i cdnuolt blveiee taht I cluod aulaclty uesdnatnrd waht I was rdanieg. The phaonmneal pweor of the hmuan mnid, aoccdrnig to a rscheearch at Cmabrigde Uinervtisy, it dseno’t mtaetr in waht oerdr the ltteres in a wrod are, the olny iproamtnt tihng is taht the frsit and lsat ltteer be in the rghit pclae. The rset can be a taotl mses and you can sitll raed it whotuit a pboerlm. Tihs is bcuseae the huamn mnid deos not raed ervey lteter by istlef, but the wrod as a wlohe. Azanmig huh? yaeh and I awlyas tghuhot slpeling was ipmorantt!

I was surprised to read what he had to say about substitutability.   Substitutability is the capacity to replace phrases with other phrases that are of the same type.  For example ‘the dog’ in ‘the dog ran to town’ can be replaced with ‘it’.   So in some sense, the phrase ‘the dog’ and ‘it’ can be substituted for each other.  This is one of the cornerstones of linguistic theory and is used as a basis of many parsing techniques.  The PCFG parser uses probabilities for a phrase type that can be traded out in a given context.

However, Seginer makes the claim that substitutability is not required for his incremental parser.

Substitutability, the essential idea of the Harris method, which has been seen as a starting point for the induction process for so long, turns out to be unnecessary in unsupervised parsing.

 …unlabeled parsing which only requires the parser to identify the constituents (or dependency links) but does not require them to be labeled, is purely syntagmatic (by definition).   A parser induction algorithm can therefore focus on learning to detect syntactic units while ignoring substitutability. (p20)

Another fresh idea (to me) from Seginer’s paper is the skewness of language structure.

The syntactic structure of natural language is skewed. This simply means that when the syntactic structure of an utterance is represented by a tree, each node in the tree has at least one short branch. The shorter the shortest branch is, the greater the skewness.  (p22)

Essentially how the incremental parser takes advantage of skewness is to expect skewness in the parse result.  This reduces the search size and thereby make the parsing process more efficient.

…context free grammars, allow (a-priori) any tree structure and, therefore, a learning algorithm for such representations must discover by itself the skewness property of syntactic trees. However, if this property is indeed universal, there is no need to burden the learning algorithm with its discovery and it is possible to code skewness directly into the parser.  (p23)

He claims that ‘coding the skewness’ into the syntactic representation and the parser, i.e., expecting branches to be of mixed depths, does not retract from the accuracy of the parse result.

Here is a link to Yoav’s dissertation.  Dissertation

Since graduating in 2007, it looks like Yoav Seginer is working at a  small company in Amsterdam, Mondria Technologies Ltd (according to LinkedIn.com).  The company website doesn’t say anything yet.  I wonder if they are working on a project that uses the incremental parser.

Incremental Learning.  The system adds information to the grammar with each new sentence.  In other approaches (CCM by Manning and Klein, UDOP by Bod), the entire corpus is processed as a block to collect the parameters of the grammar and the entire corpus is repeatedly processed until the learning converges on final result.  In Seginer’s approach, the corpus is processed one sentence at a time.  After each sentence, the grammar weights are updated.  That sentence is not revisited for further training.

Simpler Math.  The approach by Seginer uses much simpler math for computing the parameters of the resulting grammar.   There are a few ratios, some accumulation of values into other, and some comparison of weights to choose which one to apply for a given step.  There are no long chains of probabilities to compute the best parse.  I don’t have any principled reason why this makes more sense for a model of how the brain works – it just fits my gut feeling better.

Not Restricted to Binary Trees.  The CMM approach and others approaches that are referenced in Seginer’s paper all give binary parse trees as the result.  But natural languages aren’t limited to binary trees.  Although it is possible to represent any non-binary tree as a binary tree, forcing a binary representation onto natural language is not adding to the value of the result.  Seginer’s process gives parse trees that include constructions of more than two nodes.

Exocentric Constructions.   A phrase like ‘the boy’ has links going both directions between the two words.  Either word can be considered the head of the phrase – which in some way matches the disagreement between linguists about which word is actually the head of the phrase.

Link depths.  His algorithm result include a depth value on dependency links.  This value can be used to distinguish between internal and external arguments to a phrase.

Incremental Parsing.  For each word read into the sentence, links can only be added to or from the new word.

             I know the boy (sleeps)

When the algorithm encounters the word sleeps it can only add a link to or from this word.  No links that were generated at earlier steps can be affected.   This reduces the search space and contributes to the overall speed of the parser.  Another advantage is that the links in ‘I know the boy’ are an exact subset of the links in ‘I know the boy sleeps’.  There is no decision necessary about dropping links that were previously discovered.

No Clustering.  Seginer’s approach groups words with class labels that in some respect are part of speech classes.  However, the approach does not require clustering of words into specific classes.  It finds similarities between the and a but it does not require that all determiners are grouped together in a well defined class.  This makes the approach better able to deal with noise in the training data.

