I've studied some simple semantic network implementations and basic techniques for parsing natural language. However, I haven't seen many projects that try and bridge the gap between the two.
For example, consider the dialog:
"the man has a hat"
"he has a coat"
"what does he have?" => "a hat and coat"
A simple semantic network, based on the grammar tree parsing of the above sentences, might look like:
the_man = Entity('the man')
has = Entity('has')
a_hat = Entity('a hat')
a_coat = Entity('a coat')
Relation(the_man, has, a_hat)
Relation(the_man, has, a_coat)
print the_man.relations(has) => ['a hat', 'a coat']
However, this implementation assumes the prior knowledge that the text segments "the man" and "he" refer to the same network entity.
How would you design a system that "learns" these relationships between segments of a semantic network? I'm used to thinking about ML/NL problems based on creating a simple training set of attribute/value pairs, and feeding it to a classification or regression algorithm, but I'm having trouble formulating this problem that way.
Ultimately, it seems I would need to overlay probabilities on top of the semantic network, but that would drastically complicate an implementation. Is there any prior art along these lines? I've looked at a few libaries, like NLTK and OpenNLP, and while they have decent tools to handle symbolic logic and parse natural language, neither seems to have any kind of proabablilstic framework for converting one to the other.
There is quite a lot of history behind this kind of task. Your best start is probably by looking at Question Answering.
The general advice I always give is that if you have some highly restricted domain where you know about all the things that might be mentioned and all the ways they interact then you can probably be quite successful. If this is more of an 'open-world' problem then it will be extremely difficult to come up with something that works acceptably.
The task of extracting relationship from natural language is called 'relationship extraction' (funnily enough) and sometimes fact extraction. This is a pretty large field of research, this guy did a PhD thesis on it, as have many others. There are a large number of challenges here, as you've noticed, like entity detection, anaphora resolution, etc. This means that there will probably be a lot of 'noise' in the entities and relationships you extract.
As for representing facts that have been extracted in a knowledge base, most people tend not to use a probabilistic framework. At the simplest level, entities and relationships are stored as triples in a flat table. Another approach is to use an ontology to add structure and allow reasoning over the facts. This makes the knowledge base vastly more useful, but adds a lot of scalability issues. As for adding probabilities, I know of the Prowl project that is aimed at creating a probabilistic ontology, but it doesn't look very mature to me.
There is some research into probabilistic relational modelling, mostly into Markov Logic Networks at the University of Washington and Probabilstic Relational Models at Stanford and other places. I'm a little out of touch with the field, but this is is a difficult problem and it's all early-stage research as far as I know. There are a lot of issues, mostly around efficient and scalable inference.
All in all, it's a good idea and a very sensible thing to want to do. However, it's also very difficult to achieve. If you want to look at a slick example of the state of the art, (i.e. what is possible with a bunch of people and money) maybe check out PowerSet.
Interesting question, I've been doing some work on a strongly-typed NLP engine in C#: http://blog.abodit.com/2010/02/a-strongly-typed-natural-language-engine-c-nlp/ and have recently begun to connect it to an ontology store.
To me it looks like the issue here is really: How do you parse the natural language input to figure out that 'He' is the same thing as "the man"? By the time it's in the Semantic Network it's too late: you've lost the fact that statement 2 followed statement 1 and the ambiguity in statement 2 can be resolved using statement 1. Adding a third relation after the fact to say that "He" and "the man" are the same is another option but you still need to understand the sequence of those assertions.
Most NLP parsers seem to focus on parsing single sentences or large blocks of text but less frequently on handling conversations. In my own NLP engine there's a conversation history which allows one sentence to be understood in the context of all the sentences that came before it (and also the parsed, strongly-typed objects that they referred to). So the way I would handle this is to realize that "He" is ambiguous in the current sentence and then look back to try to figure out who the last male person was that was mentioned.
In the case of my home for example, it might tell you that you missed a call from a number that's not in its database. You can type "It was John Smith" and it can figure out that "It" means the call that was just mentioned to you. But if you typed "Tag it as Party Music" right after the call it would still resolve to the song that's currently playing because the house is looking back for something that is ITaggable.
I'm not exactly sure if this is what you want, but take a look at natural language generation wikipedia, the "reverse" of parsing, constructing derivations that conform to the given semantical constraints.
Related
I have a noob question, go easy on me — I'll probably get the terminology wrong. I'm hoping someone can give me the "here's what to google next" explanation for how to approach creating a CoreML model that can identify tokens within spans. Since my question falls between the hello world examples and the intellectual papers that cover the topics in detail, it has been hard to google for.
I'm taking my first stab at doing some natural language processing, specifically parsing data out of recipe ingredients. CreateML supports word tagging, which I interpret to mean Named Entity Recognition — split a string into tokens (probably words), annotate them, feed them to the model.
"1 tablespoon (0.5 oz / 14 g) baking soda"
This scenario immediately breaks my understanding of word tagging. Tokenize this by words, this includes three measurements. However, this is really one measurement with a clarification that contains two alternate measurements. What I really want to do is to label "(0.5 oz / 14 g)" as a clarification which contains measurements.
