|
This page answers frequently asked questions (FAQs) for CS224N / Ling 237. Buggy command lines in PA3 handout?16 May 2008As some students have noticed, the example command lines in PA3 are full of bugs, some of the legacy (such as missing the "ir/" or having three 2's in "224n"), and some of them new (such as where we tell you to append "genia" et al. to the path information). They command lines have all been corrected in the PA3 handout which is available from the Assignments tab. What was the distribution of grades for PA1?5 May 2008Mean: 8.38 A minor bug in Hypothesis28 April 2008There's actually a minor bug in Hypothesis.java Line 121 (the assertion statement). Instead of assert(targetSentence.size() != newTargetSentence.size());It really should be: assert(targetSentence.size() == newTargetSentence.size());This shouldn't affect the way your code works. Previously we said it works with Sun JDK; it's probably because Sun JDK doesn't check the assertion by default, while gcj does. I will change this in the starter code. But don't panic. Whatever works for you should still work. To use the decoder in PA2, we have to implement getAlignmetProb which takes an "Alignment". What is its format? How's NULL represented?25 April 2008
In Model 1 and 2, each foreign word has to align to one English word (or NULL). In the internal representation of Alignment, the indices of sentences start from 0. If a foreign word is aligned to NULL, the index for the English word (NULL) will be -1.
If you're on bramble (or hedge), you should check if your "java" points to the sun jdk one. If not, you should do this: bramble13:~> setenv JAVA_HOME /usr/lib/j2sdk1.5-sun(Thanks Tague for noticing the problem and providing the commands) What kind of numbers should I expect for PA224 April 2008Here are some AER numbers to give you some general idea of what you might get from PA2: PMI: 0.79 Why is
24 April 2008
|
Unigram | Bigram | Trigram | Linearly-interpolated Bigram | Linearly-interpolated Trigram | |
Training set perplexity: | 849.9011 | 71.8341 | 9.853 | 91.9442 | 21.9336 |
Test set perplexity: | 817.5552 | 435.7855 | 2429.7497 | 228.0201 | 195.6814 |
HUB perplexity: | 3070.1622 | 10366.2218 | 33935.5494 | 2258.6445 | 2092.19 |
HUB Word Error Rate: | 0.0986 | 0.0757 | 0.0872 | 0.078 | 0.0734 |
Run "setenv JAVA_HOME /usr/lib/j2sdk1.5-sun/" and ant should work on hedge.
You do not have to create your own. Uncomment lines 227-228 in LanguageModelTester.java (after loading the training sentences, but before loading the test pairs) and you will then have a validation set to use.
You can use this set to choose parameters, such as the lambdas for linear interpolation (trying out hand-picked values is fine).
Some students who have been trying to investigate learning curves have reported seeing test-set perplexity increase as the amount of training data grows. This is counter-intuitive: shouldn't more training data yield a better model, which is therefore able to attain lower perplexity? Chris came up with a possible explanation involving the handling of the <UNK> token. Remember that the <UNK> token actually represents an equivalence class of tokens. As more training data is added, this equivalence class shrinks. Because the meaning of the <UNK> token is changing, model perplexities are not directly comparable. Especially when the amount of training is small, adding more data will rapidly lower the model probability of <UNK>, causing the entropy and perplexity of the model distribution to grow.
If you've been looking at learning curves, an interesting investigation — not specifically required for the assignment — would be to measure the learning curve while holding the definition of <UNK> constant. This would mean allowing the models trained on small data sets to "know about" all the words in the largest training set. All known words in this sense would get explicit counts, which could be 0, and then you'd still have an <UNK> token representing all words which did not appear in even the largest training set.
Generally speaking, smoothing should be the last step of training a model: first you collect counts from your training data, and then you compute a smoothed model distribution which can be applied to (that is, used to make predictions about) any test data.
In principle, it would be possible to postpone the computation of a smoothed probability until test time. But (a) it's not very efficient, because most smoothing algorithms require iterating through all the training data, which you shouldn't have to do more than once, and (b) if you're wanting to do this because your smoothing computation depends upon something in the test data, then you're doing things wrong. (For example, model probabilities should not depend on how many unknown words appear in the test data.)
Real sentences are not infinite; they begin and end. To capture this in your n-gram model, you'll want to use so-called "stop" tokens, which are just arbitrary markers indicating the beginning and end of the sentence.
It's typically done as follows. Let <s>
and
</s>
(or whatever) be arbitrary tokens
indicating the start and end of a sentence, respectively. During
training, wrap these tokens around each sentence before
counting n-grams. So, if you're building a bigram model,
and the sentence is
I like fish tacos
you'll change this to
<s> I like fish tacos </s>
and you'll collect counts for 5 bigrams, starting with
(<s>
, I
) and ending with
(tacos
, </s>
). If you
encountered the same sentence during testing, you'd predict its
probability as follows:
P(<s> I like fish tacos </s>
) = P(<s>
) · P(I
|<s>
) · ... · P(tacos
|fish
) · P(</s>
|tacos
)
where P(<s>
) = 1. (After all, the sentence
must begin.)
