(r66) GrmNGramForum < Forum

Forum Web>GrmNGramForum (revision 66) (raw view)~~EditAttach~~
---+ [[GRM.NGramLibrary][OpenGrm NGram]] Forum

You need to be a  registered user to participate in the discussions. <br /> 
%IF{"context authenticated" then='%MAKETEXT{"You are logged in as [_1]." args="[[%WIKIUSERNAME%][%SPACEOUT{%WIKINAME%}%]]"}%%BR%%LOGOUT%' else='%LOGIN% %MAKETEXT{"or"}% [[%TWIKIWEB%.TWikiRegistration][%MAKETEXT{"Register"}%]]'}%

%TOC{depth="2"}%

You can start a new discussion here:
<blockquote>
%ICON{help}% You can use the formatting commands describes in TWiki.TextFormattingRules in your comment. <br /> 
%ICON{tip}% If you want to post some code, surround it with =&lt;verbatim&gt;= and =&lt;/verbatim&gt;= tags. <br /> 
%ICON{warning}% Auto-linking of TWiki.WikiWords is now disabled in comments, so you can type !VectorFst and it won't result in a broken link.<br /> 
%ICON{warning}% You now need to use =&lt;br&gt;= to force new lines in your comment (unless inside =verbatim= tags). However, a blank line will automatically create a new paragraph. </blockquote> 
%COMMENT{type="newtopic"}%



---++ Negative weights associated with epsilon transitions 
---+++ Main.VinodhR - 2014-04-02 - 16:36
<noautolink>
I understand that the transition weights in the ngram model are negative log probabilities. (correct me if I am wrong)

But I can see that the epsilon transitions also take negative weights. What do they actually represent ?
</noautolink>


---+++ Main.BrianRoark - 2014-04-03 - 12:53
<noautolink>
epsilon transitions are for backoff, and the weights on those arcs are negative log alpha, where alpha is the backoff constant that ensures normalization over all words leaving that state.  Ideally, those backoff transitions are interpreted as failure transitions, i.e., only traversed if the symbol does not label a transition leaving the state.  See the ngramapply command for an example of the use of a phi matcher to handle failure transitions.

Brian
</noautolink>
%COMMENT{type="mythreadmode"}% 



---++ Constraining the minimum length of sentences in random generating 
---+++ Main.VivekKulkarni - 2014-03-18 - 21:31
<noautolink>
While sampling sentences using ngramrandgen do we have an option to constrain the minimum length of sentence generated just as we have for maximum length 
</noautolink>


---+++ Main.BrianRoark - 2014-03-19 - 09:10
<noautolink>
Sorry, no, we don&#39;t have that option.  The solution is to oversample, then throw away strings with length less than your minimum using grep or something like that.
</noautolink>
%COMMENT{type="mythreadmode"}% 



---++ Models without &lt;s&gt; or &lt;/s&gt; 
---+++ Main.PaidiCreed - 14 Jan 2014 - 12:34
<noautolink>
Is it possible to build models from raw ngram counts which do not contain the symbols for start and end of sentence?
</noautolink>


---+++ Main.BrianRoark - 16 Jan 2014 - 14:26
<noautolink>
Hi,

To build an automaton that has valid paths, you need to specify some kind of termination cost, which is what the final symbol allows for.  So, you can't really do without having an entry for end-of-string in your raw n-gram counts; otherwise, it will represent this as an automaton that accepts no strings (empty) since no string can ever terminate.  If you have a large set of counts without start or stop symbols, you should just add a unigram of the end-of-string symbol with some count (maybe just 1).  The start symbol you can do without (though it is often an important context to include).  Hope that helps.

brian
</noautolink>
%COMMENT{type="mythreadmode"}% 



---++ Perplexity without &lt;s&gt; or &lt;/s&gt;
---+++ Main.ErKn - 05 Sep 2013 - 03:35
<noautolink>
I have been using ngramapply and ngramperplexity tools. I am wondering if there is an option to specify not to use the ending state </s> in the calcul of logprob.
</noautolink>

---+++ Main.ErKn - 05 Sep 2013 - 03:36
<noautolink>
ending state = &lt;/s&gt;
</noautolink>

---+++ Main.BrianRoark - 05 Sep 2013 - 09:44
<noautolink>
Hi,

if you run it in verbose mode (ngramperplexity --v=1 as shown in the quick tour) then you can see the -log probability of each item in the string.  You can then read this text into your own code to calculate perplexity in whatever way you wish.  However, typically the end of string marker is included in standard perplexity calculations, and should be included if you are comparing with other results.  No options to exclude it in the library.  Hope that helps.

brian
</noautolink>

---+++ Main.ErKn - 05 Sep 2013 - 10:10
<noautolink>
Hi,

Thanks you for the hint. Could you tell me why I get different perplexity values when running ngramapply and ngramperplexity tools (in verbose mode) with same data/parameters.
</noautolink>

---+++ Main.JosefNovak - 06 Sep 2013 - 06:28
<noautolink>
=ngramapply= only applies the LM weights to your input token sequence, it does not compute perplexity.

