Science for SEO: machine translation

January 05, 2009

Query expansion using MT

The Google patent entitled "Machine Translation for Query Expansion" (25/12/08) is a really interesting read. It describes a really exciting new method for query expansion.

"Query expansion" is when a users query is modified before the search is performed. This is done to improve the search results. To do this techniques such as stemming, spelling correction, and the adding of synonyms are used.The method described deals with query expansion using synonyms. Usually this is done using thesauri or lexical ontologies but here it is proposed that machine translation be used - ingenious.

Synonym selection is really not that easy at all. WordNet and such resources have helped us a lot, but there's room for improvement. Sometimes a word can have several different meanings and choosing the wrong one would completely change the query.

"The method includes receiving a search query and selecting a synonym of a term in the search query based on a context of occurrence of the term in the received search query, the synonym having been derived from statistical machine translation of the term. The method also includes expanding the received search query with the synonym and using the expanded search query to search a collection of documents."

Google uses statistical machine translation (as opposed to the rule-based approach). This type of system includes a language model which is used to figure out which bit of text is in the target language and a translation model which uses certain probabilities to determine the translation. So it looks at the likelihood of a particular string being the translation of another. The language model tells the system which proposed translation coming out of the translation model is likely to be right.

"In general, in another aspect, a method is provided. The method includes receiving a request to search a corpus of documents, the request specifying a search query, using statistical machine translation to translate the specified search query into an expanded search query, the specified search query and the expanded search query being in the same natural language, and in response to the request, using the expanded search query to search a collection of documents."

The end result is that there is an increased likelihood hat the search results are more accurate. Also it limits the expansion of the query with erroneous words (which is nice technically speaking).

"Statistical correlations between the occurrences of words in the source language and words in the target language are expressed as alignments between particular words or phrases. When the target language and source language are the same natural language, the principal meaning of an aligned pair is the same. The aligned word or phrase pair is presumed to have similar meaning, i.e., they are presumed to be synonymous. For example, the word "ship" can be aligned under certain circumstances (e.g., in a particular context) with the word "transport". Thus, for those circumstances, "ship" is synonymous with "transport".

The Google translate system is far from accurate though, using their example:

User query: "How to ship a box"

Google translate French: "Comment une boîte de livraison"

Google translate German: "Wie Sie,ein Feld"

French:

Comment = How

une boîte = a box

livraison = delivery

German:

Wie = how

Sie = you

ein feld = a field

The synonyms and the context are pretty hard to get from these translations - obviously this is a really really simplistic and short test. It just gives and idea of the thing. It works better with "Achilles heel running injury" btw but...evaluation is not done like this it's a bit more complex.

Here basically the idea is to add more context awareness to the search system. I like it, it's very clever indeed. The Google translate engine is therefore capable of being put to use in other ways than just translation.

Why should you care?

Well this method shows that queries are being expanded to include far more words than are actually present in the query. This means that going after particular keywords may be useful at the basis but is a very limited approach. As an SEO expert, you should be seeking to create content rich with not only your top target keywords but also terms and concepts that belong to that topic. It's time to look at things in more dimensions than one.

November 11, 2008

About machine translation

Machine translation (MT) is all about translating text (or speech even) from one language to another. It's part of computational linguistics, and uses a lot of NLP methods as well as statistical methods, rule-based methods, corpus techniques, some AI too, amongst other things. Apparently it was started in the 17th century and in the 1950's the Georgetown experiment wen on, but it didn't really work so funding was really reduced, meaning that a lot of research in this area was terminated. In the 1980's it made a comeback.

It's important today, in the age of the Internet, because a lot of data is in different languages, and when we can't understand another language, we are deprived from what may be the most relevant content to our query.

First you have to pull apart the source text to make sense of it, and then you have to re-engineer it into the target language so it makes perfect sense to a target language reader. Not only do you have to understand all the grammatical elements, the syntax, the idioms, the semantics, and so on, you also have to have a good grasp of the culture associated with the target language.

Different systems use different approaches, here is a brief description:

Rule-based systems:

It's basically made up of a load of rules relating to translation between the two languages. It can use a dictionary and map to that. You can use a parallel corpus to find those rules, which means that you map between ready made translations and pick out the common patterns, then feed these into a machine. I did this and it wasn't very precise. Google used SYSTRAN for many years, and this is a rule-based system.

Statistical methods:

Google translate now works with these. It involves generating a load of statistics derived from a large corpus. The problem is finding a very large corpus. this isn't too much of a problem for Google, but not very many corpora exist, but even Google used the united nations corpus to add 200 billion words to it's system. These are used to train the system.

The main issues:

Word disambiguation is very difficult. This is when words can have more than one meaning. Google doesn't do so well in this area. There are 2 methods that are known to deal with this, the shallow approach (looking at surrounding words and drawing statistical information from this), and the deep approach (providing a comprehensive definition to the system). The deep approach takes a lot of time, and isn't so precise, so statistical methods tend to do better.

Consider this for example: "Cleaning fluids can be dangerous" - does it mean that cleaning fluids IS dangerous or that they ARE dangerous?

There are so many difficult issues in handling language anyway, seeing as it requires natural language understanding, which is far from performing right now. There is a lot of research going on though, and eventually machine translation will work, but I'm not so sure how soon that will be.

What does it mean for SEO?

Well your keywords and your content is going to look a lot different in other languages, and the text may also be modified and re-written in places. This means that you have a lot less control over how these pages rank in other languages. The solution? Maybe it would be worth having multi-lingual staff :)

Read more here, from the University of Essex.

There's also good information at Microsoft research (MT labs).

John Hutchins is a great source of information.

And check out Carnegie Mellon University MT labs too.