For a toy project, I want to implement an automated question answering system with Lucene and I'm trying to figure out a reasonable way to implement it. The basic operation is as follows:
1) The user will enter a question.
2) The system will identify the keywords in the question.
3) The keywords will be searched in a large knowledgebase and matching sentences will be shown as answers.
My knowledgebase (i.e., corpus) is not structured. It is just a large, continuous text (say, a user manual without any chapters). I mean that the only structure is that sentences and paragraphs are identified.
I plan to treat each sentence or paragraph as a separate document. To present the answer in a context, I may consider keeping one sentence/paragraph before/after the indexed one as payload. I would like to know if that makes sense. Also, I'm wondering if there are other tried and well-known approaches for that kind of systems. As an example, another approach that comes to mind is to index large chunks of the corpus as documents with the token positions, then process the vicinity of found keywords to construct my answers.
I would appreciate direct recommendations based on experience or intuition, but also tutorials or introductory materials to question-answering systems with Lucene in mind.
Thanks.
It's not an unreasonable approach to take.
One enhancement you might consider is incorporating learning feedback, so that you can continually improve the scoring of content vs search terms. To do this you would ask users to rate the answers that come back ('helpful vs unhelpful'), that way you can start to rank documents against keywords based on the historical data. You could classify potential documents as helpful/unhelpful for given keywords by using a simple Bayesian classifier.
Indexing each sentence as a document will give you some problems. You've pointed out one: you would need to store the surrounding texts a payloads. That means you'll need to store each sentence three times (before, during and after), and you'll have to manually get into the payload.
If you want to go the route of each sentence being a document, I would recommend coming up with an ID for each sentence and storing that as a separate field. Then you can display [ID-1, ID, ID+1] in each result.
The bigger question though is: how should you break up the text into documents? Identifying semantically related areas seems difficult, so doing it by sentence/paragraph might be the only way to go. A better way would be if you could find which text is the header of a section, and then put everything in that section as a document.
You might also want to use the index (if your corpus has one). The terms there could be boosted, as they are presumably more important.
Instead of luncene which does text indexing, search and retrieval, I think using something like Apache Mahout would help with this. Mahout considers text as knowledge and doing that makes the answering the question better than just text matching. Mahout is a machine learning and data mining f/w which fits this domain better. Just a very high level thought.
--Sai
Related
I was reading some interesting questions about the topic "Can we make a program that, given a particular sequence, produces the next terms", like this one, and I really like the detailed answer of this one. I understand that the answer is "That's impossible without more restrictions", and that given some restrictions (polynomials, rational function or boolean map) we know some good algorithms, as the second answer I linked explains.
Now, a natural question is how much can we solve, trying our best even if we can't always solve it, to answer the original, general question. What I usually do when facing a hard sequence is trying to see if it's in OEIS, and if it seems to be there, seeing if there is any formula or algorithm to produce it in there. You can download a small version of OEIS with the first terms of each sequence, and you can make queries to find formulas or maple algorithms for a particular sequence. My question is, do you think it's feasible to download a small version of OEIS that includes, with the first terms, a little algorithm to produce it?
The natural problem here is that I haven't seen any link to download the entire database of OEIS with all the details, which maybe deserves its own question. Even if we had this, you need to read the formulas/algorithms (that can be written in different languages, from what I've seen) and interpret them correctly. But I thought maybe someone here knows how to solve this, in any case thanks in advance.
