I have a set of requirements and I'm looking for the best Java-based strategy / algorthm / software to use. Basically, I want to take a set of recipe ingredients entered by real people in natural english and parse out the meta-data into a structured format (see requirements below to see what I'm trying to do).
I've looked around here and other places, but have found nothing that gives a high-level advice on what direction follow. So, I'll put it to the smart people :-):
What's the best / simplest way to solve this problem? Should I use a natural language parser, dsl, lucene/solr, or some other tool/technology? NLP seems like it may work, but it looks really complex. I'd rather not spend a whole lot of time doing a deep dive just to find out it can't do what I'm looking for or that there is a simpler solution.
Requirements
Given these recipe ingredient descriptions....
"8 cups of mixed greens (about 5 ounces)"
"Eight skinless chicken thighs (about 1ΒΌ lbs)"
"6.5 tablespoons extra-virgin olive oil"
"approximately 6 oz. thinly sliced smoked salmon, cut into strips"
"2 whole chickens (3 .5 pounds each)"
"20 oz each frozen chopped spinach, thawed"
".5 cup parmesan cheese, grated"
"about .5 cup pecans, toasted and finely ground"
".5 cup Dixie Diner Bread Crumb Mix, plain"
"8 garlic cloves, minced (4 tsp)"
"8 green onions, cut into 2 pieces"
I want to turn it into this....
|-----|---------|-------------|-------------------------|--------|-----------|--------------------------------|-------------|
| | Measure | | | weight | weight | | |
| # | value | Measure | ingredient | value | measure | preparation | Brand Name |
|-----|---------|-------------|-------------------------|--------|-----------|--------------------------------|-------------|
| 1. | 8 | cups | mixed greens | 5 | ounces | - | - |
| 2. | 8 | - | skinless chicken thigh | 1.5 | pounds | - | - |
| 3. | 6.5 | tablespoons | extra-virgin olive oil | - | - | - | - |
| 4. | 6 | ounces | smoked salmon | - | - | thinly sliced, cut into strips | - |
| 5. | 2 | - | whole chicken | 3.5 | pounds | - | - |
| 6. | 20 | ounces | forzen chopped spinach | - | | thawed | - |
| 7. | .5 | cup | parmesean cheese | - | - | grated | - |
| 8. | .5 | cup | pecans | - | - | toasted, finely ground | - |
| 9. | .5 | cup | Bread Crumb Mix, plain | - | - | - | Dixie Diner |
| 10. | 8 | - | garlic clove | 4 | teaspoons | minced | - |
| 11. | 8 | - | green onions | - | - | cut into 2 pieces | - |
|-----|---------|-------------|-------------------------|--------|-----------|--------------------------------|-------------|
Note the diversity of the descriptions. Some things are abbreviated, some are not. Some numbers are numbers, some are spelled out.
I would love something that does a perfect parse/translation. But, would settle for something that does reasonably well to start.
Bonus question: after suggesting a strategy / tool, how would you go about it?
Thanks!
Joe
Short answer. Use GATE.
Long answer. You need some tool for pattern recognition in text. Something, that can catch patterns like:
{Number}{Space}{Ingredient}
{Number}{Space}{Measure}{Space}{"of"}{Space}{Ingredient}
{Number}{Space}{Measure}{Space}{"of"}{Space}{Ingredient}{"("}{Value}{")"}
...
Where {Number} is a number, {Ingredient} is taken from dictionary of ingredients, {Measure} - from dictionary measures and so on.
Patterns I described are very similar to GATE's JAPE rules. With them you catch text that matches pattern and assign some label to each part of a pattern (number, ingredient, measure, etc.). Then you extract labeled text and put it into single table.
Dictionaries I mentioned can be represented by Gazetteers in GATE.
So, GATE covers all your needs. It's not the easiest way to start, since you will have to learn at least GATE's basics, JAPE rules and Gazetteers, but with such approach you will be able to get really good results.
It is basically natural language parsing. (You did already stemming chicken[s].)
So basically it is a translation process.
Fortunately the context is very restricted.
You need a supportive translation, where you can add dictionary entries, adapt the grammar rules and retry again.
An easy process/work flow in this case is much more important than the algorithms.
I am interested in both aspects.
