University Homepage School Homepage School Contacts School Search

Sample paper responses

Below are examples of paper responses, written by Sharon Goldwater, a previous instructor for the course. Samples 1 and 2 were written before this course and are probably a bit more technical than your responses will be. Nonetheless, they illustrate the basic idea of summarising the important contributions of the paper and responding to them, and are approximately the right length (on the upper end; your responses should not be longer than these). Sample 3 refers to the same paper as Sample 2, and is more similar to the level of technical detail you will be expected to show. Sample 4 also refers to the same paper, but is an example of a poor response.

Sample 1

This paper investigates the learnability of a particular linguistic phenomenon: the restriction of anaphoric "one" to refer to an N', rather than N0, constituent. The authors argue that this restriction need not be innate (as many linguists have argued), but instead can be learned from the input children actually receive, assuming that various other linguistic knowledge is already present (whether through innate endowment or learning). They use a Bayesian computational model to show that the syntactic category of "one" can be learned from child-directed speech. As the model sees more data, if no evidence has occurred to support one of the choices, then it becomes more and more probable that the other choice is correct. The data comes from post-nominal complements and modifiers, which the authors chose because previous computational work in this area had focused on pre-nominal adjectives, which presented some difficulties due to the ambiguity between N' and N0 in that case.

While the particular phenomenon in question is very specific, the broader questions addressed (What kinds of linguistic knowledge must be innate? Which can be learned?) are central to the study of language and cognitive science. Of course, showing that anaphoric "one" is learnable does not prove that domain-specific constraints are unnecessary - in fact, the authors' model includes some fairly domain-specific representations and assumptions. However, as the authors state, their goal is to reduce the learning problem from a very specific one (learn what anaphoric "one" refers to) to a more general one (learn that pronouns refer to antecedents, that language contains structure, etc.). This will then allow researchers to focus on the more general problem, asking what kinds of constraints are necessary to acquire this more general knowledge. I personally like this approach, and I particularly appreciate the authors' explicitness about the linguistic assumptions that are and are not built into the model.

Sample 2

In this paper, the authors compare the predictions of two different computational models of word segmentation (SRN and Parser) against experimental data. The models differ in their predictions of how sub-word units are treated: Parser predicts that competition between words and sub-words (sequences of syllables that form incomplete words) causes words to become more familiar with more exposure to training stimuli in a Saffran-style word segmentation experiment. Therefore discrimination between words and part-words (sequences containing the end of one word and the beginning of another) will improve with longer exposure, but discrimination between sub-words and part-words will not. In contrast, the SRN predicts that representations of both words and sub-words are strengthened, so both words and sub-words will be distinguished from part-words more easily after more exposure. These predictions are shown through simulations in Experiment 1. Experiment 2 presents the same data to human subjects, with the result that there is a trend toward improved discrimination between words and part-words after longer exposure, and no improvement for sub-words vs. part-words. Although the improvement in word vs. part-word discrimination is only a trend (p < .09), there is a significant interaction between exposure duration and type of item (word or sub-word vs. part-word). This indicates that words and sub-words do not respond to exposure duration in the same way, contrary to predictions of the SRN.

One thing I liked about this paper is that it explicitly compares two models within the same paper, and uses both computational modelling and experiments with humans. Also the authors connect the two specific models to two broader classes of models, which suggests broader implications: namely, that models that perform segmentation by "chunking" may be a better representation of human behavior than models that segment by "bracketing".

Sample 3

This is a rewritten version of Sample 2, to give you an idea of a response that isn't quite as technical, but would also be considered an acceptable response. It is clearly written, explains the important terminology and concepts, and provides a brief synopsis of the experiments and results. It mentions a broader implication about representation in memory, and it also includes a question that came to mind while reading the paper, but could have gone into more detail.

This paper is about modeling how infants might segment words from continuous speech. The authors look at two different computational models (SRN and Parser) and compare their results to experimental data. These models are treated as examples of two bigger classes of models: "chunking" models, which look for coherent sequences of input syllables to treat as words (Parser), and "bracketing" models, which look for incoherent pairs of syllables and insert boundaries between them (SRN).

The Parser model assumes direct competition in memory between words and sub-word units (sequences of syllables that form incomplete words), and makes the prediction that words will become more familiar as exposure to the training data increases. The SRN model has no notion of competition, and predicts that both words and sub-word units will become strengthened in memory with more exposure. The predictions are tested by having both the models and human subjects try to discriminate sub-words from part-words (sequences of syllables containing the end of one word in the beginning of another) after a training phase. The results support the predictions of Parser and not of the SRN, which the authors use to support the idea that word-like units compete with each other in memory, and more generally that chunking models are more plausible than bracketing models.

One question I thought of is how these models would respond to noise in the input data, for example if a syllable isn't heard correctly. Would they still work? Would they still make the same predictions?

Sample 4

This is another rewritten version of Sample 2, to give you an idea of a poor response. Without the second paragraph, it's probably worth a 0 out of 5; with the second paragraph, it might merit a 1 of 5. Note the lack of detail, and the fact that several terms are used without being explained. There is also no attempt to connect the results of the paper to any broader implications.

This paper is about modeling how infants might segment words from continuous speech. The authors look at two different computational models (SRN and Parser) and compare their results to experimental data. SRN is a neural network model and Parser uses chunks. The authors compare the models using training data from a nonsense language and compare sub-word and part-word units. Parser makes predictions that are more similar to the human responses.

One question I thought of is how these models would respond to noise in the input data, for example if a syllable isn't heard correctly. Would they still work? Would they still make the same predictions?