You could, as you note, download the sequences and their A-numbers from the link mentioned here: https://oeis.org/wiki/Welcome#Compressed_Versions
After searching that and finding one sequence (or a small number of sequences) of interest, you could scrape the respective page(s) for formulas. There are specific fields for Maple and Mathematica, which may be helpful, and otherwise, an entry in the PROGRAM field should include identifying information when it is not one of the standard languages with its own field in the database. See: http://oeis.org/wiki/Style_Sheet
Unofficially, but with the interests of the OEIS in mind, I would not recommend trying to download or scrape the OEIS in its entirety. Whether it's one person, or a whole host of people, we would certainly recommend using the compressed version of the database to identify sequences of interest by A-number first, then pulling their entire entry by scraping the site or querying the OEIS using methods that you have already mentioned: Programmatic access to On-Line Encyclopedia of Integer Sequences
If this sounds laborious, perhaps an alternative is the Wolfram Cloud, which actives this through other means. For example, you can navigate to the cloud (you may have to register just to get access) at: https://www.wolframcloud.com/
Typing in something like FindSequenceFunction[{1, 2, 3, 5, 17, 305, 34865}] will give you a formula, if Wolfram/Mathematica can find one. The documentation for FindSequenceFunction can be found here: https://reference.wolfram.com/language/ref/FindSequenceFunction.html
Wolfram/Mathematica can also invoke the OEIS using packages like the one described here: https://mathematica.stackexchange.com/questions/40/is-it-possible-to-invoke-the-oeis-from-mathematica
I'm in the planning phase of developing a very tag heavy website. Everything will essentially be associated with tags and the entire site would be based on searching these tags.
Now, I've been thinking a lot about going the nosql route here, since from what I read and understand, it makes the most sense for something like this.
Would it be best to go with this database system? Would it makes sense to go with the relational database system? Should I think incorporating something like SOLR?
What would the ideal setup be?
UPDATE:
Ideally they would be user generated, but we all know how that would turn out with giving users that much power. So, let’s change up the requirements and say that users WILL NOT have the power to create tags.
Searching on tags based on text matches is something that would probably be useful and needed. If the tag is “garage sale”, the search for “sale” should also pick this up, at a lower relevance for sure.
I can’t imagine the usage being so much that scaling would be an issue.
Thanks
I would spend a bit of time thinking about these tags. For example, are these tags going to be user generated or will you provide a few tags and let users select which ones they want?
Will you need to search on tags based on text matches? For example if a tag is "garage sale" do you want to search for "sale" to also pick this up? Maybe at a lower relevance?
Also, what kind of usage are you looking at? One good thing about Solr is that it's super easy to scale and synchronize data, it is easy to deploy multiple nodes, shard collections and replicate data to other nodes, something that traditional databases struggle with.
Another thing to keep in mind is that most of the time, Solr is not the official "repository of record", most of the time the data gets fed to it from a DB somewhere, but all reading activities are done from Solr.
See this answer for a SQL solution. Offhand I can't think of any advantage to using most NoSQL databases (i.e. key-value, columnar, or document) as the SQL solution will be more compact and ought to give good performance; a graph database may be appropriate if you're doing a lot of navigational type queries on your tags, but it doesn't sound like that's the case.
Use of Solr (or ElasticSearch or whatever) is orthogonal to your primary database; it may be appropriate to incorporate a search tool if users are typing inexact tags for search, but I recommend integrating a stemming library or something along those lines before turning to a full blown search tool.
Association mining seems to give good results for retrieving related terms in text corpora. There are several works on this topic including well-known LSA method. The most straightforward way to mine associations is to build co-occurrence matrix of docs X terms and find terms that occur in the same documents most often. In my previous projects I implemented it directly in Lucene by iteration over TermDocs (I got it by calling IndexReader.termDocs(Term)). But I can't see anything similar in Solr.
So, my needs are:
To retrieve the most associated terms within particular field.
To retrieve the term, that is closest to the specified one within particular field.
I will rate answers in the following way:
Ideally I would like to find Solr's component that directly covers specified needs, that is, something to get associated terms directly.
If this is not possible, I'm seeking for the way to get co-occurrence matrix information for specified field.
If this is not an option too, I would like to know the most straightforward way to 1) get all terms and 2) get ids (numbers) of documents these terms occur in.
You can export a Lucene (or Solr) index to Mahout, and then use Latent Dirichlet Allocation. If LDA is not close enough to LSA for your needs, you can just take the correlation matrix from Mahout, and then use Mahout to take the singular value decomposition.
I don't know of any LSA components for Solr.
Since there are still no answers to my questions, I have to write my own thoughts and accept it. Nevertheless, if someone propose better solution, I'll happily accept it instead of mine.
