. quite an undertaking! In any case, the field in question (and the task at hand) is very interesting, and I hope that lots of interesting stuff will be done in comp.linguistics / nlp in the near future.
femto 14y ago
From my reading, context and interpretation were two domains that IBM specifically had to target with Watson [1], to get it to process natural language. If Jeopardy is solved, legal advocacy might be a next harder problem? Like Jeopardy, the task involves a lot of "soft" information, but results in an easy to measure pass/fail decision. It could be run in real-time, as an advocate in a courtroom, or off-line as a legal adviser. Also like Jeopardy, it offers the opportunity of a public contest, and the associated PR. [1] http://www.research.ibm.com/deepqa/faq.shtml http://www.research.ibm.com/deepqa/faq.shtml --- Edit: It seems like at least one person inside IBM is already thinking this way. To quote [2]: "Law, Patent and Trademarks" "Not only could a Watson-like capability minimize filing through existing databases on laws, prior cases, rulings, hearings, opinions, it could also be used as a method of testing witness questions. or suggest a series of inquiries and questions for litigation. It could be used to simulate certain judges, prosecution and defense lawyers, based on prior cases." "A Watson-like system could generate questions for a prospective patent claim based on it's ingestion of the entire patent and trademark database." [2] https://www-304.ibm.com/connections/blogs/davidian/entry/where_does_ibm_watson_go_after_jeopardy7?lang=en_us https://www-304.ibm.com/connections/blogs/davidian/entry/whe...

wfn 14y ago

If the rules of interpretation (one might even say that this is the definition of 'context' in this sense) can be strictly defined (formalized), e.g. what you wrote: > for some states, you only have to register is you maintain an office presence; for some states you have to register only if you spend a significant amount of time in the state, [...] then this is doable, in principle. But the overall project and required workload would indeed be vast.. I wonder how much one might have to work to have some kind of a working PoC.. NLTK and their computational linguistics book / python tutorial might be a good start. e.g. http://nltk.org/book/ch10.html http://nltk.org/book/ch10.html - we can do things like dt = nltk.DiscourseTester(['A student dances', 'Every student is a person']) dt.add_sentence('No person dances', consistchk=True) - throws error 'Inconsistent discourse ..[list of predicates etc]' dt.retract_sentence('No person dances', verbose=True) - outputs Current sentences are s0: A student dances s1: Every student is a person But scaling this / making it work for something more than these simple linguistic tasks.. quite an undertaking! In any case, the field in question (and the task at hand) is very interesting, and I hope that lots of interesting stuff will be done in comp.linguistics / nlp in the near future.

femto 14y ago

From my reading, context and interpretation were two domains that IBM specifically had to target with Watson [1], to get it to process natural language. If Jeopardy is solved, legal advocacy might be a next harder problem? Like Jeopardy, the task involves a lot of "soft" information, but results in an easy to measure pass/fail decision. It could be run in real-time, as an advocate in a courtroom, or off-line as a legal adviser. Also like Jeopardy, it offers the opportunity of a public contest, and the associated PR. [1] http://www.research.ibm.com/deepqa/faq.shtml http://www.research.ibm.com/deepqa/faq.shtml --- Edit: It seems like at least one person inside IBM is already thinking this way. To quote [2]: "Law, Patent and Trademarks" "Not only could a Watson-like capability minimize filing through existing databases on laws, prior cases, rulings, hearings, opinions, it could also be used as a method of testing witness questions. or suggest a series of inquiries and questions for litigation. It could be used to simulate certain judges, prosecution and defense lawyers, based on prior cases." "A Watson-like system could generate questions for a prospective patent claim based on it's ingestion of the entire patent and trademark database." [2] https://www-304.ibm.com/connections/blogs/davidian/entry/where_does_ibm_watson_go_after_jeopardy7?lang=en_us https://www-304.ibm.com/connections/blogs/davidian/entry/whe...