The total weight output by =ngramapply= for each string should be equal to the logprob(base 10) generated in the verbose output for =ngramperplexity= (but note that the ngram apply value is negative log using base =e=).

If you want to convince yourself you can try this:
<verbatim>
#Apply the model to your test data
$ ngramapply toy.mod test.far test.scored.far
#Extract the individual results
$ farextract --filename_prefix=tst test.scored.far
#Print out one of the sentences, which now have individual arc weights
$ fstprint tsttest.sent-1
0	1	a	a	0.750776291
1	2	b	b	1.20206916
2	3	a	a	1.43548453
3	4	a	a	0.864784181
4	5	a	a	0.864784181
5	1.27910674
#Add all the values up and compute the perplexity in the normal manner
$ python
>>> import math
>>> -1.*(0.750776291+1.20206916+1.43548453+0.864784181+0.864784181+1.27910674)/math.log(10)
-2.778184008253953
>>> math.pow(10, ((0.750776291+1.20206916+1.43548453+0.864784181+0.864784181+1.27910674)/math.log(10))/(5+1))
2.9042277315216354
</verbatim>
Those values should match whatever you get out of =ngramperplexity= (again careful about the log base) in verbose mode for each individual sentence.
In my toy example case above:
<verbatim>
$ ngramperplexity --v=1 toy.mod test.far
WARNING: OOV probability = 0; OOVs will be ignored in perplexity calculation
a b a a a
                                                ngram  -logprob
        N-gram probability                      found  (base10)
        p( a | <s> )                         = [2gram]  0.326058
        p( b | a ...)                        = [2gram]  0.522052
        p( a | b ...)                        = [1gram]  0.623423
        p( a | a ...)                        = [2gram]  0.375571
        p( a | a ...)                        = [2gram]  0.375571
        p( </s> | a ...)                     = [2gram]  0.555509
1 sentences, 5 words, 0 OOVs
logprob(base 10)= -2.77818;  perplexity = 2.90423
</verbatim>
</noautolink>
%COMMENT{type="mythreadmode"}% 



---++  Experimenting with =ngramnormalize=
---+++ Main.JosefNovak - 02 Sep 2013 - 08:38
<noautolink>
I have been experimenting a bit with the new =ngrammarginalize= tool - very cool.  I noticed that in some cases it can get stuck in the =do{}while= loop in the =StationaryStateProbs= function.  

I trained a model using a small amount of data and witten_bell smoothing + pruning.  It may be worth noting that the data is very noisy.  
<br>
I have been digging around in the paper and source code a bit to figure out what might be the issue.  The issue seems to be here:
<verbatim>
if (fabs((*probs)[st] - last_probs[st]) > converge_eps * last_probs[st] )
</verbatim>

where =last_probs[st]= can occasionally be a very small negative value.  If I am following the paper correctly, this corresponds to the epsilon value mentioned at the end of Section 4.2.  (I'm not entirely convinced I followed correctly to this point though so please correct me if I'm wrong).  
<br>
Anyway, if =last_probs[st]= turns out to be negative, for whatever reason, then there is a tendency to get stuck in this block.  This also seems to be affected by the fact that the comparison delta is computed relative to =last_probs[st]= , so if =last_probs[st]= happens to get smaller with each iteration, then the comparison also becomes more 'sensitive', in the sense that a smaller absolute difference will still evaluate to 'true'.  So as we iterate, it seems that some values become more sensitive to noise (maybe this is the numerical error you mention?) and the process gets stuck.

I noticed that if I set the comparison to an absolute:
<verbatim>
if (fabs( fabs((*probs)[st]) - fabs(last_probs[st])) > converge_eps )
</verbatim>
then it always terminates, and does so quickly for each iteration, even for very small values of =converge_eps=.

I have not convinced myself that this is a theoretically acceptable solution, nor have I validated the resulting models, but it does some reasonable at a superficial level.

I would definitely appreciate some feedback if available.
</noautolink>


---+++ Main.JosefNovak - 02 Sep 2013 - 11:38
<noautolink>
EDIT: this second issue may be ignored.  It was my fault.

Also, I notice that I am still getting 'unplugged' 'holes' some of my pruned models when I run ngram-read after pruning with SRILM.