If you need a programming hand for an initial prototype, feel free to contact me. I did see, you are already working quite structured.
Unfortunately I do not know of fitting frameworks. You are doing something, that Mathematica wants to do with its Alpha (natural language commands yielding results).
Data mining? But simple natural language parsing with a manual adaption process should give fast and easy results.
You also can try Gexp.
Then you have to write rules as Java class such as
seq(Number, opt(Measure), Ingradient, opt(seq(token("("), Number, Measure, token(")")))
Then you have to add some group to capture (group(String name, Matcher m)) and extrat parts of pattern and store this information into table.
For Number, Measure you should use similar Gexp pattern, or I would recommend some Shallow parsing for noun phrase detection with words from Ingradients.
If you don't want to be exposed to the nitty-gritty of NLP and machine learning, there are a few hosted services that do this for you:
Zestful (disclaimer: I'm the author)
Spoonacular
Edamam
If you are interested in the nitty-gritty, the New York Times wrote about how they parsed their ingredient archive. They open-sourced their code, but abandoned it soon after. I maintain the most up-to-date version of it and I wrote a bit about how I modernized it.
Do you have access to a tagged corpus for training a statistical model? That is probably the most fruitful avenue here. You could build one up using epicurious.com; scrape a lot of their recipe ingredients lists, which are in the kind of prose form you need to parse, and then use their helpful "print a shopping list" feature, which provides the same ingredients in a tabular format. You can use this data to train a statistical language model, since you will have both the raw untagged data, and the expected parse results for a large number of examples.
This might be a bigger project than you have in mind, but I think in the end it will produce better results than a structured top-down parsing approach will.
Related
My team is building an application which has to solve many user defined formulas. It is a replacement for a huge spreadsheet that our customers use. For e.g. Each formula uses simple arithmetic (mostly) and a few math functions. We are using an expression evaluation library called Parsii to do the actual formula evaluation. But among all the formulas we have to evaluate them in the order of their dependent formula. For e.g.
F1 = a + b
F2 = F1 * 10%
F3 = b / 2
F4 = F2 + F3
In the example above a, b are values input by users. The system should compute F1 & F3 initially since they are directly dependent on user input. Then F3 should be computed. And finally F4.
My question is that what data structure is recommended to model these dependencies of formula evaluation?
We have currently modeled it as a DIRECTED GRAPH. In the example above, F1 & F3 being the root node, and F3 being connected to both, and F4 connected to F3, F4 being the leaf node. We've used the Tinkerpop3 graph implementation to model this.
Any data structure used to model this should have following characteristics.
- Easy to change some input data of few top level root nodes (based on user input)
- Re-calculate only those formulas that are dependent on the root nodes that got changed (since we have 100s of formulas in a specific calculation context and have to respond back to the GUI layer within 1-2 secs)
- Minimize the amount of code to create the data structure via some existing libraries.
- Be able to query the data structure to query/lookup the root nodes by various keys (name of formula object, id of the object, year etc.) and be able to edit the properties of those keys.
Do you store this in a flat file currently?
If you wish to have better queryability, and easier modification, then you could store it as a DAG on database tables.
Maybe something like this (I expect the real solution to be somewhat different):
+-----------------------------------------------------------+
| FORMULA |
+------------+--------------+----------------+--------------+
| ID (PK) | FORMULA_NAME | FORMULA_STRING | FORMULA_YEAR |
+============+==============+================+==============+
| 1 | F1 | a + b | |
+------------+--------------+----------------+--------------+
| 2 | F2 | F1 * 10% | |
+------------+--------------+----------------+--------------+
| 3 | F3 | b / 2 | |
+------------+--------------+----------------+--------------+
| 4 | F4 | F2 + F3 | |
+------------+--------------+----------------+--------------+
+--------------------------------------+
| FORMULA_DEPENDENCIES |
+-----------------+--------------------+
| FORMULA_ID (FK) | DEPENDS_ON_ID (FK) |
+=================+====================+
| 2 | 1 |
+-----------------+--------------------+
| 4 | 2 |
+-----------------+--------------------+
| 4 | 3 |
+-----------------+--------------------+
With this you can also have the security of easily knowing if a formula depends on a non-existent formula because it would violate the DEPENDS_ON_ID foreign key. Also the database can detect if any of the formulas form a cycle of dependencies. Eg where F1 depends on F2 depends on F3 depends on F1.
