When I realized that a dictionary, as a whole, could be viewed as a non -syntactical knowledge container, in other words, that there was  no new knowledge in it, I wondered how something that was such a perfect auto-referential source could actually be useful.  For since every word is defined by other words in the same language, looking up the meaning of a given word would lead you to other words and yet others and eventually back to the first word, the word whose meaning you set out to learn. This would be the case even if  the dictionary were bilingual, and the meaning of the word you wished to check was given in a second language. Thus all dictionaries are perfectly circular, closed, auto-referential sources. However, I discovered that the analytical knowledge in a dictionary comes from the net supporting words connected with each other. Thus  if “chair” is strongly connected with “table”, “office”, “dining room”, etc. it should be easy to map it to its equivalent in any other language. Of course this overlooks languages in which the word “chair” does not exist,  reflecting the lack of  a comparable object in the culture and thus the lack of a cognitive representation of it. But such instances are rare since most cultures share a basic physical reality and human experience.

Of course problems arise with  words like  “personne” in French, which  maps onto “persona” and “nadie” in Spanish,  a noun and an adjective respectively with completely different connections and different supporting nets. Or conversely, when the verb “gustar” from Spanish maps onto”gouter” and “plaire” in French.  So even when it seems that all words are surjective, the general case is not bijective, and that applies to homonyms too, which often creates ambiguities in translation. In other words, any word in any language has its equivalent in any other language, whether a single word in one langauge becomes two or more words in another, or whether two or more words become one after mapping– but one word could mean two completely different things in another language. However, the supporting network  would be able to uncover this fact and solve a possible ambuiguity based on  context by extending the word network to encompass the ambiguity. In other words, if a subnet cannot be uniquely mapped, extending it should eventually solve the ambiguity. What one would need  is a corpus big enough to build such a network once and for all and then simply make comparisons at the network level. This could work even for completely new or unknown languages, either dead or living, assuming that they share a part of our actual reality and hence some part of our mental representations  (In a sense this is what Champollion did when he deciphered the Rosetta stone– he discovered a partial mapping of a subnetwork of words from an unknown language - Egyptian - to a subnetwork of a known one - Greek ). In the final analysis, each language has a single unique network (changing slightly through time but remaining well connected and strong enough to make  it unique and recognizable while being isomorphic with that of any other language).  Thus an entire language could be identified by its fingerprint -its network.

Having  established that, what about mining the world of all possible meanings, the world of all possible translations, and the world of all possible ideas? We wouldn’t have the problem of distinguishing between a coherent idea and a non-coherent one since the network would provide some minimal coherence. Thus the net-into-the -net approach would give us a way of translating from word to word and from phrase to phrase and from idea to idea.

The above ideas would apply to unilingual dictionaries, lets say English-English. The analytical knowledge in them again comes from the net connecting the words, so even if someone does not know English at all I would say that he would be able, albeit with considerable difficulty, to learn English just by deducing the net connecting objects, in other words, by mapping his own mental  representations of objects onto words in the  English dictionary. In the process he could encounter some ambiguities, but the further he goes, the more of these he would be able to resolve. On the other hand,  speakers of those languages in which “chair” does not exist, both in the language itself and as a real object in the culture,   would be able to deduce what  a chair is by tracking its  relations with the objects they know and for which they do have mental representations and the phonemes to externalize them. So the problem of translation, which began with the mapping of word onto word and then phrase onto phrase  with statistical tools,  becomes with this approach a matter of  mapping net to net.  Indeed this seems to be the approach adopted  by Meaningful Machines http://www.meaningfulmachines.com/-.  Such ideas have been around for a while, for example at WordNet: wordnet.princeton.edu/, but they somehow remain old-fashioned even as they are shifting the paradigm.

These ideas could be carried to the limit by taking the sum total of human languages and enquiring into the mapping between such a network and our cognitive representations. Such a move would provide grounds  for rebutting the Chinese room argument, since in the end it does not matter whether someone inside the room has no knowledge at all of a language; he would be able to map what he is mechanically translating onto his own mental representations, generating what, according to the argument, could not be generated: understanding. Because Searle’s  idea was, as I recall, to build up a case against A.I. in terms of the Turing test.

My conclusion: Meaning resides in the net.

This post is  related to a previous post titled “Meaning against A.I.”: http://www.mathrix.org/liquid/?p=29

Last Modified : May 17th, 2008 Filed under : Complexity, Minds and Machines, Natural Language and Computational Linguistics, New Ideas Navigate : Previous post / Next post