Specifically, I prune with srilm, then run: =ngramread= it terminates without issue. If I run the info command I get:

<verbatim>
$ ngraminfo l2p5k.train.p5.mod
FATAL: NGramModel: bad ngram model topology
</verbatim>

and if I run the marginalize command (without any of the modifications I mentioned above) I get:
<verbatim>
$ ngrammarginalize --v=3 --iterations=2  --norm_eps=0.001 l2p5k.train.p5.mod  l2p5k.train.p5.marg.mod
INFO: Iteration #1
INFO: FstImpl::ReadHeader: source: l2p5k.train.p5.mod, fst_type: vector, arc_type: standard, version: 2, flags: 3
INFO: CheckTopology: bad backoff arc from: 557 with order: 4 to state: 0 with order: 1
FATAL: NGramModel: bad ngram model topology
</verbatim>

Maybe there is a flag I am missing?  Or I should just pre-process and plug them manually prior to conversion?
</noautolink>

---+++ Main.BrianRoark - 02 Sep 2013 - 22:21
<noautolink>
Hi Josef,

For the second issue, I would be interested in seeing the original ARPA format file and follow this through to see what the potential issue might be.  So please email or point me to that if you can.  For the first issue, I have seen some problems with the convergence of the steady state probs for Witten-Bell models in particular, which is due, I believe, to under-regularization (as we say in the paper).  Your tracing this to the value of last_probs[st] is something I hadn't observed, so that could be very useful.  In answer to your question, I wonder if there's not a way to maintain the original convergence criterion, but control for the probability falling below zero (presumably due to floating point issues).  One way to do this would be to set a very small minimum probability for the state and set a state's probability to that minimum if the calculated value falls below it.  Since last_probs[st] gets its value from (*probs)[st], then it would never fall below 0 either.  I'll look into this, too, thanks.

brian
</noautolink>

---+++ Main.JosefNovak - 04 Sep 2013 - 03:27
<noautolink>
Hi, The 2nd issue was my own embarrassing mistake.  I was setting up a test distribution to share, and realized that I still had an old, duplicate version of =ngramread= on my =$PATH=.  This was the source of the conversion issue.  When I switched it things worked as expected for all models.

I will share an example of the other via email.  
</noautolink>
%COMMENT{type="mythreadmode"}% 



---++ How to handle unknown words? 
---+++ Main.GeorgePetasis - 29 Aug 2013 - 07:18
<noautolink>
Hi, I am new to this librbary. I have created a small language model using a corpus. Is there a solution in applying the model in user input, which may contain an unknown word?

How can I handle unknown words?
</noautolink>


---+++ Main.BrianRoark - 29 Aug 2013 - 10:02
<noautolink>
Hi,

when you apply the model to new text, you are typically using farcompilestrings to compile each string into FSTs.  It has a switch (--unknown_symbol) to map out of vocabulary (OOV) items to a particular symbol.  The language model then needs to include that symbol, typically as some small probability at the unigram state.  There are a couple of ways to do this -- see the discussion topic below entitled: Model Format: OOV arcs.

brian
</noautolink>
%COMMENT{type="mythreadmode"}% 



---++ ngramrandgen with a prefix 
---+++ Main.GiulianoLancioni - 17 Aug 2013 - 11:45
<noautolink>
Is it possible to constraint the output of ngramrandgen by requiring that it continues a (well-formed) prefix, as it is possible in SRILM? Thank you?
</noautolink>


---+++ Main.BrianRoark - 17 Aug 2013 - 14:40
<noautolink>
Hello,

that is certainly possible, and we'll add it to the list of features for a future release.  Likely not in the soon-to-be-released latest version, but perhaps in the next update after that.

brian
</noautolink>

---+++ Main.GiulianoLancioni - 17 Aug 2013 - 15:36
<noautolink>
Brian,

thank you for your answer. I am no expert of FSTs, but I guess It should be possible to obtain the same result by somewhat filtering the complete FST, through composition (?) with another FST. After all, it should look like a FST where some paths have already been walked through. Please excuse my vagueness.
</noautolink>

---+++ Main.GiulianoLancioni - 17 Aug 2013 - 15:49
<noautolink>
For instance, fstintersect of the ngram FST with an acceptor for the prefix could do the trick, couldn't it?
</noautolink>

---+++ Main.BrianRoark - 18 Aug 2013 - 18:41
<noautolink>
yes, this is the way to think about it.  The complication comes from the random sampling algorithm, and the way that it interprets the backoff arcs in the model.  The algorithm assumes a particular model topology during the sampling procedure, which will not be preserved in the composed machine that you propose.  But, yes, essentially that is what would be done to modify the algorithm.  

brian
</noautolink>

---+++ Main.BrianRoark - 18 Aug 2013 - 18:42
<noautolink>
and, of course, the benefit of open-source is that you could play around with the algorithm yourself, if you like.
</noautolink>

---+++ Main.GiulianoLancioni - 20 Aug 2013 - 06:38
<noautolink>
Thanks!
</noautolink>
%COMMENT{type="mythreadmode"}% 



---++ skip-grams 
---+++ Main.GezaKiss - 24 Jul 2013 - 14:12
<noautolink>
Do you support or plan to support skip-grams? Thanks!
</noautolink>


---+++ Main.BrianRoark - 25 Jul 2013 - 10:23
<noautolink>
Hi Geza,

Skip k-grams generally require a model structure that is trickier to represent compactly in a WFST than standard n-gram models.  This is because there are generally more than one state to backoff to.  For example, in a trigram state, if the specific trigram 'abc' doesn't exist, a standard backoff n-gram model will backoff to the bigram state and look for 'bc'.  With a skip-gram, there is also the skip-1 bigram 'a_c' to look for, i.e., there are two backoff directions.  So, to answer your question, it is something we've thought about and are thinking about, but nothing imminent.

brian
</noautolink>
%COMMENT{type="mythreadmode"}% 



---++ calculating perplexity for >10 utterances using example command 
---+++ Main.GezaKiss - 19 Jul 2013 - 14:09
<noautolink>
I am a newbie who wants to calculate perplexity for a text file consisting of more than 10 lines. The examples provided for that only works for < 10, and it is not obvious to me how to bypass that. (Yes, I know I can use a loop over the utterances, I just hope I do not have to.)

