Question SEMPRE vs SPF

[silentrob]

Can anyone explain the differences semantic or high level goals between SEMPRE and SPF[1]

They both seem to use a CCG and Lambda calculus to convert natural language to a logical form. I know SPF has a ML layer for getting to the solution while hiding some of the parsing internals.

Perhaps they projects target sets of use-cases and I simply overlooked it.

[1] https://bitbucket.org/yoavartzi/spf

[yonatansito]

At a very high-level both are semantic parsers that are developed by different research groups, so there is some substantial overlap. I don't know how to exactly compare the two, I guess we focus on training parsers from denotations rather than logical forms, but I recommend going through our tutorial to get a better sense of what's available.

On Wed, Mar 11, 2015 at 10:56 AM, Rob Ellis notifications@github.com wrote:

Can anyone explain the differences semantic or high level goals between SEMPRE and SPF[1]

They both seem to use a CCG and Lambda calculus to convert natural language to a logical form. I know SPF has a ML layer for getting to the solution while hiding some of the parsing internals.

Perhaps they projects target sets of use-cases and I simply overlooked it.

[1] https://bitbucket.org/yoavartzi/spf

— Reply to this email directly or view it on GitHub https://github.com/percyliang/sempre/issues/36.

[percyliang]

Things on which SPF excels:

• Better support for inducing lexical entries (e.g., GENLEX).
• Better support for lambda calculus and CCG.

SEMPRE has the following that SPF lacks:

• You can use different linguistic formalisms (you can use CCG, a more conventional S -> NP VP grammar, or custom rules).
• You can use different logical forms (you can use lambda DCS or Java programs).
• There is more support for execution of logical forms (e.g., Virtuoso or Java).
• There is a powerful "grammar" file which allows you to flexibly control the space of logical forms you generate, and more generally tweak your experimental setup.
• There is a interactive repl and a web interface which allows you to play around with the semantic parser, change the model on the fly, call Java, etc. (see the TUTORIAL).
• More documentation.

There are also other differences ranging from software engineering decisions to semantic parsing modeling decisions, which are probably too technical to go into.

Finally, SEMPRE is under active development and is the foundation of several semantic parsing projects at Stanford, so expect it to improve and become more featureful over the next few years.

[silentrob]

Great answer - thank you! Big plus one for the Documentation, Tutorial and REPL.

[evbo]

@percyliang any thoughts you can share on comparison of sempre vs ccg2lambda?