I was hoping to brainstorm a little bit on the subject of storing n-gram data. In my project, I am trying to solve linguistic problems where I know all (n-1) data items and want to statistically guess my n using linear interpolation over all applicable n-grams. (Yes, there is a tagger that assigns tags to known words according to its lexicon and a suffix tree that tries to guess the word kind for unknown words; the n-gram component discussed here will be tasked with resolving ambuguity.)

My initial approach would be to simply store all observed n-grams (for n = 1..3, i.e. monogram, bigram, trigram) data in respective SQL databases and call it a day. But the requirements of my project may change to include other vector lengths (n), and I would like my application to adapt to 4-gram without a lot of work (updating schema, updating application code, etc.); ideally, I would simply tell my application to work with 4-grams now without having to change code much (or at all) and train its data from a given data source.

To sum up all requirements:

  • Ability to store n-gram data (initially for n = {1, 2, 3}
  • Ability to change what kinds of n-grams should be used (between application runs)
  • Ability to (re-)train n-gram data (between application runs)
  • Ability to query the data store (e.g. if I have observed A, B, C, I'd like to know the most frequently observed item for what might follow using my trained 4-, 3-, 2-, 1-gram data sets)

    The application will most likely be read-heavy, data sets most likely won't be retrained that often

  • The solution employs the .NET Framework (up to 4.0)

Now what design would be better fit for such a task?

  • A fixed table managed by a SQL server (MSSQL, MySQL, ...) for each n (eg. dedicated tables for bi-grams, tri-grams, etc.)
  • Or a NoSQL document database solution that stores the first n-1 as the key of the document, and the document itself contains the n-th value and observed frequencies?
  • Or something different?
  • 3
    I think this would be better suited on Stack Overflow. Commented Apr 1, 2011 at 13:45
  • 1
    Perhaps a trie (prefix tree) data structure would fit your requirements?
    – Schedler
    Commented Apr 1, 2011 at 14:09
  • 1
    I'd suggest Stack Overflow or even cstheory.stackexchange.com
    – Steve
    Commented Apr 1, 2011 at 16:04
  • Okay, thanks. I'll try to get the question up over there.
    – Manny
    Commented Apr 4, 2011 at 11:36
  • 5
    This question is perfectly suited for programmers.stackexchange.com and should not be migrated to stackoverflow, IMO. It's exactly the kind of “whiteboard situation“ question that should be asked here. Check meta for details.
    – user281377
    Commented Jul 10, 2011 at 8:56

4 Answers 4


Given that you won't know the optimal range of N, you definitely want to be able to change it. For example, if your application predicts the likelihood that a certain text is English, you would probably want to use character N-grams for N 3..5. (That's what we found experimentally.)

You haven't shared details about your application, but the problem is clear enough. You want to represent N-gram data in a relational database (or NoSQL document-based solution). Before suggesting a solution of my own, you may want to take a look at the following approaches:

  1. How to best store Google ngrams in a database?
  2. Storing n-grams in database in < n number of tables
  3. Managing the Google Web 1T 5-gram with Relational Database

Now, having not read any of the above links, I suggest a simple, relational database approach using multiple tables, one for each size of N-gram. You could put all of the data in a single table with the maximum necessary columns (i.e. store bigrams and trigrams in ngram_4, leaving the final columns null), but I recommend partitioning the data. Depending on your database engine, a single table with a large number of rows can negatively impact performance.

  create table ngram_1 (
      word1 nvarchar(50),
      frequency FLOAT,
   primary key (word1));

  create table ngram_2 (
      word1 nvarchar(50),
      word2 nvarchar(50),
      frequency FLOAT,
   primary key (word1, word2));

  create table ngram_3 (
      word1 nvarchar(50),
      word2 nvarchar(50),
      word3 nvarchar(50),
      frequency FLOAT,
   primary key (word1, word2, word3));

  create table ngram_4 (
      word1 nvarchar(50),
      word2 nvarchar(50),
      word3 nvarchar(50),
      word4 nvarchar(50),
      frequency FLOAT,
   primary key (word1, word2, word3, word4));

Next, I'll give you a query that will return the most probable next word given all your ngram tables. But first, here is some sample data that you should insert into the above tables:

  INSERT [ngram_2] ([word1], [word2], [frequency]) VALUES (N'building', N'with', 0.5)
  INSERT [ngram_2] ([word1], [word2], [frequency]) VALUES (N'hit', N'the', 0.1)
  INSERT [ngram_2] ([word1], [word2], [frequency]) VALUES (N'man', N'hit', 0.2)
  INSERT [ngram_2] ([word1], [word2], [frequency]) VALUES (N'the', N'bat', 0.7)
  INSERT [ngram_2] ([word1], [word2], [frequency]) VALUES (N'the', N'building', 0.3)
  INSERT [ngram_2] ([word1], [word2], [frequency]) VALUES (N'the', N'man', 0.4)
  INSERT [ngram_2] ([word1], [word2], [frequency]) VALUES (N'with', N'the', 0.6)
  INSERT [ngram_3] ([word1], [word2], [word3], [frequency]) VALUES (N'building', N'with', N'the', 0.5)
  INSERT [ngram_3] ([word1], [word2], [word3], [frequency]) VALUES (N'hit', N'the', N'building', 0.3)
  INSERT [ngram_3] ([word1], [word2], [word3], [frequency]) VALUES (N'man', N'hit', N'the', 0.2)
  INSERT [ngram_3] ([word1], [word2], [word3], [frequency]) VALUES (N'the', N'building', N'with', 0.4)
  INSERT [ngram_3] ([word1], [word2], [word3], [frequency]) VALUES (N'the', N'man', N'hit', 0.1)
  INSERT [ngram_3] ([word1], [word2], [word3], [frequency]) VALUES (N'with', N'the', N'bat', 0.6)
  INSERT [ngram_4] ([word1], [word2], [word3], [word4], [frequency]) VALUES (N'building', N'with', N'the', N'bat', 0.5)
  INSERT [ngram_4] ([word1], [word2], [word3], [word4], [frequency]) VALUES (N'hit', N'the', N'building', N'with', 0.3)
  INSERT [ngram_4] ([word1], [word2], [word3], [word4], [frequency]) VALUES (N'man', N'hit', N'the', N'building', 0.2)
  INSERT [ngram_4] ([word1], [word2], [word3], [word4], [frequency]) VALUES (N'the', N'building', N'with', N'the', 0.4)
  INSERT [ngram_4] ([word1], [word2], [word3], [word4], [frequency]) VALUES (N'the', N'man', N'hit', N'the', 0.1)

To query the most probable next word, you would use a query like this.

  DECLARE @word1 NVARCHAR(50) = 'the'
  DECLARE @word2 NVARCHAR(50) = 'man'
  DECLARE @word3 NVARCHAR(50) = 'hit'
  DECLARE @bigramWeight FLOAT = 0.2;
  DECLARE @trigramWeight FLOAT = 0.3
  DECLARE @fourgramWeight FLOAT = 0.5

  SELECT next_word, SUM(frequency) AS frequency
  FROM (
    SELECT word2 AS next_word, frequency * @bigramWeight AS frequency
    FROM ngram_2
    WHERE word1 = @word3
    SELECT word3 AS next_word, frequency * @trigramWeight AS frequency
    FROM ngram_3
    WHERE word1 = @word2
      AND word2 = @word3
    SELECT word4 AS next_word, frequency * @fourgramWeight AS frequency
    FROM ngram_4
    WHERE word1 = @word1
      AND word2 = @word2
      AND word3 = @word3
    ) next_words
  GROUP BY next_word
  ORDER BY SUM(frequency) DESC

If you add more ngram tables, you will need to add another UNION clause to the above query. You might notice that in the first query I used word1 = @word3. And in the second query, word1 = @word2 AND word2 = @word3. That's because we need to align the three words in the query for the ngram data. If we want the most likely next word for a sequence of three words, we'll need to check the first word in the bigram data against the last word of the words in the sequence.

You can tweak the weight parameters as you wish. In this example, I assumed that higher ordinal "n" grams will be more reliable.

P.S. I would structure the program code to handle any number of ngram_N tables via configuration. You could declaratively change the program to use N-gram range N(1..6) after creating the ngram_5 and ngram_6 tables.