Using stdin with farcompilestrings results in an error message:
FATAL: STListWriter::Add: key disorder: 10
FATAL: STListReader: error reading file: stdin

And specifying a text file as the first parameter for farcompilestrings does not work either:
FATAL: FarCompileStrings: compiling string number 1 in file test.txt failed with token_type = symbol and entry_type = line
FATAL: STListReader::STListReader: wrong file type:

I could not find related usage info. I'd appreciate some help.
Thanks!

</noautolink>


---+++ Main.GezaKiss - 19 Jul 2013 - 14:40
<noautolink>
The example command would be this:

echo -e "A HAND BAG\nA HAND BAG\nA HAND BAG\nA HAND BAG\nA HAND BAG\nA HAND BAG\nA HAND BAG\nA HAND BAG\nA HAND BAG\nBAG HAND A" |\
    farcompilestrings -generate_keys=1 -symbols=earnest.syms --keep_symbols=1 |\
    ngramperplexity --v=1 earnest.mod -

</noautolink>

---+++ Main.BrianRoark - 20 Jul 2013 - 14:43
<noautolink>
Hi Geza,

the -generate_keys=N switch for farcompilestrings creates numeric keys for each of the FSTs in the FAR file using N digits per FST.  (One FST for each string in your case.)  So, with N=1 you can index 0-9 strings, with N=2 you can index 0-99 strings, etc.  So, for your example, you just need to up your -generate_keys argument with the number of digits in the total corpus count.

brian
</noautolink>

---+++ Main.GezaKiss - 24 Jul 2013 - 13:56
<noautolink>
Thanks, Brian!

One more question, just so that I can see more clearly how ngramperplexity works. 
When I specify the "--OOV_probability=0.00001 --v=1" arguments, I get "[OOV]    9" for "ngram  -logprob" in the output when using unigrams, and somewhat higher values for longer n-grams. What is the reason for his? (I guess I do not yet see how exactly perplexity calculation works.)

Thanks, Géza
</noautolink>

---+++ Main.GezaKiss - 24 Jul 2013 - 14:22
<noautolink>
Sorry, I was too quick to write. (The solution is of course --OOV_class_size=10000, hence the 4 difference from the expected.)
</noautolink>

---+++ Main.GiulianoLancioni - 17 Aug 2013 - 04:48
<noautolink>
As to the FarCompileStrings fatal error signaled by Géza, it depends from file encoding: text files must be encoded with Unix, noth Windows EOL.
</noautolink>
%COMMENT{type="mythreadmode"}% 



---++ Model Format: OOV arcs 
---+++ Main.SamS - 15 Jul 2013 - 00:31
<noautolink>
When I load a pre-trained language model into c++ and iterate over arcs, I see that there are no arcs labeled with the given OOV symbol, although it is contained in the symbol table.  

Are OOV words handled somewhere other than the FST itself, or is the absence of these arcs likely due to a quirk in my particular language model?  

Any insight or ideas would be greatly appreciated.  Thanks!
</noautolink>


---+++ Main.BrianRoark - 15 Jul 2013 - 22:37
<noautolink>
Hi,

The model will only have probability mass for the OOV symbol if there are counts for it in the training corpus.  ngramperplexity does have a utility for including an OOV probability, but this is done on the fly, not in the model structure itself.  If you want to provide probability mass at the unigram state for the OOV symbol, you could create a corpus consisting of just that symbol, train a unigram model, then use ngrammerge to mix (either counts or model) with your main model.  Then there would be explicit probability mass allocated to that symbol.  You can use merge parameters to dictate how much probability that symbol should have.  Hope that helps.

brian
</noautolink>

---+++ Main.SamS - 16 Jul 2013 - 13:09
<noautolink>
This absolutely helps.  Thanks for the answer!
</noautolink>
%COMMENT{type="mythreadmode"}% 



---++ OpenGrm in c++ application 
---+++ Main.EricYoo - 03 Jul 2013 - 08:42
<noautolink>
Hi,
I built 3-gram language model on few English words.
My c++ program receive streaming character one by one. 
I would like to use the 3-gram model to score the up-coming character with history context. What example can I start with ? Is it possible not to convert to farstrings each time ?
</noautolink>