Additionally you can easily add whatever metadata you wish to the tables and index on whatever you might query on.
this is quite long, and I am sorry about this.
I have been trying to implement the Minhash LSH algorithm discussed in chapter 3 by using Spark (Java). I am using a toy problem like this:
+--------+------+------+------+------+
|element | doc0 | doc1 | doc2 | doc3 |
+--------+------+------+------+------+
| d | 1 | 0 | 1 | 1 |
| c | 0 | 1 | 0 | 1 |
| a | 1 | 0 | 0 | 1 |
| b | 0 | 0 | 1 | 0 |
| e | 0 | 0 | 1 | 0 |
+--------+------+------+------+------+
the goal is to identify, among these four documents (doc0,doc1,doc2 and doc3), which documents are similar to each other. And obviously, the only possible candidate pair would be doc0 and doc3.
Using Spark's support, generating the following "characteristic matrix" is as far as I can reach at this point:
+----+---------+-------------------------+
|key |value |vector |
+----+---------+-------------------------+
|key0|[a, d] |(5,[0,2],[1.0,1.0]) |
|key1|[c] |(5,[1],[1.0]) |
|key2|[b, d, e]|(5,[0,3,4],[1.0,1.0,1.0])|
|key3|[a, c, d]|(5,[0,1,2],[1.0,1.0,1.0])|
+----+---------+-------------------------+
and here is the code snippets:
CountVectorizer vectorizer = new CountVectorizer().setInputCol("value").setOutputCol("vector").setBinary(false);
Dataset<Row> matrixDoc = vectorizer.fit(df).transform(df);
MinHashLSH mh = new MinHashLSH()
.setNumHashTables(5)
.setInputCol("vector")
.setOutputCol("hashes");
MinHashLSHModel model = mh.fit(matrixDoc);
Now, there seems to be two main calls on the MinHashLSHModel model that one can use: model.approxSimilarityJoin(...) and model.approxNearestNeighbors(...). Examples about using these two calls are here: https://spark.apache.org/docs/latest/ml-features.html#lsh-algorithms
On the other hand, model.approxSimilarityJoin(...) requires us to join two datasets, and I have only one dataset which has 4 documents and I would like to figure out which ones in these four are similar to each other, so I don't have a second dataset to join... Just to try it out, I actually joined my only dataset with itself. Based on the result, seems like model.approxSimilarityJoin(...) just did a pair-wise Jaccard calculation, and I don't see any impact by changing the number of Hash functions etc, left me wondering about where exactly the minhash signature was calculated and where the band/row partition has happened...
The other call, model.approxNearestNeighbors(...), actually asks a comparison point, and then the model will identify the nearest neighbor(s) to this given point... Obviously, this is not what I wanted either, since I have four toy documents, and I don't have an extra reference point.
I am running out of ideas, so I went ahead implemented my own version of the algorithm, using Spark APIs, but not much support from MinHashLSHModel model, which really made me feel bad. I am thinking I must have missed something... ??
I would love to hear any thoughts, really wish to solve the mystery.
Thank you guys in advance!
The minHash signatures calculation happens in
model.approxSimilarityJoin(...) itself where model.transform(...)
function is called on each of the input datasets and hash signatures
are calculated before joining them and doing a pair-wise jaccard
distance calculation. So, the impact of changing the number of hash
functions can be seen here.
In model.approxNearestNeighbors(...),
the impact of the same can be seen while creating the model using
minHash.fit(...) function in which transform(...) is called on
the input dataset.
We're using GWT. I have a DataGrid with many repeated values in the column on the left. I would like to hide these. For example:
I have:
Town | Address | Color
------------------------------------------
Springfield | Springfield Heights | Blue
Springfield | Bum town | Red
Springfield | Little Italy | Blue
Shelbyville | Manhattan Square | Green
Shelbyville | Chinatown | Red
I would like to have:
Town | Address | Color
------------------------------------------
Springfield | Springfield Heights | Blue
| Bum town | Red
| Little Italy | Blue
Shelbyville | Manhattan Square | Green
| Chinatown | Red
I tried a few things, but they don't work well with sortable columns. Is there a standard way to do this?