Homophony.  When a word has more than one meaning, it can confuse any machine algorithm.  Seginer’s algorithm deals with homophony by comparing labels between target pairs of words.  In the example that follows, words in brackets are possible labels for the target words.  When an underscore appears in the label, it means look at  the label of the preceding or following word.

             This[the] year[_the]    (this is a determiner)

            This[is_] was [is]          (this is a pronoun).

In the example given above, in one case this is a determiner.  One of its labels is [the] because this and the occur in similar places.  In the second line a label for this is [is_] which means that this is frequently followed by is.  The algorithm finds pairs of labels such as [the] and [_the] or [is_] and [is] to decide where to put links.

I think Seginer’s approach has some real merit.  It is a greedy algorithm – it learns as it goes and does not need to revisit previous sentences.  It uses simple math.  It has results that seem to match psycholinguistic models.   It would be great to see this approach extended to take advantage of other linguistic phenomena such as morphology.

 A link to the paper can be found here.  Paper

 A link to a video presentation on the approach can be found here.    Video

However, it occurred to me that infants could rely on the heavy weighting of head directionality when they are first learning a language.  By choosing the direction that is predominant and ignoring the incoming utterances of the opposite direction, it would simplify the initial learning phase.

Perhaps someone has already proposed this little insight into child language acquisition.

As I was pondering this, it also occurred to me that the single most important characteristic of a language is that it can be learned by an infant.  If the majority of infants in a culture can’t learn the language of their parents, then that language will not persist.  It will die out.

Perhaps someone else has already proposed this little insight as well.

Chris Manning gave this talk at MLCS 2007.  It is a fairly detailed discussion of the work that Chris did with Dan Klein on learning syntax structure.  Dan used the material for his dissertation.

Chris Manning Video

Dan Klein gave this talk at UAI 2008.  It is more high level, and describes how Dan and his students have applied their work to a few other areas of unsupervised learning.

Dan Klein Video

Both these videos are worth watching.

However, this week I was reminded that there are limits to how much you can learn in a degree program that lasts only 5 quarters.   A blog at LingPipe (a review of my blog) (lingpipe-blog.com) pointed out several errors in what I had written about here in one of my posts.  Well, lucky for me I hadn’t intended the text to be submitted to a conference jury for review.  The paper was just a one-week assignment.  As I remember, I passed the class, so the paper probably met the requirements for the course even though it has very clear short comings in the eye of the LingPipe reviewer.

Maybe I will find the time to dig out the code that was used for that paper, and compare it against the evaluation at LingPipe and see if I can understand what the LingPipe objections were.  I might even try to re-implement the code following the expert’s directions.

Another way I look at this is: the more you know, the more you know that you don’t know.  It has been healthy for me to receive some constructive criticism of my post.  I have been writing software for only 30 years, so maybe in the next 30 years, I can learn how to fix the problems that LingPipe pointed out.  For now, I am just an average software engineer.  Code reviews are part of the process.

“Languages tend to have terms for contact, vertical alignment, attachment, containment, and proximity, as if there were a cognitive alphabet of spatial relationships more basic than the prepositions of a given language.”  (p. 178)

“A light bulb is considered to be in a socket when its base is inserted, since that allows it to be illuminated, but a person is not in a car if only his arm extends in through a window, since that doesn’t allow the car to move him or even shelter him.” (p. 187)  The meaning of the preposition in depends on the objects that are being described.

“If Sally has one big stone and Jenny three much smaller stones, who has more?  The question by itself is unanswerable: it depends on whether you mean “more stone”, or “more stones.”  (p. 173)   The meaning of more depends on whether it is referring to the number of objects (stones) or the mass/volume/weight of the object (stone).

“The part of the mind that interfaces with language treats objects schematically.  … Every morsel of matter has a length, a width, and a thickness, but when we speak of these morsels we pretend that some of the dimensions aren’t there. … A road, a river, or a ribbon is conceptualized as an unbounded line (its length which serves as its single primary dimension) fattened out by a bounded line (its width which serves as a secondary dimension), resulting in a surface.”  (pp. 179, 180)

 ”Since words and syllables aren’t free, languages economize when they can. … Imagine you are in a rainstorm, ten feet away from an overhanging ledge.  Move one foot toward it, you still get wet.  Move over another foot; you still get wet.  Keep moving, and at some point you no longer get wet.  Continue to move another foot in the same direction, you don’t get any dryer.  So nature has set up a discontinuity between the segment of the path where gradual changes of position leave you equally wet and the segment where gradual changes leave you equally dry.  And it is exactly at that discontinuity that one would begin to describe your position using under rather than near.”  (P 186)

“Spatial terms quantize space at the cusps where causal events play out differently on each side.  As your palm gradually [wraps] around a marble, the curvature at which you stop saying the marble is on the hand and start saying it’s in is more or less the shape that would prevent it from rolling off when you jiggle it.”  (P 186)

Dr. Pinker’s premise in this book is that language reflects our thoughts.  By disecting our language, we get a glimpse of how the thought engine behind the language works.  We use count nouns and mass nouns in language because our minds see countable items such as chairs or dogs and our minds also see non-countable mass objects such as water or furniture.  We use a preposition like along to describe proximity to a one-dimensional line and we use inside to describe containment in a two or three dimensional object. 