Or how about "Olive oil". If I were tokenizing by words, I'd probably get two tokens labeled as "ingredient" which I'd interpret to mean I have two ingredients, but they go together as one.
I've been looking at https://prodi.gy/ which does span categorization, and seemingly handles this scenario — tokenize, then name the entities, then categorize them into spans. However, as far as I understand it, spans are an entirely different paradigm which wouldn't convert over to CoreML.
My naive guess for how you'd do this in CoreML is that I use multiple models, or something that works recursively — one pass would tokenize "(0.5 oz / 14 g)" as a single token labeled as "clarification" and then the next pass would tokenize it into words. However, this smells like a bad idea.
So, how does one solve this problem with CoreML? Code is fine, if relevant, but I'm really just asking about how to think about the problem so I can continue my research.
Thanks for your help!
My Motivations I'm trying to learn German and realized there's a confounding fact with the structure of German: every noun has a gender which seems unrelated to the noun itself in many cases.
Unlike languages such as English, each noun has a different definite article, depending on gender: der (masculine), die (feminine), and das (neuter). For example:
das Mädchen ("the girl"), der Rock ("the skirt), die Hose ("the trousers/pants"). So, there seems to be no correlation between gender assignment of nouns and their meanings.
The Data
I gathered up to 5000 German words with 3 columns (das, der, die) for each word with 1's and 0's. So, my data is already clustered with one hot encoding and I'm not trying to predict anything.
Why I'm here I am clueless on where to start, how to approach this problem as the concept of distance in clustering doesn't make sense to me in this setting. I can't think of a way to generate an understandable description of these clusters. The mixed data makes it impossible for me to think of some hard-coded metrics for evaluation.
So, my question is:
I want to find some patterns, some characteristics of these words that made them fall in a specific cluster. I don't know if I'm making any sense but some people managed to find some patterns already (for example word endings, elongated long objects tend to be masculine etc., etc) and I believe ML/AI could do a way better job at this. Would it be possible for me to do something like this?
Some personal thoughts
While I was doing some research (perhaps, naive), I realized the potential options are decision trees and cobweb algorithms. Also, I was thinking if I could just scrape a few images (say 5) for every word and try to run some image classification and see the intermediate NN's to see if any specific shapes support a specific object gender. In addition to that, I was wondering whether scraping the data of google n-gram viewers of these words could help in anyway. I couldn't think of a way to use NLP or its sub domains.
Alternatives If everything I just wrote sounds nonsensical, please suggest me a way to make visual representations of my dataframe (more like nodes and paths with images at nodes, one for each cluster) in Python so that I could make pictorial mind maps and try to by heart them.
The ultimate purpose is to make learning German simpler for myself and possibly for others
I am experimenting with machine learning in general, and Bayesian analysis in particular, by writing a tool to help me identify my collection of e-books. The input data consist of a set of e-book files, whose names and in some cases contents contain hints as to the book they correspond to.
Some are obvious to the human reader, like:
Artificial Intelligence - A Modern Approach 3rd.pdf
Microsoft Press - SharePoint Foundation 2010 Inside Out.pdf
The Complete Guide to PC Repair 5th Ed [2011].pdf
Hamlet.txt
Others are not so obvious:
Vsphere5.prc (Actually 'Mastering VSphere 5' by Scott Lowe)
as.ar.pdf (Actually 'Atlas Shrugged' by Ayn Rand)
Rather than try to code various parsers for different formats of file names, I thought I would build a few dozen simple rules, each with a score.
For example, one rule would look in the first few pages of the file for something resembling an ISBN number, and if found would propose a hypothesis that the file corresponds to the book identified by that ISBN number.
Another rule would look to see if the file name is in 'Author - Title' format and, if so, would propose a hypothesis that the author is 'Author' and the title is 'Title'. Similar rules for other formats.
I thought I could also get a list of book titles and authors from Amazon or an ISBN database, and search the file name and first few pages of the file for any of these; any matches found would result in a hypothesis being suggested by that rule.
In the end I would have a set of tuples like this:
[rulename,hypothesis]
I expect that some rules, such as the ISBN match, will have a high probability of being correct, when they are available. Other rules, like matches based on known book titles and authors, would be more common but not as accurate.
My questions are:
Is this a good approach for solving this problem?
If so, is Bayesian analysis a good candidate for combining all of these rules' hypotheses into compound score to help determine which hypothesis is the strongest, or most likely?
Is there a better way to solve this problem, or some research paper or book which you can suggest I turn to for more information?
It depends on the size of your collection and the time you want to spend training the classifier. It will be difficult to get good generalization that will save you time. For any type of classifier you will have to create a large training set, and also find a lot of rules before you get good accuracy. It will probably be more efficient (less false positives) to create the rules and use them only to suggest title alternatives for you to choose from, and not to implement the classifier. But, if the purpose is learning, then go ahead.