---+++ Main.BrianRoark - 15 Jul 2013 - 22:32
<noautolink>
Hi,

This is one of the benefits of having the open-source library interface in C++, you can write functions of your own.  We choose to score strings (when calculating perplexity for example) when encoding the strings as fars, but you could perform a similar function in your own C++ code.  I would look at the code for ngramperplexity as a starting point, and learn how to use the arc iterators.  Once you understand the structure of the model, you should be able to make that work.  Alternatively, print out the model using fstprint and read it into your own data structures.  Good luck!

brian
</noautolink>
%COMMENT{type="mythreadmode"}% 



---++ How to understand the order one count result? 
---+++ Main.HeTianxing - 08 Jun 2013 - 10:06
<noautolink>
I tried order-1 ngram count with this simple text <br>
Goose is hehe <br>
Goose is hehe <br>
Goose is <br>
Goose <br>
But I don't understand the resulting count fst.<br>
 0 -1.3863 <br>
0 0 Goose Goose -1.3863 <br>
0 0 is is -1.0986 <br>
0 0 hehe hehe -0.69315 <br>
The document says Transitions and final costs are weighted with the negative log count of the associated n-gram, but I can't make sense with these numbers, can someone help me out? Thx!!
</noautolink>


---+++ Main.BrianRoark - 09 Jun 2013 - 10:12
<noautolink>
Hi,

The counts are stored as negative natural log (base e), so -0.69315 is -log(2), -1.0986 is -log(3) and -1.3863 is -log(4).  The count of each word is kept on arcs in a single state machine (since this is order 1) and the final cost is for the end of string (which occurred four times in your example).  You printed this using fstprint, but you can also try ngramprint which in this case yields:<br>

&lt;s&gt;	4<br>
Goose	4<br>
is	3<br>
hehe	2<br>
&lt;/s&gt;	4<br>

where &lt;s&gt; and &lt;/s&gt; are the implicit begin of string and end of string events.  These are implicit because we don't actually use the symbols to encode them in the fst.

Hope that clears it up for you.  If not, try the link in the 'Model Format' section of the quick tour, to the page on 'precise details'.

best,

brian
</noautolink>
%COMMENT{type="mythreadmode"}% 



---++  
---+++ Main.HuanliangWang - 15 May 2013 - 03:02
<noautolink>
Hi, 

    I try to merge two LM by following command line:

    ./tools/opengrm-ngram-1.0.3/bin/ngrammerge --alpha=0.2 --beta=0.8 --normalize --use_smoothing A.fst B.fst AB.mrg.fst

    But I get a error:

    FATAL: NGramModel: input not deterministic

    A.fst is normal fst LM trained by SRILM. B.fst is class-expanded LM by fstreplace command. 

     I also try to convert fst LM into arpa LM by command line:

      ./tools/opengrm-ngram-1.0.3/bin/ngramprint --ARPA B.fst > B.arpa

      But I got a same error:

      FATAL: NGramModel: input not deterministic    

      How to fix the error?

Thanks,

Huanliang wang
</noautolink>


---+++ Main.BrianRoark - 16 May 2013 - 22:17
<noautolink>
Hi,

you've introduced non-determinism into the ngram models via your replace class modification.  The ngrammerge (and ngramprint) commands are simple operations that expect a standard n-gram topology, hence the error messages.  For more complex model topologies of the sort you have, you'll have to write your own model merge function that does the right thing when presented with non-determinism.  The base library functions don't handle these complex examples, but the code should give you some indication of how to approach such a model mixture.  Such is the benefit of open source!

brian
</noautolink>
%COMMENT{type="mythreadmode"}% 



---++ error when converting LM genereted by HTK into fst format 
---+++ Main.HuanliangWang - 07 May 2013 - 03:08
<noautolink>
Hi, 

   I try to convert arpa LM genereted by HTK tool into fst format. The command is : 

   ./tools/opengrm-ngram-1.0.3/bin/ngramread --ARPA test.arpa > test.lm.fst

   But I get a error:  

         FATAL: NGramInput: Have a backoff cost with no state ID!
   
   How to fix the error?

Thanks,

Huanliang wang
</noautolink>


---+++ Main.BrianRoark - 07 May 2013 - 22:18
<noautolink>
Hi,

it appears that you have n-grams ending in your stop symbol (probably &lt;/s&gt;) that have backoff weights, i.e., the ARPA format has an n-gram that looks like:

-1.583633       XYZ &lt;/s&gt;       -0.30103

But &lt;/s&gt; means end-of-string, which we encode as final cost, not an arc leading to a new state.  Hence there is no state where that backoff cost would be used.  (Think of it this way: what's the next word you predict after &lt;/s&gt;?  In the standard semantics of &lt;/s&gt;, it is the last term predicted, so nothing comes afterwards.)  Do you also have n-grams that start with &lt;/s&gt;?

So, one fix on your ARPA format is just to remove the backoff weight after n-grams that end in &lt;/s&gt;.

hope that helps,

brian
</noautolink>

---+++ Main.HuanliangWang - 07 May 2013 - 23:08
<noautolink>
Hi, Brian
Thank you! I get it. 

