As part of an ambitious project, I am attempting to better understand the legislative text that is written into bills introduced in the U.S. Congress. I have electronic versions of recent bills, and am attempting to implement an algorithm that would compare a bill with prior bills, looking for similarities. The hypothesis is that many bills that fail end up getting co-opted into other bills.
Obviously, this is a large task. Many questions exist regarding difference engines, but my issue is slightly different. Many times bills are introduced that package several ideas together. So the difference engine would need to compare portions of bills, not the entire bills.
Any recommendations on difference algorithms or a method to go about doing this? I have access to serious computational power, but do keep in mind that I will be using a dataset of about 100,000 bills.
Take a look at Simian - Similarity Analyser. It works for plain text as well as code.
Very interesting idea. I would start by looking into longest common subsequence algorithms, and see about adapting them to (1) report any sequence over some threshold, say, 20 words, and (2) see if you can get them to handle a bit of fuzziness, in case a word or two gets changed. I'd suggest looking at the diff code to start.
Related
I've written a GA to model a handful of stocks (4) over a period of time (5 years). It's impressive how quickly the GA can find an optimal solution to the training data, but I am also aware that this is mainly due to it's tendency to over-fit in the training phase.
However, I still thought I could take a few precautions and and get some kind of prediction on a set of unseen test stocks from the same period.
One precaution I took was:
When multiple stocks can be bought on the same day the GA only buys one from the list and it chooses this one randomly. I thought this randomness might help to avoid over-fitting?
Even if over-fitting is still occurring,shouldn't it be absent in the initial generations of the GA since it hasn't had a chance to over-fit yet?
As a note, I am aware of the no-free-lunch theorem which demonstrates ( I believe) that there is no perfect set of parameters which will produce an optimal output for two different datasets. If we take this further, does this no-free-lunch theorem also prohibit generalization?
The graph below illustrates this.
->The blue line is the GA output.
->The red line is the training data (slightly different because of the aforementioned randomness)
-> The yellow line is the stubborn test data which shows no generalization. In fact this is the most flattering graph I could produce..
The y-axis is profit, the x axis is the trading strategies sorted from worst to best ( left to right) according to there respective profits (on the y axis)
Some of the best advice I've received so far (thanks seaotternerd) is to focus on the earlier generations and increase the number of training examples. The graph below has 12 training stocks rather than just 4, and shows only the first 200 generations (instead of 1,000). Again, it's the most flattering chart I could produce, this time with medium selection pressure. It certainly looks a little bit better, but not fantastic either. The red line is the test data.
The problem with over-fitting is that, within a single data-set it's pretty challenging to tell over-fitting apart from actually getting better in the general case. In many ways, this is more of an art than a science, but here are some general guidelines:
A GA will learn to do exactly what you attach fitness to. If you tell it to get really good at predicting one series of stocks, it will do that. If you keep swapping in different stocks to predict, though, you might be more successful at getting it to generalize. There are a few ways to do this. The one that has had perhaps the most promising results for reducing over-fitting is imposing spatial structure on the population and evaluating on different test cases in different cells, as in the SCALP algorithm. You could also switch out the test cases on a time basis, but I've had more mixed results with that sort of an approach.
You are correct that over-fitting should be less of a problem early on. Generally, the longer you run a GA, the more over-fitting will be possible. Typically, people tend to assume that the general rules will be learned first, before the rote memorization of over-fitting takes place. However, I don't think I've actually ever seen this studied rigorously - I could imagine a scenario where over-fitting was so much easier than finding general rules that it happens first. I have no idea how common that is, though. Stopping early will also reduce the ability of the GA to find better general solutions.
Using a larger data-set (four stocks isn't that many) will make your GA less susceptible to over-fitting.
Randomness is an interesting idea. It will definitely hurt the GA's ability to find general rules, but it should also reduce over-fitting. Without knowing more about the specifics of your algorithm, it's hard to say which would win out.
That's a really interesting thought about the no free lunch theorem. I'm not 100% sure, but I think it does apply here to some extent - better fitting some data will make your results fit other data worse, by necessity. However, as wide as the range of possible stock behaviors is, it is much narrower than the range of all possible time series in general. This is why it is possible to have optimization algorithms at all - a given problem that we are working with tends produce data that cluster relatively closely together, relative to the entire space of possible data. So, within that set of inputs that we actually care about, it is possible to get better. There is generally an upper limit of some sort on how well you can do, and it is possible that you have hit that upper limit for your data-set. But generalization is possible to some extent, so I wouldn't give up just yet.
