-
Notifications
You must be signed in to change notification settings - Fork 4
LkbSppp
The Simple PreProcessor Protocol (SPPP) provides a way of hooking up to an external tokenizer and morphological analyzer; optionally, SPPP also allows inclusion of part of speech (PoS) information in the preprocessor output, but as of March 2005 back-end support to take advantage of PoS assignments on input tokens (e.g. in the treatment of unknown words) is lacking. SPPP assumes that the preprocessor runs as an external process to the LKB that communicates with its caller through its standard input and output channels. Communication is in XML, using a designated control character for synchronization, and always in UTF-8 encoding (in other words, the encoding value in the XML document header is hard-wired for communication in both directions). SPPP allows ambiguous preprocessing output, for example alternative morphological analyses for one token. All references to LKB-internal parameters and functions in the following are within the lkb package, i.e. users need to make sure they are setting things up in the right package (aka namespace).
For hooking up external tokenizers and morphological analyzers, you may also be intrested in SMAF (see SmafTop).
Below is an example of the SPPP representation that might be returned from the preprocessor for the input kim sleeps. (and assuming the inflectional rules as found in the ERG). The top-level XML document is a segment object that contains a sequence of token entities. Each token encodes the string that is its surface form and the character start and end ranges (from and to, respectively) corresponding to it; character ranges are zero-based, counting characters (not bytes) in the original input string. Preprocessing results are keyed off individual tokens in the form of one or more analysis entities embedded in tokens. Where appropriate, an analysis can embed one or more rule entities corresponding to orthographemic rules.
<?xml version="1.0" encoding="utf-8"?>
<segment>
<token form="kim" from="0" to="3">
<analysis stem="kim"/>
</token>
<token form="sleeps." from="4" to="11">
<analysis stem="sleep">
<rule id="plur_noun_infl_rule" form="sleeps"/>
<rule id="punct_period_rule" form="sleeps."/>
</analysis>
<analysis stem="sleep">
<rule id="third_sg_fin_verb_infl_rule" form="sleeps"/>
<rule id="punct_period_rule" form="sleeps."/>
</analysis>
</token>
</segment>
^L
Each analysis needs to supply at least one bit of information, viz. the stem resulting from morphological analysis. Where zero morphology suggests that the stem and form values are identical, nothing else is required. For morphologically more interesting cases, the rule elements embedded in each analysis encode a chain of one more orthographemic operations that relate the analysis stem and token (surface) form. Each rule element needs to provide the rule identifier (the name of a lexical rule in the LKB) and the corresponding intermediate form. In the example above, one of the analyses is to derive the surface form sleeps. from the underlying stem sleep by virtue of chaining the punct_period_rule and third_sg_fin_verb_infl_rule lexical rules.
Other, optional attributes of the analysis entity are probability and tag to encode PoS tagger output, but in the current LKB back-end processing these remain unused.
To enable SPPP-based preprocessing for the LKB, you will need to provide a binary of the preprocessor that the LKB can invoke as an external process. For each sentence being analyzed in the LKB, the system will send an XML document to the standard input of the preprocessor, terminate the input with a formfeed character (`^L' or ASCII code 0x0c), and then wait for the preprocessor output to be printed on the standard output of the external binary. Again, to allow synchronization between the two processes, the LKB will look for another formfeed character to terminate the XML document returned by the preprocessor. Following is an example input
<?xml version="1.0" encoding="utf-8"?>
<text>kim sleeps</text>
^L
Note that the formfeed characters sent to the preprocessor input and required following its output are, strictly speaking, not part of the XML itself. They rather serve to simplify the protocol, in that each side of the SPPP protocol can rely on a formfeed to mark the end of a turn. Also, remember that SPPP communication is hard-wired to UTF-8, i.e. in generating XML output for a preprocessor, the external binary will need to make sure to use the UTF-8 encoding on its standard output stream.
Assuming a suitable preprocessor binary is available, its location in the file system needs to be declared to the LKB. Assuming the binary was called sppp and added to the platform-specific LKB binary directory, then the following would point the LKB to it:
(setf *sppp-application* (namestring (merge-pathnames bin-dir "sppp")))
Alternatively, users could make sure to have the binary directory in their shell search path (and then just specify the name of the executable file itself, without directory information), or just use a fully-qualified, valid path name, e.g.
(setf *sppp-application* "/usr/local/bin/sppp")
Once the external preprocessing binary is know to the LKB, the function call
(initialize-sppp)
will activate external preprocessing. Try parsing a sentence and observe the results. Typically, the setting of *sppp-application* could be done as part of the globals.lsp in an LKB grammar, and calling initialize-sppp() might be done in the grammar load script. To stop using SPPP (and go back to the default LKB-internal processing), use the function shutdown-sppp().
Given the asynchronous communication, it can be tedious to debug the initial creation of an SPPP interface. Once the communication works, it will likely be very stable, but getting both sides of the protocol to cooperate can be a bit tricky at first. Following are a few suggestions for debugging the initial SPPP set-up, in the unlikely event that it did not just work out of the box.
There is a sample shell script to implement a fake external preprocessor for the ERG. The script is called sppp and resides in the lkb/src/glue/ sub-directory of the DELPH-IN tree (assuming your installation includes source code for the LKB; see the LkbInstallation pages). Maybe start out by loading the ERG and making sure you understand the SPPP basics by activating the fake SPPP script; note that, no matter what input you parse, this script will always return the example XML shown above as the preprocessor output.
Once you know that the LKB is actually calling your own external binary, make sure the formfeed-based synchronization works. It may be worth augmenting your own binary with some code to log its activity to a separate file (note that logging must not occur on standard output, as that will be read by the LKB as preprocessor results). On the LKB side, try calling something like
(sppp "kim sleeps.")
Should the sppp() call not return, then most likely the preprocessor is not including a trailing formfeed to terminate its output. In case the function does return but reports errors, e.g. problems parsing the XML, then inspect the value of the global variable *sppp-input-buffer* to confirm the output from the external process actually adheres to the SPPP protocol. Should everything else fail, consult the source code in sppp.lsp, which is in the same directory as the fake sppp script.
(inferred from the above example - UlrichSchaefer):
<!ELEMENT segment ( token )* >
<!ELEMENT token ( analysis )* >
<!ATTLIST token form CDATA #REQUIRED
from NMTOKEN #REQUIRED
to NMTOKEN #REQUIRED >
<!ELEMENT analysis ( rule )* >
<!ATTLIST analysis stem CDATA #REQUIRED
probability NMTOKEN #IMPLIED
pos NMTOKEN #IMPLIED >
<!ELEMENT rule EMPTY >
<!ATTLIST rule id NMTOKEN #REQUIRED
form CDATA #REQUIRED >
Home | Forum | Discussions | Events