In standard Earley parsing the parser expects to find exactly one instance of an S nonterminal generating the entire input. This expectation is reflected by the fact that the chart is initialized with the dummy start state
For robust parsing, we want to identify all nonterminals that can possibly generate any substring of the input. This can be accomplished by also placing dummy states
for all positions k and nonterminals X, in the Earley chart prior to the start of normal operation. (In practice, dummy states need to be added only for those nonterminals X whose expansion can start with the current input symbol. This technique is discussed in Appendix B.3.2.)
The immediate effect of these extra states is that more predictions will be generated, from which more completions follow, etc. After finishing the processing of the jth state set, the chart will contain states
indicating that nonterminal X generates the substring 
.
Table:
Robust parsing using the simple grammar from
Table 1.
(a) State sets generated from parsing the ungrammatical string
a circle touches above a square. Dummy states (those with empty
LHS) represent partial parses. States representing ``maximal'' partial
parses are marked with MAX.
Predictions that don't lead to
completions have been omitted to save space.
(b) Trace of wildcard state completions resulting in a list of
complete partial parses for this input.
Table 
(a) illustrates the robust parsing process
using the example grammar from Table 1
(p. ).
Probabilities in the extra states are handled as follows. The initial dummy states are initialized with a forward probability of zero. This will ensure that the forward probabilities of all extra states remain at zero and don't interfere with the computation of prefix probabilities from the regular Earley states.
Inner probabilities on dummy states are initialized to unity just as for the S start state, and processed in the usual way. The inner probabilities for the each substring/nonterminal pair can then be read off of the complete dummy states.
Viterbi probabilities and Viterbi back-pointers can also be processed unchanged. Applying the Viterbi-parse procedure from Section 5.1 to the complete dummy states yields Viterbi parses for all substring/nonterminal pairs.