Bottom line: I think that varying the test cases shows the most promise (although I'm biased, because that's one of my primary areas of research), but it is also the most challenging solution, implementation-wise. So as a simpler fix you can try stopping evolution sooner or increasing your data-set.
I am currently working on a search ranking algorithm which will be applied to elastic search queries (domain: e-commerce). It assigns scores on several entities returned and finally sorts them based on the score assigned.
My question is: Has anyone ever tried to introduce a certain level of randomness to any search algorithm and has experienced a positive effect of it. I am thinking that it might be useful to reduce bias and promote the lower ranking items to give them a chance to be seen easier and get popular if they deserve it. I know that some machine learning algorithms are introducing some randomization to reduce the bias so I thought it might be applied to search as well.
Closest I can get here is this but not exactly what I am hoping to get answers for:
Randomness in Artificial Intelligence & Machine Learning
I don't see this mentioned in your post... Elasticsearch offers a random scoring feature: https://www.elastic.co/guide/en/elasticsearch/guide/master/random-scoring.html
As the owner of the website, you want to give your advertisers as much exposure as possible. With the current query, results with the same _score would be returned in the same order every time. It would be good to introduce some randomness here, to ensure that all documents in a single score level get a similar amount of exposure.
We want every user to see a different random order, but we want the same user to see the same order when clicking on page 2, 3, and so forth. This is what is meant by consistently random.
The random_score function, which outputs a number between 0 and 1, will produce consistently random results when it is provided with the same seed value, such as a user’s session ID
Your intuition is right - randomization can help surface results that get a lower than deserved score due to uncertainty in the estimation. Empirically, Google search ads seemed to have sometimes been randomized, and e.g. this paper is hinting at it (see Section 6).
This problem describes an instance of a class of problems called Explore/Exploit algorithms, or Multi-Armed Bandit problems; see e.g. http://en.wikipedia.org/wiki/Multi-armed_bandit. There is a large body of mathematical theory and algorithmic approaches. A general idea is to not always order by expected, "best" utility, but by an optimistic estimate that takes the degree of uncertainty into account. A readable, motivating blog post can be found here.
Help by editing my question title and tags is greatly appreciated!
Sometimes one participant in my corpus of "conversations" will refer to another participant using a nickname, usually an abbreviation or misspelling, but hereafter I'll just say "nicknames". Let's say I'm willing to manually tell my software whether or not I think various possible nicknames are in fact nicknames, but I want software to come up with a list of possible matches between the handle's that identify people, and the potential nicknames. How would I go about doing that?
Background on me and then my corpus: I have no experience doing natural language processing but I'm a competent data analyst with R. My data is produced by 70 teams, each forecasting the likelihood of 100 distinct events occurring some time in the future. The result that I have 70 x 100 = 7000 text files, containing the stream of forecasts participants make and the comments they include with their forecasts. I'll paste a very short snip of one of these text files below, this one had to do with whether the Malian government would enter talks with the MNLA:
02/12/2013 20:10: past_returns answered Yes: (50%)
I hadn't done a lot of research when I put in my previous
placeholder... I'm bumping up a lot due to DougL's forecast
02/12/2013 19:31: DougL answered Yes: (60%)
Weak President Traore wants talks if MNLA drops territorial claims.
Mali's military may not want talks. France wants talks. MNLA sugggests
it just needs autonomy. But in 7 weeks?
02/12/2013 10:59: past_returns answered No: (75%)
placeholder forecast...
http://www.irinnews.org/Report/97456/What-s-the-way-forward-for-Mali
My initial thoughts: Obviously I can start by providing the names I'm looking to match things up with... in the above example they would be past_returns and DougL (though there is no use of nicknames in the above). I wouldn't think it'd be that hard to get a computer to guess at minor misspellings (though I wouldn't personally know where to start). I can imagine that other tricks could be used, like assuming that a string is more likely to be a nickname if it is used much much more by one team, than by other teams. A nickname is more likely to refer to someone who spoke recently than someone who spoke long ago, or not at all on regarding this question. And they should be used in sentences in a manner similar to the way the full name/screenname is typically used in the corpus. But I'm interested to hear about simple approaches, as well as ones that try to consider more sophisticated techniques.