All languages take slightly different approaches to describing space, but there are similarities that can possibly be used to infer an underlying brain structure that helps define our language.  “Most of the world’s languages divide the space around the speaker into just two regions, though about a quarter of them (including Spanish) make a three-way distinction among ‘near me’, ‘far from me’, and ‘in between,’ and a very few go to four, adding ‘very far from me’.” (p. 178)  He is referring to the English terms here (near me) and there (far from me).

“Not all languages carve [spatial relationships] up in the same way.  Presumably this is because each language trades off expressiveness, precision, word length, and vocabulary size in a different way.  But the quantization of spatial relations is universal, and causally important relations like contact, attachment, alignment, verticality and proximity make their appearance in all the spatial vocabularies of the world.”  (p. 187)

The book is an excellent example of Dr. Pinker’s writing – it is entertaining while at the same time being specific and to the point.  He digs into the issues and comes at them from all aspects – cognitive psychology, neuro-science, pathology, and child language acquisition.  He is an academic, and at the same time he presents his material in a way that is concise and engaging.

So I chose to look at the TREC Question Answering challenge and read about some of the techniques that were used for submissions to that evaluation.  The kinds of questions that are used for this include simple factoid questions such as “How many calories are in a Big Mac?”, list questions such as “Which past and present NFL players have the last name of Johnson?”, and definition question such as “What is a Golden Parachute?”.

Many of the system descriptions submitted to TREC have pipelined architectures.  Here are some examples of components that were described:

• Question Type Classification – Decide what the answer should be like.  Is it asking for a name, a date, a description or a list, etc.
• Question Rewrite – Convert the question to search terms that are used for the IR search.  This step typically expands the list of words used in the search by adding synonyms.  This gives a broader list of pages returned that can then be further analyzed.
• IR Search – generally a simple search using established technologies.
• Passage Selection – deciding which portions of the text returned by the IR search to include in the results.
• Answer extraction and ranking – creating the actual text that will be returned as answers.

The paper is not intended to be a thorough description of the field.  It helped me get a handle on what current designs are being developed.

Here is the link to my exploration paper: QuestionAnsweringAtTrec

As part of the Master’s program in Computational Linguistics at the University of Washington, we are preparing for internship appointments in the summer of 2008. As part of that preparation, we were asked to pick a topic that concerns Computational Linguistics and write a short summary of a few papers that covered the topic. This is a PDF of that paper. PreInternship Topic

This test was run by ‘training’ a classifier on a set of data instances. Each instance has a vector of features. In the Bernoulli case, we treat the features as binary. In the multinomial case, we treat the features as a numeric value from 0 to n where n is the number of instances the given word was found in the instance. Both sets of data have a ‘class’ assigned to them. After we train on the data, we check our system results with the actual class value assigned to each instance. This is where the percent accuracy comes from.

Of course these numbers are from a single test on one set of training and test data, but the fact that the Multinomial results are 91% accurate compared to 88% accurate for the Bernoulli is pretty telling. Also you can see that the elapsed runtime for the test is significantly different as well. The ‘cross_prob_delta’ is a parameter for tweaking the ‘add-one’ smoothing in the training stage.

Bernoull

Cross prob delta	Training Accuracy	Test Accuracy	Wall Clock (seconds)
0.1	0.9303	0.8800	303.23
0.5	0.9103	0.8633	318.10
1.0	0.8970	0.8400	305.45
2.0	0.8796	0.8233	305.56

Multinomial

Cross prob delta	Training Accuracy	Test Accuracy	Wall Clock (seconds)
0.1	0.9570	0.9133	8.33
0.5	0.9503	0.9066	8.50
1.0	0.9448	0.9000	8.29
2.0	0.9400	0.8966	8.26

While we were working on this evaluation, we were brainstorming on other ways to compare the Bernoulli and Multinomial approaches. This is what we came up with.

• Office chair comparison. Put signs on two contestants that are sitting in office chairs. The signs are ‘Bernoulli’ and ‘Multinomial’. The contestants race down the hallway without lifting themselves out of the chair. The first sign to the finish line is the method of choice.
• Date getting comparison. Again, put signs on two contestants. Have them stand back to back in the center of the student lounge and randomly ask girls for dates. The sign that gets the most dates is the method of choice.

All kidding aside, the NLP stats processing class is challenging and fun. We are gaining insight into how these methods can be used for basic classification of data.