Wednesday, May 09, 2012

Google Translate met my challenge?

In the last footnote of the previous entry, I wondered if the infamous ambiguity of natural language would easily defeat algorithms that don't have deep comprehension of the symbols. I contrived two different English sentences with distinct meanings despite sharing many words: 1) "I'm quitting my run because I can't stand the sun directly overhead," and 2)"Too much overhead is required to run a stand". I'm feeling chagrin after I stuck the two into Google Translate. It produced French translations that don't appear to have grievous errors. Then again, my first language is English and my French is...limited. But my consultations of online French-to-English dictionaries haven't shown any serious problems. For an English speaker who doesn't know any French, the words' discernible resemblances aren't hard to see. I'll admit it. I'm impressed by Google Translate's use of context. Will Google's automated cars do well on Canadian roads?


UPDATE later: And...here's a counterexample of GT not working as well. "The child put the toy in the pen" is a classic conundrum for machine translation. "Stylo" is not the appropriate French word in this case. Changing "pen" to "playpen" works better, though. L'enfant mis le jouet dans le stylo.

Tuesday, May 08, 2012

the renovated Chinese Room

I have already posted about the Chinese Room two times. But after reading some comments in The Symbolic Species by Terrence W. Deacon, I think I can offer some more clarification or express the same perspective in a different way. Although Deacon states his belief that artificial intelligence is possible, he also opines that the Chinese Room casts doubt on the equivalency of sentience and algorithms for manipulating dissociated symbol tokens. These algorithms belong to systems of logical deductions, which function well in many domains but not as realistic models of brain computations. I agree with him. In fact, I shall renovate the Room around his observation.

As always, the bemused human alone in the closed room receives a sequence of unknown1 tokens (i.e. symbolic representations), applies a preexisting set of manipulation rules onto that sequence to form a response sequence, and finally sends back the response. However, this time, the testers outside the room are shrewd in their determination to detect the presence or absence of a sentient conversationalist. They carefully note the close correspondences of questions and responses. Gradually the growing collection of notes resembles the collection of originating rules in the room, which the unknowing manipulator can neither disobey nor adjust.

Eventually, the testers select and "combine" two suitable questions from out of the notes. By "suitable" I mean that via the syntactic and semantic structure of the unknown language, one of the selected questions modifies the meaning of the second. The meaning of the combined question isn't a "lossless sum" of the two questions' independent meanings2. Therefore an appropriate and believable response to the combined question isn't in the set of acceptable responses to either of the two independent questions. If the testers receive a believable response, then they may try combining two other questions. Or combine the combined question with a third. Or combine it with another combined question. For mischievous testers, the possibilities multiply.

For instance, since the algorithm in the room is intended to mimic a human3, it surely has a prepared response to the question, "What food does your father like to eat?" The alleged algorithm somehow uses the tokens of that question to produce the tokens of a response like "Licorice". Again, the algorithm surely has a prepared response to a distinct question of the same general category, "What food do you like to eat for breakfast?" Will the algorithm produce a reasonable response to the modified question, "What food does your father like to eat for breakfast?" Or, "What food does the father of your father like to eat for breakfast?" Or, "What food does your father not like to eat and you like to eat for breakfast?"

Regardless of the questions' triviality, the interpretive complexity can increase dramatically through tweaks that are rather simple (for humans!). As always, the Chinese Room asks, "Assuming that the fraud in the room achieves a perfect outcome from the testers' perspective, isn't it a self-contradiction for them to claim that anything in the room understands?" But after my renovation of the testers' methods, successful deception is much more difficult for the proposed algorithm! No matter how long the algorithm's finite list of prearranged question and responses, the testers can still escape it through valid combinations of any suitable questions in the list. And recall that the tokens of the right responses for the combinations aren't derivable from the tokens for the responses for the independent questions. If the algorithm is one big list of token transformations, then the testers can confound it by constructing an explosive quantity of more and more questions. It cannot yield passable words in response to the combined question by transforming passable words of the responses to independent questions4.

Thus the algorithm cannot successfully defeat the challenge unless it consists of more than a preset list of token manipulations. Instead it must think of the tokens as symbols (and qualify as Deacon's second "Symbolic Species"). It must expand and condense information. It must translate: 1) from the source sequence to reconstituted isomorphic reference "content", 2) from the restored "content" to an appropriate response, 3) from the response content to an adequate sequence of symbolic tokens to represent it.

My own past reactions to the Chinese Room have assumed that the algorithm in the room was doing these tasks in order to work convincingly: hence my view that the meanings, i.e. isomorphisms, of the unknown language simply resided in the smart algorithm as opposed to the algorithm's human "scribe". With a renovated Chinese Room in which the testers exploit the tireless elaboration permitted by language, the algorithm is forced to take the form of symbolic thought which I assumed; direct token transformation is a hopeless strategy5.


1 As other commentators have remarked, the unknown characters are "Chinese" in order to emphasize mystery for native speakers who use Latin-derived alphabets, so Aurebesh and Tengwar would fill the role too. In any case, with the right talent, tools, and educated skills, a linguist in the Room might speculate accurately about the unknown language itself after study of the algorithm and a large quantity of sample sequences. Nevertheless, the experience of archaeologists demonstrates that lack of context would prevent discovery of the majority of the tokens' meanings.
2 Grammatical modification is a significant characteristic of natural/unbounded language. Changes to tense, mood, gender and so forth are less like addition or multiplication and more like shifting ("rotating" or "translating") a vector's direction along many axes. The testers outside the room can certainly perform these language operations because they're fluent humans without the restriction of pure token manipulations. Unfortunately, they couldn't mechanize this type of processing until AI has a long-sought breakthrough: guided re-combination of old ideas is one of the leading definitions of creativity.
3 It's questionable whether mimicry of humanity is an essential proof of intelligence. I chose this instance to highlight that aspect. "Thinking like me" is a narrow and shallow definition of intelligence. Deacon acknowledges that this definition is already applied to answer the similar and basic question, "How do I know any human besides me has the same intelligence, sentience, and consciousness?"
4 In more formalized language, if the algorithm requires an explicit individualized mapping from every source sequence of tokens to a response sequence of tokens, then the algorithm is not computable. The combinatorial capability to create new meaningful source sequences implies that no theoretical algorithm can enumerate or count all the source sequences or contain a novel response sequence for each. (Nobody really speculates that human brains handle natural language in this way.)
5 Perhaps ambiguity would stymie direct token transformation sooner than I suggest. Without deep information about all the numerous definitions, the initial categorization of words as "nouns" or "verbs" could be problematic. For example, what if the question were to guess the speakers of the following quotations? 1) "Too much overhead is required to run a stand." 2) "I'm quitting my run because I can't stand the sun directly overhead." The first is an entrepreneur pondering where to sell, and the second is a dehydrated athlete.