systems that know and understand and think and learn
What is the part of speech of “subject” in the sentence:
Related questions might include:
Vitamins is the direct object of the verb in each of these sentences, so, perhaps you would think “subject” is a verb in the subject sentence…
Continue reading “Are vitamins subject to sales tax in California?”
I regularly build deep learning models for natural language processing and today I gave one a try that has been the leader in the Stanford Question Answering Dataset (SQuAD). This one is a impressive NLP platform built using PyTorch. But it’s still missing the big picture (i.e., it doesn’t “know” much).
Generally, NLP systems that emphasize Big Data (e.g., deep learning approaches) but eschew more explicit knowledge representation and reasoning are interesting but unintelligent. Think Siri and Alexa, for example. They might get a simple factoid question if a Google search can find closely related text, but not much more.
Here is a simple demonstration of problems that the state of the art in deep machine learning is far from solving…
Here is a paragraph from a Wall Street Journal article about the Fed today where the deep learning system has “found” what the pronouns “this” and “they” reference:
The essential point here is that the deep learning system is missing common sense. It is “the need”, not “a raise” that is referenced by “this”. And “they” references “officials”, not “the minutes”.
Bottom line: if you need your natural language understanding system to be smarter than this, you are not going to get there using deep learning alone.
The following is motivated by Section 6359 of the California Sales and Use Tax. It demonstrates how knowledge can be acquired from dictionary definitions:
Here, we’ve taken a definition from WordNet and prefixed it with the word followed by a colon and parsed it using the Linguist.
A Linguist user recently had a question about part of a sentence that boiled down to something like the following:
The question was whether “many” was an adjective, cardinality, or noun in this sentence. It’s a reasonable question!
The Winograd Challenge is an alternative to the Turing Test for assessing artificial intelligence. The essence of the test involves resolving pronouns. To date, systems have not fared well on the test for several reasons. There are 3 that come to mind:
As an example, one of the simpler exemplary problems is:
A heuristic system (or a deep learning one) could infer that “it” does not refer to “me” or “I” and toss a coin between “pillar” and “stage”. A system worthy of the passing the Winograd Challenge should “know” it’s the pillar.
Even this simple sentence presents some NLP challenges that are easy to overlook. For example, does “between” modify the pillar or the verb “is”?
This is not much of a challenge, however, so let’s touch on some deeper issues and a more challenging problem…
Just a quick note about a natural language interpretation that came up for the following sentence:
The NLP system comes up with many hundreds of plausible parses for this sentence (mostly because it’s considering lexical and syntactic possibilities that are not semantically plausible). Among these is “meaning” as a nominalization.
From Wikipedia:
It’s quite common to use the present participle of a verb as a noun. In this case, Google comes up with this definition for the noun ‘meaning’:
The NLP system has a definition of “meaning” as a mass or count noun as well as definitions for several senses of the verb “mean”, such as these:
Text analytics and natural language processing have made tremendous advances in the last few years. Unfortunately, there is a lot more to understanding natural language that TA/NLP.
I was reading a paper today about NLP pipelines for question answering that used machine learning to find what tools are good at what tasks and to configure a pipeline by selecting the best tool for a given task from each of the types of components in the pipeline. The paper has a long list of various components, so I checked a few out. Most of those of interest were available on the web so that they could be easily composed into pipelines without a lot of software setup. Looking at these I quickly tired in disappointment. Here are some of the reasons.
I am not surprised by these results. NLU is hard. But they are not particularly strong results either. I’m surprised that people find such results useful (if they do).
A nationwide physical therapy business is renowned for eliminating chronic pain. One unique aspect of this business is that it eliminates pain not by manipulation but by providing clients with expertly selected sequences of exercises that address problems in their functional anatomy. In effect, the business helps people fix themselves and teaches them to maintain their musculoskeletal function for pain-free life.
The business has dozens of clinics with many more therapists. Its expertise comes primarily from its founder and a number of long-time employees who have learned through experience. We have been engaged to assist with several challenges on several occasions.
Working on translating some legal documentations (sales and use tax laws and regulations) into compliance logic, we came across the following sentence (and many more that are even worse):
Natural language processing systems choke on sentences like this because of such sentences’ combinatorial ambiguity and NLP’s typical lack of knowledge about what can be conjoined or complement or modify what.
This sentences has many thousands of possible parses. They involve what the scopes of each of the the ‘or’s are and what is modified by conditional, otherwise, or whatsoever and what is complemented by in, by, of, and for.
The following shows 2 parses remaining after we veto a number of mistakes and confirm some phrases from the 400 highest ranking parses (a few right or left clicks of the mouse):
Recently, we “read” ten thousand recipes or so from a cooking web site. The purpose of doing so was to produce a formal representation of those recipes for use in temporal reasoning by a robot.
Our task was to produce ontology by reading the recipes subject to conflicting goals. On the one hand, the ontology was to be accurate so that the robot could reason, plan, and answer questions robustly. On the other hand, the ontology was to be produced automatically (with minimal human effort).[1]
In order to minimize human effort while still obtaining deep parses from which we produce ontology, we used more techniques from statistical natural language processing than we typically do in knowledge acquisition for deep QA, compliance, or policy automation. (Consider that NLP typically achieves less than 90% syntactic accuracy while such work demands near 100% semantic accuracy.)[2]
In the effort, we refined some prior work on representing words as vectors and semi-supervised learning. In particular, we adapted semi-supervised, active learning similar to Stratos & Collins 2015 using enhancements to the canonical correlation analysis (CCA) of Dhillon et al 2015 to obtain accurate part of speech tagging, as conveyed in the following graphic from Stratos & Collins:
