Aspects of Cognitive Parsing

With respect to grammatical processing, it seems to be evident that humans do not compute in advance all possible readings of an utterance before some sort of disambiguation is performed in order to determine the ``best'' reading. Rather it seems to be the case that humans prefer a reading as most appropriate in a given situation. Therefore, most of the cognitive processing strategies follow some kind of deterministic strategy. These approaches can further be classified into those which are strongly deterministic and those which are quasi-deterministic. In quasi-deterministic models, a first reading of a sentence (or phrase) is constructed. If at a later point this preferred first reading is rejected because of syntactic, semantic or pragmatic reasons a second analysis is computed by means of re-analysis. In strongly deterministic models, no re-analysis is done.

In [Hemforth1993] several prominent cognitive models are discussed and empirically evaluated with respect to the following questions (under particular consideration of the German language):

• How do the models deal with structural ambiguities?
• How are word order preferences explained?
• When is lexical information taken into account?

Based on the empirical analysis presented in Hemforth's work, she deduces the following conditions for a cognitive plausible model of language processing:

• The human language process is competence-based; i.e., grammatical knowledge is fundamentally used for the analysis of language. Processing of grammatical knowledge is performed automatically and efficiently.
• Humans seem to process the grammatical structure of an utterance incrementally from left to right, i.e., words are integrated into the existing structure as fast as possible.
• Based on the experimental results presented in [Marslen-Wilson1976] , there is evidence that during the processing of a single utterance syntactic and semantic knowledge is already used in common to handle local disambiguations. Together with the incremental nature of the parsing process Hemforth concludes that sign-based approaches seem to be most appropriate, especially ([Hemforth1993], page 180).
• In the case of structural ambiguities only one alternative is considered during further processing using preferences. If the chosen alternative has to be rejected the process of re-analysis is activated. Furthermore, the re-analysis is purposive and efficient.

Hemforth concludes that the only strategy that is consistent with the empirical evidence she had obtained through her experiments is the mixed bottom-up/top-down left-corner strategy of the Earley algorithm (see [Hemforth1993], page 176).

Purely data-driven (bottom-up) and goal-directed (top-down) strategies are not consistent with the empirical results provided in Hemforth's work. For bottom-up strategies there is no prediction possible about items of the next input structure, although experiments have shown that humans actually use such predictions for disambiguation. Next, for right-branching structures no incremental processing is possible, because a new constituent can only be built if all parts are completely determined. Therefore the necessary amount of memory of keeping partial results is very high.

A top-down strategy allows the integration of new items into the sentence structure immediately, and each lexical item can be predicted on the basis of knowledge constructed so far. Therefore, disambiguation of lexical ambiguities as well the processing of right-branching structures can be made efficiently. However, simple top-down backtracking strategies have the well-known left-recursive problems and the use of backtracking would not explain why the re-analysis process is so efficient.

Hemforth's results support the uniform architecture as follows. First, we also follow the Earley strategy. Processing is performed left-to-right if the grammar defines string concatenation in that way. The algorithm is performed as an on-line strategy, which means that words of a string are consumed next by next. However, the incremental behaviour is restricted, because we do not pass the words of a string one by one as input to the uniform tabular algorithm (see also the notes about fully incremental text processing made in the previous section).

The uniform algorithm is directed by an agenda, and we have shown that this allows for depth-first and even preference-based strategies, although the latter has not yet been used. The underlying grammars used are sign-based. Lexical access is done as early as possible by taking into account a huge number of top-down predictions. In summary, our uniform algorithm reflects the requirements worked out in Hemforth's thesis and actually can be used as a basis for cognitive models - at least for parsing.