Q. Is it necessary to give strict mathematical proof that the smoothing we've done is proper probability distribution? Or is it enough to just give a brief explanation?
A. You should give a concise, rigorous proof. No hand-waving. I'll show an example on Friday. Note that it's important that your proof applies to your actual implementation, not some ideal abstraction.
Do you have questions regarding details of various smoothing methods? (For example, maybe you're wondering how to compute those alphas for Katz back-off smoothing.)
You might benefit from looking at a smoothing tutorial Bill put together last year.
For greater detail, an excellent source is the Chen & Goodman paper, An empirical study of smoothing techniques for language modeling.
Some people have the wrong idea about how to combine smoothing with conditional probability distributions. You know that a conditional distribution can be computed as the ratio of a joint distribution and a marginal distribution:
P(x | y) = P(x, y) / P(y)
What if you want to use smoothing? The wrong way to
compute the smoothed conditional probability distribution
The problem is that steps 1 and 2 do smoothing separately, so it
makes no sense to divide the results. (In fact, doing this might
even yield "probabilities" greater than 1.) The right way
to compute the smoothed conditional probability distribution
Here, there is only one smoothing operation. We compute a smoothed joint distribution, and compute everything else from that.
If there's interest, I can show a worked-out example of this in Friday's section.
(It would also be correct to compute all the conditional distributions before doing any smoothing, and then to smooth each conditional distribution separately. This is a valid alternative to smoothing the joint distribution, and because it's simple to implement, this is often the approach used in practice. However, the results might not be as good, because less information is used in computing each smoothing function. This would be an interesting question to investigate in your PA1 submission.)
A few people have inquired about smoothing and unknown words (or more generally, n-grams). The basic idea of smoothing is to take some probability mass from the words seen during training and reallocate it to words not seen during training. Assume we have decided how much probability mass to reallocate, according to some smoothing scheme. The question is, how do we decide how to allocate this probability mass among unknown words, when we don't even know how many unknown words there are? (No fair peeking at the test data!)
There are multiple approaches, but no perfect solution. (This is an opportunity for you to experiment and innovate.) A straightforward and widely-used approach is to assume a special token <UNK> which represents (an equivalence class of) all unknown words. All of the reallocated probability mass is assigned to this special token, and any unknown word encountered during testing is treated as an instance of this token.
Another approach is to make the (completely unwarranted) assumption that there is some total vocabulary of fixed size B from which all data (training and test) has been drawn. Assuming a fixed value for B allows you to fix a value for N0, the number of unknown words, and the reallocated probability mass can then be divided equally (or according to some other scheme) among the N0 unknown words. The question then arises: how do you choose B (or equivalently, N0)? There is no principled way to do it, but you might think of B as a hyperparameter to be tuned using the validation data (see M&S p. 207).
Both of these approaches have the shortcoming that they treat all unknown words alike, that is, they will assign the same probability to any unknown word. You might think that it's possible to do better than this. Here are two unknown words you might encounter, say, on the internet: "flavodoxin" and "B000EQHXQY". Intuitively, which should be considered more probable? What kind of knowledge are you applying? Do you think a machine could make the same judgment?
A paper by Efron & Thisted, Estimating the number of unseen species: How many words did Shakespeare know?, addresses related issues.
Here are some estimates for what kind of numbers you should be seeing
as you do PA1:
Model | Test Set Perplexity | HUB WER |
Unigram | 1300 | .09 |
Bigram | 1200 | .075 |
Trigram | 300 | .06 |
Follow the instructions on the first page of the PA1 handout. On the second page, the instruction to "cd cs224n/java" is incorrect - there is no such directory (the handout has now been fixed).
You can run ant in one of two ways. From the directory ~/cs224n/pa1/java/ you can type "./ant" to use the symbolic link, or you can add the cs224n bin directory to your path variable: "setenv PATH ${PATH}:/afs/ir/class/cs224n/bin/apache-ant-1.6.2/bin/". Once you change the PATH variable, you can invoke ant by simply typing "ant" (note, however, that to compile the files you should still make sure you are in directory ~/cs224n/pa1/java/ so it can find build.xml).
ant will compile the .java files in the ~/cs224n/pa1/java/src folder and put the resulting .class files in the ~/cs224n/pa1/java/classes folder. To run these classes, you must make sure java knows to look in the classes folder:
Make sure you are in ~/cs224n/pa1/java/. Check to see whether you have anything in the CLASSPATH variable by typing "printenv CLASSPATH". If nothing was printed out, you can set the CLASSPATH by simply typing "setenv CLASSPATH ./classes". If something was printed, type "setenv CLASSPATH ${CLASSPATH}:./classes". This will preserve whatever was already stored in the variable.
You should now be able to run the class files. From ~/cs224n/pa1/java/ try typing "java cs224n.assignments.LanguageModelTester" or try out the the included shell script "./run".
No. You can use Perl, C, C++, or any other widely used programming language. Extra credit if you design a Turing machine to compute your final project. Double extra credit if you build a diesel-powered mechanical computer to compute your final project. Triple extra credit if you build a human-level AI capable of autonomously conceiving, executing, and presenting your final project.
|
Site design by Bill MacCartney |