  • With this query, i see only the frequency score you have here. How do i select the next predictive word. Which is the most relevance to the sentence?
    – TomSawyer
    Commented Dec 1, 2015 at 21:48
  • Good point @TomSawyer. I added sample data to the answer and gave a sample query that returns the most probable next word. Commented Dec 1, 2015 at 22:41
  • Tks for your update. But how can we calculate the frequency here? ie: in ngram_2 , the phrase building with has freq is 0.5. Same question with @bigramWeight , what is that?. I though freq is the field will be update everytime we update the database. Ie if user enter more string, the frequency for this string will be recalculate? 0.5 is 0.5 percent in total used times or appearance rate of each phrase?
    – TomSawyer
    Commented Dec 2, 2015 at 6:49
  • The bigramWeight and trigramWeight (etc) are how to weight the different n-grams in the overall calculation. It's a simplistic way of saying that longer n-grams have higher entropy and you may want them to "count" more than shorter n-grams. Commented Dec 4, 2015 at 14:08
  • In terms of updating the database, obviously I haven't covered all the details and there is lots of room for improvement. For example, rather than storing nvarchars in the ngram tables, you'd probably want to tokenize into a words table (word_id INT, word NVARCHAR) and then refer to word_ids in the ngram tables. To update the tables on retraining, that's right -- you'd just update the frequency field. Commented Dec 4, 2015 at 14:10

Contrary to what the others are suggesting, I'd suggest to avoid any data structures more complex than a hashmap or a key-value store.

Keep in mind your data access requirements: a) 99% requests - query ngram "aaa-bbb-ccc" and retreive the value (or 0) b) 1% requests - inserting/updating a count of specific ngram c) there is no (c).

The most effective way is to retrieve it with a single lookup. You can use an out-of-bounds (or escaped) separator to combine the full n-gram in a single string (e.g. "alpha|beta|gamma" for 3gram, "alpha" for unigram, etc) and just fetch that (by the hash of that). That's how quite a lot of NLP software does it.

If your ngram data is small (say, < 1 gb) and fits in memory , then I'd suggest to use an efficient in-program memory structure (hashmaps, trees, tries, etc) to avoid overhead; and just serialize/deserialize to flat files. If your ngram data is terabytes or more, then you may choose NoSQL key-value stores split on multiple nodes.

For extra performance, you may want to replace all words everywhere with integer ids so that your core algorithm doesn't see any (slow) strings at all; then it's slightly different to implement the same idea.


Not the most efficient, but simple and wedded to the database like you want:

Table: word
word (int, primary key) - a unique identifier for each word
text (varchar) - the actual word

Table: wordpos
document (int) - a unique identified for the document of this word
word (int, foreign key to word.word) - the word in this position
pos (int) - the position of this word (e.g., first word is 1, next is 2, ...)

wordpos should have indexes on document and pos.

bigrams are:

select word1.text as word1, word2.text as word2
from wordpos as pos1, wordpos as pos2, word as word1, word as word2
where pos1.document = pos2.document
      and pos1.pos = pos2.pos - 1
      and word1.word = pos1.word
      and word2.word = pos2.word

Then you can count() and group your way to frequencies and stuff.

To change to trigrams, it is easily to generate this string to include a word3.

I've done this before actually (even though the SQL up there is probably a little rusty). I settled on a set of flat files that could be seeked into easily then streamed off of disk. Kinda depends on your hardware how to do it better.


While trying to improve my applications' simple searches to bigrams and trigrams from unigrams, essentially, I saw your question.

If one of requirements is ability to query a distributed file system or database, then this might be interesting for you too: the paper Pibiri and Venturini 2018 "Handling Massive N-Gram Datasets Efficiently" outlines an efficient way to store n-gram data in terms of runtime and space. They've offered their implementation at https://github.com/jermp/tongrams

Each "n" of n-grams is held in a separate table accessed by a minimum perfect hash function with very fast select and query abilities. The tables are static and built by the main code using input in format of Google n-grams text files.

I haven't used the code yet, but there are many ways you could with your open requirements of where your queries are from.

One way: if the .NET equivalent of a servlet is used with a database or datastore, and if you need to conserve storage space, then storing each ngram table in binary form in the database/datastore as a table is one option (one database/datastore table for the resulting static file of the efficient ngram code for all 1-grams, another for all 2-grams, etc). Queries would be run by invoking the efficient n-gram code (wrapped to be accessible by your servlet). It's a work-around to making a distributed database that is using the efficient n-gram code to access the files on a distributed file system. Note that the binary database/datastore tables each have the file-size restriction of the underlying file-system.

Not the answer you're looking for? Browse other questions tagged or ask your own question.