Another case is there are n-grams that start with &lt;/s&gt;  in my HTK LM.  I think it is a bug of HTK tool, but It is a the only choice to train class-based LM. How do I fix it? Is it reasonable to remove directly these n-grams?

Thanks,

Huanliang Wang

</noautolink>

---+++ Main.HuanliangWang - 07 May 2013 - 23:10
<noautolink>
Hi, Brian

Thank you! I get it.

Another case is there are some n-grams that start with &lt;/s&gt; in my HTK LM. I think it is a bug of HTK tool, but it is my only choice to train class-based LM with automatic class clustering from large plain data . How do I fix it? Is it reasonable to remove directly these n-grams?

Thanks,

Huanliang Wang
</noautolink>

---+++ Main.BrianRoark - 08 May 2013 - 09:02
<noautolink>
yes, you might try just removing those n-grams.  In the ARPA format, you'll have to adjust the n-gram counts at the top of the file to match the number you have at each order.
</noautolink>

---+++ Main.HuanliangWang - 09 May 2013 - 05:06
<noautolink>
Hi, Brian

Thank you! I got it.

Could you give me an example how to use fstreplace to replace e nonterminal label in a fst by another fst? 

Thanks,

Huanliang Wang

</noautolink>

---+++ Main.BrianRoark - 09 May 2013 - 10:08
<noautolink>
You might try the forum over at openfst.org, that's an fst command.

brian
</noautolink>
%COMMENT{type="mythreadmode"}% 



---++ error when converting LM genereted by HTK into fst format 
---+++ Main.HuanliangWang - 07 May 2013 - 03:06
<noautolink>
Hi, 
   I try to convert arpa LM genereted by HTK tool into fst format. The command is : 
   ./tools/opengrm-ngram-1.0.3/bin/ngramread --ARPA test.arpa > test.lm.fst
   But I get a error:  
         FATAL: NGramInput: Have a backoff cost with no state ID!
   
   How to fix the error?

Thanks,
Huanliang wang
</noautolink>

%COMMENT{type="mythreadmode"}% 



---++ Infinity values / ill formed FST 
---+++ Main.RolandSchwarz - 03 Apr 2013 - 05:50
<noautolink>
Hi, 

I'm currently playing around with a test example and I noticed than after ngrammake if I call fstinfo (not ngraminfo) on the resulting language model fstinfo complains about the model being ill-formed. This is due to transitions (typically on epsilons) that have "Infinity" weight, which does not seem to be supported by openFST. Is that "working as intended"? The problem is later if I call fstshortestpath to get e.g. the n most likely sentences from the model the result contain not only "Infinity" weights but also "BadNumber" which might be a result of the infinite values.

Thanks,

Roland

</noautolink>


---+++ Main.BrianRoark - 03 Apr 2013 - 10:56
<noautolink>
Hi Roland,

yes, under certain circumstances, some states in the model end up with infinite backoff cost, i.e., zero probability of backoff.  In many cases this is, in fact, the correct weight to assign to backoff.  For example, with a very small vocabulary and many observations, you might have a bigram state that has observations for every symbol in the vocabulary, hence no probability mass should be given to backoff.  Still, this does cause some problems with OpenFst.  In the next version (due to be released in the next month or so) we will by default have a minimum backoff probability of some very small epsilon (i.e., very large negative log probability).  As a workaround in the meantime, I would suggest using fstprint to print the model to text, then use sed or perl or whatever to replace Infinity with some very large cost -- I think SRILM uses 99 in such cases, which would work fine. 

hope that helps,

brian
</noautolink>

---+++ Main.BrianRoark - 03 Apr 2013 - 10:58
<noautolink>
oh, I forgot to say that, once you have replaced Infinity with 99, use fstcompile to recompile the model into an FST.

brian
</noautolink>

---+++ Main.RolandSchwarz - 04 Apr 2013 - 05:12
<noautolink>
Hey, thanks a lot for the quick reply, I'll give that a go!

Roland
</noautolink>

---+++ Main.RolandSchwarz - 04 Apr 2013 - 05:32
<noautolink>
If I may add another quick question, when running fstshortestpath on the ngram count language model (i.e. after ngramcount but before ngrammake) I was expecting to get the most frequent n-gram, but instead the algorithm never seems to terminate. Any idea why that is? I though that shortestpath over the tropical sr should always terminate anyway.

Thanks,

Roland

</noautolink>

---+++ Main.BrianRoark - 04 Apr 2013 - 09:22
<noautolink>
The ngram count Fst contains arcs with negative log counts.  Since the counts can be greater than one, the negative log counts can be less than zero.  Hence the shortest path is an infinite string repeating the most frequent symbol.  Each symbol emission shortens the path, hence non-termination.

brian
</noautolink>

---+++ Main.RolandSchwarz - 05 Apr 2013 - 10:48
<noautolink>
Oh I should have thought of that myself, thanks for clarifying. I somehow assumed the counts model would be acyclic.
</noautolink>
%COMMENT{type="mythreadmode"}% 