This could get about as complicated as you want to make it. From the semi-linguistic side of things, research topics would include Levenshtein Distance (for detecting minor misspellings of known names/nicknames) and Named Entity Recognition (for the task of detecting names/nicknames in the first place). Actually, NER's worth reading about, but existing systems might not help you much in your domain of forum handles and nicknames.
The first rough idea that comes to mind is that you could run a tokenized version of your corpus against an English dictionary (perhaps a dataset compiled from Wiktionary or something like WordNet) to find words that are candidates for names, then filter those through some heuristics (do they start with the same letters as known full names? Do they have a low Levenshtein distance from known names? Are they used more than once?).
You could also try some clustering or supervised ML algorithms against the non-word tokens. That might reveal some non-"word" tokens that often occur in the same threads as a given username; again, heuristics could help rule out some false positives.
Good luck; sounds like a fun problem - hope I mentioned at least one thing you hadn't already thought of.
I have a python app with a database of businesses and I want to be able to search for businesses by name (for autocomplete purposes).
For example, consider the names "best buy", "mcdonalds", "sony" and "apple".
I would like "app" to return "apple", as well as "appel" and "ple".
"Mc'donalds" should return "mcdonalds".
"bst b" and "best-buy" should both return "best buy".
Which algorithm am I looking for, and does it have a python implementation?
Thanks!
The Levenshtein distance should do.
Look around - there are implementations in many languages.
Levenshtein distance will do this.
Note: this is a distance, you have to calculate it to every string in your database, which can be a big problem if you have a lot of entries.
If you have this problem then record all the typos the users make (typo=no direct match) and offline build a correction database which contains all the typo->fix mappings. some companies do this even more clever, eg: google watches how users correct their own typos and learns the mappings from this.
Soundex or Metaphone might work.
I think what you are looking for is a huge field of Data Quality and Data Cleansing. I fear if you could find a python implementation regarding this as it has to be something which cleanses considerable amount of data in db which could be of business value.
Levensthein distance goes in the right direction but only half the way. There are several tricks to get it to use the half matches as well.
One would be to use a subsequence dynamic time warping (DTW is actually a generalization of levensthein distance). For this you relax the start and end cases when calcualting the cost matrix. If you only relax one of the conditions you can get autocompletion with spell checking. I am not sure if there is a python implementation available, but if you want to implement it for yourself it should not be more than 10-20 LOC.
The other idea would be to use a Trie for speed up, which can do DTW/Levensthein on multiple results simultaniously (huge speedup if your database is large). There is a paper on Levensthein on Tries at IEEE, so you can find the algorithm there. Again for this you would need to relax the final boundary condition, so you get partial matches. However since you step down in the trie you just need to check when you have fully consumed the input and then return all leaves.
check this one http://docs.python.org/library/difflib.html
it should help you
I am looking for a resource similar to WordNet. However, I want to be able to look up the positive/negative connotation of a word. For example:
bribe - negative
offer - positive
I'm curious as to whether anyone has run across any tool like this in AI/NLP research, or even in linguistics.
UPDATE:
For the curious, the accepted answer below put me on the right track towards what I needed. Wikipedia listed several different resources. The two I would recommend (because of ease of use/free use for a small number of API calls) are AlchemyAPI and Lymbix. I decided to go with AlchemyAPI, since people affiliated with academic institutions (like myself) and non-profits can get even more API calls per day if they just email the company.
Start looking up topics on 'sentiment analysis': http://en.wikipedia.org/wiki/Sentiment_analysis
The are some vocabulary compilations regarding affect, aka dictionaries of affect, such as the Affective Norms of English Words (ANEW) or the Dictionary of Affect in Language (DAL). They provide a dimensional representation of affect (valence, activation and control) that may be of use in a sentiment analysis scenario (detection of positive/negative connotation). In this sense, EmoLib works with the former, by default, but may be easily extended with a more specific lexicon to tackle particular needs (for example, EmoLib provides an additional neutral label that is more appropriate than the positive/negative tag set alone in a Text-To-Speech synthesis setting).
There is also SentiWordNet, which gives you positive, negative and objective scores for each WordNet synset.
However, you should be aware that the positive and negative connotation of a term often depends on the context in which it is used. A great introduction to this topic is the book Opinion mining and sentiment analysis by Bo Pang and Lillian Lee, which is available online for free.