Click here for the full version of the presentation

Talk: Minerva 2000

My Name is Hamutal Kreiner, and the work I am going to present was performed in collaboration with prof. Asher Koriat.

A great number of reading studies concern the processes that enable comprehension to occur on-line. Text comprehension requires the coding of structure, as well as the coding of meaning. (slide 2) According to the structural precedence hypothesis, the processing of structure precedes and paves the way for the analysis of meaning. Early in text processing, readers attempt to derive a structural frame for the text. This frame is assumed, to guide the processing of content, and to facilitate its assimilation into a coherent semantic representation. The computation of structure is carried out on-line, on the basis of shallow analysis of the text, taking advantage of structural cues, such as function morphemes and punctuation marks. Once a structural frame has been extracted, attention shifts from structure to meaning. Most of the evidence in support of this hypothesis comes from work on letter detection during silent reading.

In this study, however, we focused on another reading behavior (slide 3) prosody, that is, the rhythm, intonation, and stress applied when reading aloud. Reading prosody is produced in real time, therefore we believe that the study of reading prosody might shed some light on the real time processes that enable the analysis of text on-line. The experiments I am going to present here examined the proposition that reading prosody is derived on the basis of the syntactic structure of the sentence, and is relatively independent of its content.

Experiment 1

According to the structural approach to reading, on-line text processing is guided by early structural analysis. (slide 4) The first experiment examined the hypotheses that reading prosody, produced on-line, is derived on the basis of the structural processing, and that it is relatively independent of semantic analysis. In order to do that we examined how prosodic patterns applied by readers are affected by syntactic and semantic manipulations.

Participants were recorded, when reading aloud sentences, their readings were analyzed and pause patterns at major syntactic boundaries were measured. We examined only one aspect of reading prosody -- the pause pattern. A pause-pattern was then derived for each sentence, and the effects of structural and semantic manipulations on the pause patterns were examined.

The following examples (slide 5) demonstrate how we created some of these manipulations (this is not very good English because I translated the sentences almost word by word from the Hebrew version in order to get a clear demonstration). One of the semantic manipulations, for example, consisted on creating a normal sentence like A. Then, by destroying the coherent meaning of the sentence but maintaining its structure, we created a matched morpho-syntactic nonsense sentence like B.

In this figure (slide 6), you can see averaged pause patterns - that is, pause duration as a function of its ordinal position within the sentence. These are pause patterns of normal versus morpho-syntactic nonsense sentences, and as you can see they are quite similar. The semantic manipulation practically didnt affect the reading prosody.

In contrast, (slide 7) when we manipulated syntactic structure by using sentences with completely different structures like these two sentences, (slide 8) the pause patterns applied by readers were clearly differentiated.

However, when structural information is eliminated from these two sentences, by eliminating all function morphemes and using a random word order like in these strings (slide 9) (read the sentences). Then,
(slide 10) pause patterns become monotonic the variance of pause duration is much lower they have nearly the same duration, and the differentiation between the two patterns is clearly reduced.

All together, the results indicated that the pause patterns applied by readers are highly sensitive to the structure of the sentence, yet relatively indifferent to its content. It seems that reading prosody evolves on the basis of the sentence structure, independent and possibly prior to complete analysis of content.

Experiment 2

Some additional support to this conclusion is gained in a second experiment we ran. Assuming, that reading prosody evolves on the basis of structure, we examined if it can be activated through syntactic priming.
(slide 11) In this experiment we used garden path sentences, where structure is temporally ambiguous (read sentence). The prosodic patterns produced by readers in immediate reading of sentences like this are often not fluent, or not coherent. While in non-ambiguous sentences prosody seems to evolve on the basis of structure, it appears that in ambiguous sentences in which the reader cannot figure out the sentence structure on-line, he doesnt manage to produce the appropriate prosody in real-time. Hence, we hypothesized, that in normal sentences, where structure is processed bottom-up, syntactic priming would not affect reading prosody. However in garden-path sentences where complete structural processing is postponed, syntactic priming might affect the production of reading prosody.