I am searching for information on algorithms to process text sentences or to follow a structure when creating sentences that are valid in a normal human language such as English. I would like to know if there are projects working in this field that I can go learn from or start using.
For example, if I gave a program a noun, provided it with a thesaurus (for related words) and part-of-speech (so it understood where each word belonged in a sentence) - could it create a random, valid sentence?
I'm sure there are many sub-sections of this kind of research so any leads into this would be great.
The field you're looking for is called natural language generation, a subfield of natural language processing
http://en.wikipedia.org/wiki/Natural_language_processing
Sentence generation is either really easy or really hard depending on how good you want the sentences to be. Currently, there aren't programs that will be able to generate 100% sensible sentences about given nouns (even with a thesaurus) -- if that is what you mean.
If, on the other hand, you would be satisfied with nonsense that was sometimes ungrammatical, then you could try an n-gram based sentence generator. These just chain together of words that tend to appear in sequence, and 3-4-gram generators look quite okay sometimes (although you'll recognize them as what generates a lot of spam email).
Here's an intro to the basics of n-gram based generation, using NLTK:
http://www.nltk.org/book/ch02.html#generating-random-text-with-bigrams
This is called NLG (Natural Language Generation), although that is mainly the task of generating text that describes a set of data. There is also a lot of research on completely random sentence generation as well.
One starting point is to use Markov chains to generate sentences. How this is done is that you have a transition matrix that says how likely it is to transition between every every part-of-speech. You also have the most likely starting and ending part-of-speech of a sentence. Put this all together and you can generate likely sequences of parts-of-speech.
Now, you are far from done, this will first of all not offer a very good result as you are only considering the probability between adjacent words (also called bi-grams), so what you want to do is to extend this to look for instance at the transition matrix between three parts-of-speech (this makes a 3D matrix and gives you trigrams). You can extend it to 4-grams, 5-grams, etc. depending on the processing power and if your corpus can fill such matrix.
Lastly, you need to patch up things such as object agreement (subject-verb-agreement, adjective-verb-agreement (not in English though), etc.) and tense, so that everything is congruent.
Yes. There is some work dealing with solving problems in NLG with AI techniques. As far as I know, currently, there is no method that you can use for any practical use.
If you have the background, I suggest getting familiar with some work by Alexander Koller from Saarland University. He describes how to code NLG to PDDL. The main article you'll want to read is "Sentence generating as a planning problem".
If you do not have any background in NLP, just search for the online courses or course materials by Michael Collings or Dan Jurafsky.
Writing random sentences is not that hard. Any parser textbook's simple-english-grammar example can be run in reverse to generate grammatically correct nonsense sentences.
Another way is the word-tuple-random-walk, made popular by the old BYTE magazine TRAVESTY, or stuff like
http://www.perlmonks.org/index.pl?node_id=94856
Am thinking about a project which might use similar functionality to how "Quick Add" handles parsing natural language into something that can be understood with some level of semantics. I'm interested in understanding this better and wondered what your thoughts were on how this might be implemented.
If you're unfamiliar with what "Quick Add" is, check out Google's KB about it.
6/4/10 Update
Additional research on "Natural Language Parsing" (NLP) yields results which are MUCH broader than what I feel is actually implemented in something like "Quick Add". Given that this feature expects specific types of input rather than the true free-form text, I'm thinking this is a much more narrow implementation of NLP. If anyone could suggest more narrow topic matter that I could research rather than the entire breadth of NLP, it would be greatly appreciated.
That said, I've found a nice collection of resources about NLP including this great FAQ.
I would start by deciding on a standard way to represent all the information I'm interested in: event name, start/end time (and date), guest list, location. For example, I might use an XML notation like this:
<event>
<name>meet Sam</name>
<starttime>16:30 07/06/2010</starttime>
<endtime>17:30 07/06/2010</endtime>
</event>
I'd then aim to build up a corpus of diary entries about dates, annotated with their XML forms. How would I collect the data? Well, if I was Google, I'd probably have all sorts of ways. Since I'm me, I'd probably start by writing down all the ways I could think of to express this sort of stuff, then annotating it by hand. If I could add to this by going through friends' e-mails and whatnot, so much the better.
Now I've got a corpus, it can serve as a set of unit tests. I need to code a parser to fit the tests. The parser should translate a string of natural language into the logical form of my annotation. First, it should split the string into its constituent words. This is is called tokenising, and there is off-the-shelf software available to do it. (For example, see NLTK.) To interpret the words, I would look for patterns in the data: for example, text following 'at' or 'in' should be tagged as a location; 'for X minutes' means I need to add that number of minutes to the start time to get the end time. Statistical methods would probably be overkill here - it's best to create a series of hand-coded rules that express your own knowledge of how to interpret the words, phrases and constructions in this domain.
It would seem that there's really no narrow approach to this problem. I wanted to avoid having to pull along the entirety of NLP to figure out a solution, but I haven't found any alternative. I'll update this if I find a really great solution later.