You can override getCellStyleNames for your cell used for Town column. This method gives you Context, which you can use to see where this cell is in the column (context.getIndex()). Using this information, you can compare value in this cell with value in the cell above it (if any). If it is the same, return a style to hide value in this cell.
Note that it won't work if you simply return empty value when overriding getValue for you cell, because it will make the next cell show its value even if it is the same. You can, of course, work around this by looking up until you find a non-empty cell, but overriding getCellStyleNames and simply hiding repeating values looks like a simpler solution.
Because this is a method in your cell, it should behave well with updates, sorting columns, etc.
I am searching a algorithm or general a way to solve the following problem:
Students A,...,M are inscribed for the written examniations of various modules. The inscriptions are given in the following table. If each student can do one exam per day, how many days are necessary, at minimum, for organizing the session?
|A|B|C|D|E|F|G|H|I|J|K|L|M|
Module 1 | | | |X| |X|X| |X|X| | | |
Module 2 |X| | | | |X| | | |X|X| | |
Module 3 | |X| | | | | |X| | |X| |X|
Module 4 |X| | |X| | | | | | | | | |
Module 5 | | |X| |X| | | | |X| | |X|
Module 6 | | |X| | | | |X| | | | | |
Module 7 |X|X| | | | | | |X| |X| | |
Module 8 | | |X| | | |X| | | | |X| |
How an I solve the problem?
With graph colouring.
Make a node for each module, and whenever a student has modules i and j then there is an edge between nodes i and j. Colour the graph, the colours represent days. There is an edge between nodes whenever the modules cannot be on the same day, so the colouring gives a valid schedule. A minimum colouring gives the shortest schedule.
As a suggestion for actually solving the instance (ie, an algorithm for graph colouring), for this size I would take a simple fairly brute force approach, sort of like this:
for k in 1 ..
tryColour(k, 1)
tryColour(k, i):
if i > numnodes:
found it
for c in 1 .. k:
if node i can have colour c:
colours[i] = c
tryColour(k, i+1)
I paid no attention to detail there, it's just for the idea: pick a node, give it a colour that is not immediately impossible, then recursively colour the rest. If the recursive colouring comes up empty, try again with the next colour. Do this whole thing with an increasing number of colours until you find a solution.
Once you have a table of incompatibility, which should look like:
a[1] = [2,4,5,7,8]
a[2] = [1,3,4,5,7]
a[3] = [2,3,5,6,7]
a[4] = [1,2,7,8]
a[5] = [1,2,3,6,8]
a[6] = [3,5,8]
a[7] = [1,2,3,4]
a[8] = [1,5,6]
I think it is something in the idea of:
Create a day-node, put a module in it with its incompatible modules.
Then, as long as any given node still has incompatible modules not resolved:
pop an incompatible module from that node,
either place it in a node that is compatible, or create a new day-node
then remove that module from any other day node in which it was still present
Each day-node has a list of modules that will happen that day, and a list of modules that cannot happen that day. I am not entirely sure how to prove it is optimal, though. It seems to, because it considers the incompatibilities with the modules that have been seen first.
An example of quick and dirty python implementation: https://repl.it/BY2B
I'm performing performance testing of a computer application (JAVA). The test concerns the response time (t) obtained while testing the application with a certain number of concurrent threads (th) and a certain amount of data (d).
Suppose I have the following results:
+------+-------+-----+
| th | d | t |
+------+-------+-----+
| 2 | 500 | A |
+------+-------+-----+
| 4 | 500 | B |
+------+-------+-----+
| 2 | 1000 | C |
+------+-------+-----+
| 4 | 1000 | D |
+------+-------+-----+
How can i benefit the most of these results such as knowing the limit of my app as well as creating meaningful graphs to represent these results.
I'm not a statistics person so pardon my ignorance. Any suggestions would be really helpful (even related statistics technical keywords I can Google).
Thanks in advance.
EDIT
The tricky part for me was to determine the application's performance evolution taking both the number of threads and the amount of data into consideration in one plot.
Yes there is a way, check the following example I made with paint (the numbers I picked are just random):