We recorded subjects, reading ambiguous and non-ambiguous sentences, and then asked independent judges to rate their reading prosody. However, before reading each target sentence, subjects read 4 prime sentences. (slide 12) Half of the target sentences were primed by sentences with the same syntactic structure, whereas the other half were primed by sentences consisting of the alternative syntactic interpretation.

(slide 13) As you can see, indeed, syntactic priming had no significant effect on regular, non-ambiguous sentences, however, it yielded significant effect in ambiguous sentences. It seems, then, that reading prosody was affected by the activation of syntactic representation. These results join the results of the previous experiment suggesting, that reading prosody is produced on the basis of structural representation. They also seem to support the idea that a syntactic representation that is independent of semantics is computed at some stage in the process of text analysis.

Experiment 3

It might be argued that reading prosody is only produced when reading aloud and that prosody is not inherent to the process of reading itself, and therefore cannot teach us much about text processing. (slide 14) However, 2 lines of research converge in suggesting that this is not the case. First, previous research in speech perception, suggests, that listeners dont only use the prosodic information provided in the acoustic stream as a key to other linguistic information, but rather prosodic representation is maintained in memory as part of the surface representation of the sentence. Presumably this representation serves as an initial organizing structure in the immediate memory of the sentence, until more abstract and integrative representation is construed.

The second line of research concerns the role of phonological representations in reading. Evidence such as the phonological similarity effect, the word length effect and the articulatory suppression effect were interpreted by Baddeley as an indication of the cardinal role of phonological store not just in spoken language comprehension but also in reading comprehension. Putting together these two lines of research we propose that reading prosody is not merely an end-product of reading aloud, but rather prosodic structure is construed during reading and maintained in memory as part of the phonological representation of the sentence. The third experiment takes a first step in exploring this proposition.

We asked subjects to read sentences like this one (slide 15). This sentence can be realized with two prosodic versions, however it was always presented in context that favored one of the prosodic versions (slide 16). Later they had an auditory recognition test. Half of the test sentences were new and half were old, and participants were asked to say if they read that sentence in the first phase or not. In the test, half of the old sentences wrealized in a prosodic version that is consistent with that context, while the other half were presented in a different prosodic version, both prosodies were natural. (slide 17) We found that when the auditory presentation in the test is consistent with the prosodic version appropriate to the reading context - recognition rate was higher, and reaction time was faster. Note, that half of the subjects performed the reading phase by reading aloud and the other half performed it in silent reading. Both reading fashions yielded a significant advantage to consistent prosody. This finding seems to support the hypothesis that a prosodic representation is created during reading and maintained in memory. An alternative interpretation, however, is that the prosodic structure applied in the test activates one of the possible syntactic representations, and therefore the consistency advantage that we see is due to the shared syntactic representation and not to the prosodic representation activated in the process of reading. Future experiments are planned to try to disentangle these interpretations.

Conclusion and Preview

(slide 18) Just a few words of conclusion and preview:

Reading prosody seems to evolve on the basis of the structural schema of the sentence. Therefore, it might serve as a new tool, in addition to letter-detection and eye-movement, to examine the hypothesis of structural precedence, proposed by Koriat and Greenberg. The results of the first two experiments indicate that reading prosody is produced relatively independent, and possibly prior to complete semantic analysis. Hence they seem to support the hypothesis of structural precedence.

(slide 19) We are now going a little further and proposing that reading prosody is not merely an end-product of reading aloud, but rather a prosodic representation is formed during reading (even silent reading) and maintained in memory. The third experiment was the first step in that direction. Well further examine the proposition that this prosodic representation is activated by the structural representation of the sentence. Finally, we speculate that a prosodic representation is not just an epiphenomenon, and that it has some functional value it might serve as an intermediate representation that assists in maintaining the structural frame in working memory while semantic processing is carried out. This hypothesis will be examined in future experiments.


Click here for the full version of the presentation
Back to Third Advisory Board Meeting

home5.gif (2471 bytes)