How Bad is Internet Archive OCR?

We all know that the OCR on our digital resources is pretty bad. I've often wondered if part of the reason Google doesn't share its OCR is simply it would show so much ugliness. (A common misreading, 'tlie' for 'the', gets about 4.6m results in Google books). So how bad is the the internet archive OCR, which I'm using? I've started rebuilding my database, and I put in a few checks to get a better idea. Allen already asked some questions in the comments about this, so I thought I'd dump it on to the internet, since there doesn't seem to be that much out there.

First: here's a chart of the percentage of "words" that lie outside my list of the top 200,000 or so words. (See an earlier post for the method). The recognized words hover at about 91-93 percent for the period. (That it's lowest in the middle is pretty good evidence the gap isn't a product of words entering or leaving the language).

Now, that has flaws in both directions. Here are some considerations that would tend to push the OCR error rate on a word basis lower than 8%:

• There are many real words and proper nouns that are under my 200,000 cutoff (although I suspect the actual count is not that money, and could probably prove this using Zipf's law.)
• Some books may be in entirely or partly foreign languages, and thus full of non-words.
• A lot of the error words may be fragments where a word was split up into two chunks, but only one word was missed.
• My Perl preprocessing script could be making punctuation related mistakes that create false words that aren't OCR's fault. (For instance, I run together hyphenated words).

The second is the big one there, I think.

Things might make the error rate higher than 8%:

• Some of the error words may be concatenations of two words. (This isn't as likely, I don't think, as the reverse).
• Misreadings might confuse two real words--'them' for 'then', for example.
• I'm already counting a lot of typos as words in my top 200,000. Just eyeballing some randomly selected lists, I'd guess as many as half of my 'words' are actually typos.

I'm sure there are others in both directions.

I would think the biggest outstanding reason is all the misreadings that are so common my perl script takes them to be words. Skim this list, if you like, of a random selection of the words I'm using:


> words[sample(1:nrow(words),30),]
                word count          stem
169825       thiiigs   138        thiiig <- (misreading of 'things')
4543        alcaydes   326        alcayd
135174          poww   379          poww
12068    auricularis   167    auriculari
113516         moyed   730           moi
63398         florin  1756        florin
150309        sawney   381        sawnei
120030           nye  2113           nye
128509     peajacket   458     peajacket
152493    seleucidae   237     seleucida
114066        mulder   575        mulder
127922    pastrycook   514    pastrycook
113269          moud   174          moud
134             aaoh   238          aaoh <- (I can't tell from texts what this is)
84436       imitable   324          imit
90520            ixi  3748           ixi
148794        sacris   826         sacri
53102          emere   272          emer
49575          dulls   690          dull
136241  prevaricated   207        prevar
27222         caviar   419        caviar
152210          seei   484          seei <- (misreading of 'seen' or 'seek'?)
40027            cvv   169           cvv
66494     frustrates   319      frustrat
98422        leggere   283        legger
34009  complaisantly   138 complaisantli
190102        wliieh  3762        wliieh <- (misreading of 'which')
8998         apriori   382       apriori
142431   recruitment   358       recruit
163656     subsultus   255      subsultu


Some of those are foreign words, but a lot are obviously typos. It would be _great_ for my database if I could just drop out all the typos, so I'd love advice on how to do that--I've thought about just running some clustering routines on frequent typos to see what words appear disproportionately often in books with bad OCR, but that might be risky.

So what percentage of words are typos? That's really hard to say, but let me give some examples. Perl regexes inside R let us look for some of the common misreadings I've noticed: we can look for every word that has some characters characters commonly misread as each part of 'the'; 'f' for 't', 'li' for 'h', and so forth.

(t|l|i|f)(h|b|li|ii|n|ji|ti|t|l|m|fi)(e|c):



> words[grep("^[ltif](b|li|ii|n|ji|ti|t|l|m|fi|h)[ce]$",words[,1],perl=T),c(1,2)]->local; local[order(local[2],decreasing=T),]
       word     count
169158  the 197439894
172443 tlie    738290
167183  tbe    525324
82755   ihe    218396
171326 tiie    162834
173001  tne     67567
81669   ibe     65536
172422 tlic     58502
101354 llie     51008
82737   ihc     33725
...



Putting every word that looks like "the" together, I get about 2.2 million words. Not all of those as actually misreadings of 'the' (some are probably weird words, some are misreadings of "this", "line", or whatever), but that's probably a good ballpark guess. There are about 196 million occurrences of "the" in the flesh in the database. So that's an error rate of about 1% on the (somewhat) obvious misreadings of a short word. I think 'h' is one of the harder characters to read, based on the errors I've looked at, so that would imply IA OCR is well above 99% accurate on a character basis. (One study on the IA site found 99.6-99.9% accuracy on more modern texts, so that doesn't seem unreasonable for the whole corpus).

Another way we might find errors is by looking for lone letters. "A" and "I" excepted, these make up about 1% of the words in my sample; two letters words less common than 'ye', which seems to be about the cutoff between real words and errors, make up about 0.8%. (And both those counts include a number of legitimate initials, roman numerals, etc.) Since each individual detached letter or two letter chunk probably indicates only a fraction of a missed word, this isn't terrible.

So what's the conclusion? I suspect that over 85% of words overall are correct in the IA scans, and possibly as much as 90%. The error rates are high enough, though, that using this for sentence processing is problematic--word counts should work all right, though we should be aware that the incidence of longer words is probably slightly underrepresented. Still, I don't see much that indicates using words is going to be dramatically flawed except in a few cases we should watch out for, such as when a word might be a misreading for one a hundred times more common.


And finally, these are dragged down by a few bad apples. Here's a usage example of 'ifie', which I suspected was a typo for 'the', from an 1847 edition of Scott's Waverley Novels:
"i tve flyron a lciimifiil dagger mounted with coldwbicu lifui bccn ttic ptopirty of ifie rcdouhtix klfi ucy lbt i wiie to play tlic part if dionied in the jliajj cur byron sent mt so mi"

The median percentage of words in my sample is 94.2%, which is a little higher than the mean of 93.1% see at the top. The best book is 98.2% accurate.

That's all I have to say about that.