HIRM assigns logp_score = -inf to hospital data

[ThomasColthurst]

This can be confirmed by running the integration tests at head:

• ./bazel-bin/pclean/pclean --schema=assets/hospitals.schema --obs=assets/hospital_dirty.10.csv --iters=5 ... Running inference ... Starting iteration 1, model score = -nan Starting iteration 2, model score = -inf Starting iteration 3, model score = -inf Starting iteration 4, model score = -inf Starting iteration 5, model score = -inf

[ThomasColthurst]

At least in iteration #1, the nan's are coming from a subset of the relations: Condition, MeasureName, HospitalName, CountyName, Address1, PhoneNumber, HospitalType, and City. So basically all the "~ string" variables.

This points to a problem with the bigram_string distribution, since the other part of a "~ string" variable is the bigram string emission, which is shared with several other variable types, like typo_int or stringcat.

[ThomasColthurst]

More confirming information from running on the flights data: once again in the first iteration, all of and only the "~ string" variables get -nan values -- flight, act_arr_time, act_dep_time, sched_dep_time, and sched_arr_time.

[ThomasColthurst]

Adding max_length to the bigram distribution solved most of this. Now, only two typo_int fields are causing problems in the hospital data: Hospital:zip and Hospital:provider.