I'll go with co-occurrence matrix, since it is the most principal part of association mining. In general, Solr provides all needed functions for building this matrix in some way, though they are not as efficient as direct access with Lucene. To construct matrix we need:
All terms or at least the most frequent ones, because rare terms won't affect result of association mining by their nature.
Documents where these terms occur, again, at least top documents.
Both these tasks may be easily done with standard Solr components.
To retrieve terms TermsComponent or faceted search may be used. We can get only top terms (by default) or all terms (by setting max number of terms to take, see documentation of particular feature for details).
Getting documents with the term in question is simply search for this term. The weak point here is that we need 1 request per term, and there may be thousands of terms. Another weak point is that neither simple, nor faceted search do not provide information about the count of occurrences of the current term in found document.
Having this, it is easy to build co-occurrence matrix. To mine association it is possible to use other software like Weka or write own implementation of, say, Apriori algorithm.
You can get the count of occurrences of the current term in found document in the following query:
http://ip:port/solr/someinstance/select?defType=func&fl=termfreq(field,xxx),*&fq={!frange l=1}termfreq(field,xxx)&indent=on&q=termfreq(field,xxx)&sort=termfreq(field,xxx) desc&wt=json
I'm currently configuring my schema.xml file and trying to figure out what's the best way to set up my documents. I use a RMDBS and thus many objects are relational.
Take this site for instance; a document typically consists of a question, followed by 0 or more answers. Say you'd want to set up fields for this, you would have to declare all question and answer fields in the same document, the way I see it. But given the fact that there can be more than one answer, you would have to create a document for each answer. So that means each question and each answer is stored in a separate document, which contains fields for both.
I don't see a different approach for this kind of problem, however I am relatively new to Solr and document DB's, so I may be wrong.
In short: what are the best practices if I would implement such a schema?
Another way to do it would be to have a question field and a multi-valued field for answers and have them in the same document. This is probably the best way to start unless you have specific requirements that favor the document-per-answer approach.
For example, if you needed to match individual answers as stand-alone search results, you may get better results and performance with the document-per-answer approach, as the "answer" documents will be scored, ranked and loaded in isolation.
But that would be an unconventional use of this type of data. Normally when you search a site like stack overflow, you are searching for a question and set of answers that cover a particular topic, so having everything in one document makes more sense.
Is there a way to search the web which does NOT remove punctuation? For example, I want to search for window.window->window (Yes, I actually do, this is a structure in mozilla plugins). I figure that this HAS to be a fairly rare string.
Unfortunately, Google, Bing, AltaVista, Yahoo, and Excite all strip the punctuation and just show anything with the word "window" in it. And according to Google, on their site, at least, there is NO WAY AROUND IT.
In general, searching for chunks of code must be hard for this reason... anyone have any hints?
google codesearch ("window.window->window" but it doesn't seem to get any relevant result out of this request)
There is similar tools all over the internet like codase or koders but I'm not sure they let you search exactly this string. Anyway they might be useful to you so I think they're worth mentioning.
edit: It is very unlikely you'll find a general purpose search engine which will allow you to search for something like "window.window->window" because most search engines will do some processing on the document before storing it. For instance they might represent it internally as vectors of words (a vector space model) and use that to do the search, not the actual original string. And creating such a vector involves first cutting the document according to punctuation and other critters. This is a very complex and interesting subject which I can't tell you much more about. My bad memory did a pretty good job since I studied it at school!
BTW they might do the same kind of processing on your query too. You might want to read about tf-idf which is probably light years from what google and his friends are doing but can give you a hint about what happens to your query.
There is no way to do that, by itself in the main Google engine, as you discovered -- however, if you are looking for information about Mozilla then the best bet would be to structure your query something more like this:
"window.window->window" +Mozilla
OR +XUL
+ Another search string related to what you are
trying to do.
SymbolHound is a web search that does not remove punctuation from the queries. There is an option to search source code repositories (like the now-discontinued Google Code Search), but it also has the option to search the Internet for special characters. (primarily programming-related sites such as StackOverflow).
try it here: http://www.symbolhound.com
-Tom (co-founder)