---++ Build failure on Fedora 17 
---+++ Main.JerryJames - 18 Dec 2012 - 10:42
<noautolink>
Hi.  I maintain several voice-recognition-related packages, including openfst, for the Fedora Linux distribution.  I am working on an OpenGrm NGram package.  My first attempt at building version 1.0.3 (with GCC 4.7.2 and glibc 2.15) failed:

<verbatim>
In file included from ngramrandgen.cc:32:0:
./../include/ngram/ngram-randgen.h:55:48: error: there are no arguments to 'getpid' that depend on a template parameter, so a declaration of 'getpid' must be available [-fpermissive]
./../include/ngram/ngram-randgen.h:55:48: note: (if you use '-fpermissive', G++ will accept your code, but allowing the use of an undeclared name is deprecated)
ngramrandgen.cc:39:1: error: 'getpid' was not declared in this scope
ngramrandgen.cc:39:1: error: 'getpid' was not declared in this scope
</verbatim>

It appears that an explicit =#include <unistd.h>= is needed in ngram-randgen.h.  That header was probably pulled in through some other header in previous versions of either gcc or glibc.
</noautolink>


---+++ Main.BrianRoark - 19 Dec 2012 - 17:22
<noautolink>
ok, that header file will be included in the next version.  Thanks for the heads up.

brian
</noautolink>
%COMMENT{type="mythreadmode"}% 



---++ Expected result when using a lattice with ngram perplexity? 
---+++ Main.JosefNovak - 28 Nov 2012 - 06:40
<noautolink>
I was wondering what the expected result is when feeding a lattice, rather than a string/sentence, to the ngramperplexity utility?  Is this supported?  It seems to report the perplexity of an arbitrary path through the lattice.
</noautolink>


---+++ Main.BrianRoark - 28 Nov 2012 - 20:46
<noautolink>
Hi Josef,

ngramperplexity reports the perplexity of the path through the lattice that you get by taking the first arc out of each state that you reach. (Note that this is what you want for strings encoded as single-path automata.)  Not sure what the preferred functionality should be for general lattices.  Could make sense to show a warning or an error there; but at this point the onus is on the user to ensure that what is being scored is the same as what you get from farcompilestrings - unweighted, single-path automata.  If you have an idea of what preferred functionality would be for non-string lattices, email me.

brian
</noautolink>
%COMMENT{type="mythreadmode"}% 



---++ is there a way to use NGramApply in c++ 
---+++ Main.MarkusFreitag - 19 Oct 2012 - 12:55
<noautolink>
Hi,

I do not want to print my fst and execute NGramApply in bash before reading the new fst again in c++.

Is there a method to use the method NGramApply directly in c++ ?

Thanks
</noautolink>


---+++ Main.BrianRoark - 21 Oct 2012 - 21:32
<noautolink>
Hi Markus,

there is no single method; rather there are several ways to perform composition with the model, depending on how you want to interpret the backoff arcs.  The most straightforward way to do this in your own code is to look at src/bin/ngramapply.cc and use the composition method for the particular kind of backoff arc, e.g., ngram.FailLMCompose() when interpreting the backoff as a failure transition.  In other words, write your own ngramapply method based on inspection of the ngramapply code.

Hope that helps,

brian
</noautolink>

---+++ Main.MarkusFreitag - 22 Oct 2012 - 09:58
<noautolink>
Hi,

thanks, I think yes that should work.
I am using FailureArcs and my LM fst is created, so I do not need to build a lm fst out of strings or an ARPA lm.

I first just need to read the fst lm from my disk:

#include <ngram/ngram.h>

fst::StdMutableFst *fstforNGram;
fstforNGram->Read($MYNGRAMFST);
ngram::NGramModel ngram(fstforNGram);
// that seems not to work, as: undefined reference to `ngram::NGramModel::InitModel()'

If I read the lm , I could then just add:

ngram.FailLMCompose(*lattice, &cfst, kSpecialLabel);

and the composed fst should be ready, right?

Thanks for helping
</noautolink>

---+++ Main.BrianRoark - 23 Oct 2012 - 09:05
<noautolink>
Correct, that is the method for composing with failure arcs.
</noautolink>

---+++ Main.MarkusFreitag - 23 Oct 2012 - 09:30
<noautolink>
yes, but I have a problem to read the fst lm in c++:

fst::StdMutableFst *fstforNGram;

fstforNGram->Read($MYNGRAMFST);

to that point it works.

ngram::NGramModel ngram(fstforNGram);

 that seems not to work, as: undefined reference to `ngram::NGramModel::InitModel()'

Thanks
</noautolink>
%COMMENT{type="mythreadmode"}% 



---++ Fractional Kneser Ney 
---+++ Main.JosefNovak - 09 Oct 2012 - 04:31
<noautolink>
Hi,  I have been using OpenGrm with my Grapheme-to-Phoneme conversion tools for a while now and recently added some functionality to output weighted alignment lattices in .far format.  

It is my understanding that these weighted lattices can only currently be utilized with Witten-Bell smoothing; is this correct?  

