Methods for meaning extraction

Question

Let someone wants to know what some word (concrete as "chair" or abstract as "happiness") mean. What methods, experimental techniques are there for extracting word's meaning? I found next ways:

1. Study appropriate entries at various dictionaries (defining vocabulary, philosophical/psychological/sociological dictionaries)
2. Just ask enough native speakers about what researched word mean
3. Study of context in which the word occurs
4. Make componential analysis
5. Make semantic differential experiment
6. Make linguistics association experiment

It looks as if these methods have significant disadvantages:

• Dictionaries are often too insufficient; if some really give a full description (that is rare), then same question: how the compilers get them?
• Asking native speakers looks very unreliable, subjectively way.
• Study of context is very widely. I didn't find any good examples of extracting meaning by this way.
• Componential analysis has no any rigorous algorithm.
• Semantic differential gives only connotative, emotional part of meaning.
• Linguistics association is too indirect method and requires too much interpretation.

What other methods are there for the full extracting researched word meaning? In the end I want have complete description of researched word's meaning.

Answer 1

The key point is the definition of meaning. This definition is probably not limited to linguistic. In fact many applications learn what we could call the meaning of things in completely different ways and for completely different objectives.

In purely linguistic approaches the followings come to my mind:

• Clique detection in synonymy graph. Defining a word by its set of pure synonym
• Semantic structure of word paraphrases. You can for example (among many) take a look at MTT. But also leveraging sentence semantic structure for word disambiguation...

You can also see the problem in term of image processing, vision. One can learn the meaning of the word chair only looking at pictures of chairs. Many algorithms exist for object and scene classification:

• spatial pyramid of bag of visual word, is one of them
• More recently deep learning methods like convnets...

You can also see meaning identification as a classification problem. One can argue that nobody knows the real true meaning of "chair", "conscience" or "love", but our ability to distinguish between two concepts of different meaning is essential. Applications and techniques used vary but it then falls down to a classification problem with well known algorithms (CART, logistic regression ...). The purpose can be again disambiguation of polysemic word, sense cluster identification, machine translation. In this case the meaning of a given word is a just a point in a machine learning model.

Answer 2

You listed a good number of methods, and some of their disadvantages. There are also other methods, but judging from your previous comments I'm really not sure whether you will be satisfied by any of them. Part of the reason for this, I presume, is that you first have to define what meaning is. Then you can choose a method and evaluate it. Take a look at the Wikipedia entry on semantics, and you might also want to read a good textbook or two.

Regarding you question on dictionary compilation, for example, you would have to define what you consider to be a complete entry. Maybe you would want it to list all the separate meanings of a word. But what is a separate meaning, what is metaphorical? In other words, when do speakers stop using a word in a metaphorical sense and start using it in two different meanings? Besides, what about very unusual meanings? I'm not sure there's a point in including a meaning that occurred twice in 1847 - but where do you draw the line? In 1884, the Oxford English Dictionary recognised the whole project might take a little longer than they thought - they had just reached the lemma ant, after spending five years on what was intended as a ten year project. The choices you make for a dictionary ultimately depend on the resources available and the aims (few people feel the need to consult the actual multi-volume OED, it's not a general purpose dictionary).

If you have a concrete enough definition of meaning, you can define concrete and rigorous methods. Prototype theory, for example, holds that meaning is best described in a fuzzy sense, with the prototype of a category conforming completely to the definition, while more peripheral members of the category are not prototypical but still belong to the category. Based on this you can run experiments with native speakers. For example, you present cylindrical containers made out of glass, and with different heights and diameters to people, some have a handle, some do not. If you ask native speakers for different variants of this whether they're seeing a vase or a glass, you will probably find that a handle increases the likelihood that it's a glass or cup. If the object has a very small diameter and is tall it is more likely to be a vase etc.

Distributional Semantics is an approach in Computational Linguistics. It's based on the premise that the context of a word determines its meaning. If you have a large enough corpus you can model this context, and compare it to another corpus (say, of another dialect of the same language). In this way you can compare differences in meaning between the same word in two or more dialects.