Is there any plan to support fractional counts with Kneser-Ney smoothing, for instance along the lines of,

"Correlated Bigram LSA for Unsupervised Language
Model Adaptation", Tam and Schultz. 

or would I be best advised to implement this myself?  
</noautolink>


---+++ Main.BrianRoark - 09 Oct 2012 - 09:32
<noautolink>
Hi Josef,

Witten-Bell generalizes straightforwardly to fractional counts, as you point out.  No immediate plans for new versions of other smoothing methods along those lines, so if that's something that you need urgently, you would need to implement it. 

brian
</noautolink>

---+++ Main.JosefNovak - 09 Oct 2012 - 18:21
<noautolink>
Understood, and thanks very much for your response!
</noautolink>
%COMMENT{type="mythreadmode"}% 



---++  FATAL: NegLogDiff
---+++ Main.LukeCarmichael - 25 Sep 2012 - 17:21
<noautolink>
Hello,
I run this sequence of commands with the following output.

<verbatim>
home$ ngramcounts a.far > a.cnts
home$  ngrammake --v=4 --method=katz a.cnts > katz.mod
INFO: FstImpl::ReadHeader: source: a.cnts, fst_type: vector, arc_type: standard, version: 2, flags: 3
Count bin   Katz Discounts (1-grams/2-grams/3-grams)
Count = 1   -nan/0.253709/-0.343723
Count = 2   -nan/1.26571/1.19095
Count = 3   -nan/0.467797/-0.532465
Count = 4   -nan/1.29438/1.18879
Count = 5   -nan/0.0740557/-0.489831
Count > 5   1/1/1
FATAL: NegLogDiff: undefined -10.2649 -10.2651
</verbatim>
Other methods work fine.

How can I diagnose this problem?

Thanks,
Luke

</noautolink>


---+++ Main.BrianRoark - 26 Sep 2012 - 13:12
<noautolink>
Hi Luke, 

this is basically a floating point precision issue, the system is trying to subtract two approximately equal numbers (while calculating backoff weights).   The new version of the library coming out in a month or so has much improved floating point precision, which will help.  In the meantime, you can get this to work by modifying a constant value in src/include/ngram/ngram-model.h which will allow these two numbers to be judged to be approximately equal.  Look for: 
static const double kNormEps = 0.000001;
near the top of that file.  Change to 0.0001, then recompile.

This sort of problem usually comes up when you train a model with a relatively small vocabulary (like a phone or POS-tag model) and a relatively large corpus.  The n-gram counts end up not following Good-Turing assumptions about what the distribution should look like (hence the odd discount values).  In those cases, you're probably better off with Witten-Bell smoothing with the --witten_bell_k=15 or something like that.  Or even trying an unsmoothed model.

And stay tuned for the next release, which deals more gracefully with some of these small vocabulary scenarios.

Brian
</noautolink>
%COMMENT{type="mythreadmode"}% 



---++ FATAL: NGramModel: bad ngram model topology 
---+++ Main.BenoitFavre - 10 Sep 2012 - 09:18
<noautolink>
I generated an ngram model from a .arpa file with the following command:

ngramread --ARPA lm.arpa > lm.model

ngramread does not complain, but ngraminfo and trying to load the model from C++ code generate the following error:

FATAL: NGramModel: bad ngram model topology

How can I troubleshoot the problem?
</noautolink>


---+++ Main.BenoitFavre - 10 Sep 2012 - 09:27
<noautolink>
Adding verbosity results in more mystery...

ngraminfo --v=2 lm.model 
INFO: FstImpl::ReadHeader: source: lm.model, fst_type: vector, arc_type: standard, version: 2, flags: 3
INFO: Incomplete # of ascending n-grams: 1377525
FATAL: NGramModel: bad ngram model topology
</noautolink>

---+++ Main.BrianRoark - 11 Sep 2012 - 10:33
<noautolink>
Hi,

that error is coming from a sanity check that verifies that every state in the language model (other than the start and unigram states) is reached by exactly one 'ascending' arc, that goes from a lower order to a higher order state.  ARPA format models can diverge from this, by, for example, having 'holes' (e.g., bigrams pruned but trigrams with that bigram as a suffix retained).  But ngramread should plug all of those.  maybe duplication?  I'll email you about this.
</noautolink>

---+++ Main.BrianRoark - 18 Sep 2012 - 10:50
<noautolink>
Benoit found a case where certain 'holes' from a pruned ARPA model were not being filled appropriately in the conversion.  The sanity check routines on loading the model ensured that this anomaly was caught (causing the errors he mentioned), and we were able to find the cases where this was occurring and update the code.  The updated conversion functions will be in the forthcoming version update release of the library, within the next month or two.  In the meantime, if anyone encounters this problem, let me know and I can provide a workaround.
</noautolink>
%COMMENT{type="mythreadmode"}% 


Access control:
   * Set ALLOWTOPICCHANGE = Main.AllAuthUsersGroup


-- Main.CyrilAllauzen - 09 Aug 2012
Topic revision: r66 - 2014-04-03 - BrianRoark
Forum