User:TJones (WMF)/Notes/Analysis Analysis Tools

From MediaWiki.org
Jump to navigation Jump to search

July 2017 — See TJones_(WMF)/Notes for other projects. See also T171516. For help with the technical jargon used in the Analysis Chain Analysis, check out the Language Analysis section of the Search Glossary.

Below is the README file that goes with my language analyzer analysis tools. It's been converted from markdown to HTML to wikitext, so there may be still be some formatting problems. The official version will be/has been committed to RelForge, but this is here for easier reading and comments by anyone who is interested. This version will not be updated as the live README file changes.

Trey's Language Analyzer Analysis Tools[edit]

July 2017

These are the tools I use to do analysis of Elasticsearch language analyzers and custom analysis chains. Most of my analysis write ups are available on MediaWiki.org. The older ones, naturally, used less complex versions of this code—I update it whenever something weird happens!

What's Here?[edit]

Let's see what we've got:

  • analyze_counts.pl: This is a simple program to run sample text against a language analyzer. It asks Elasticsearch to analyze the text, maps tokens in the output back to strings in the input, and keeps count of how often each mapping occurs.
  • compare_counts.pl: This is a much more complex program that does the real analysis of the language analyzer. It can perform either a "self analysis" on one counts file, or a "comparison analysis" between two counts files. See more below.
  • samples/data/: This directory contains 1MB samples of text from articles from various Wikipedias. The corpora have had markup removed, and lines have been deduped.
  • samples/output/: This directory contains output from the English examples below, so you can check them out without having to run everything yourself.

Running analyze_counts.pl[edit]

This is a pretty straightforward program to run:

./analyze_counts.pl input_file.txt > output_file.txt 
  • The input file is just a UTF-8–encoded text file with the text to be analyzed.
    • It is not strictly necessary, but it seems helpful to remove markup unless you are testing your analyzer's ability to handle markup.
    • I like to deduplicate lines to decrease the apparent importance of domain-specific patterns when I'm looking for general language behavior. For example, in Wikipedia, exact paragraphs don't repeat often, but headings like "See Also" and "References" certainly do. Deduping helps keep the counts for "see", "also", and "references" at more typical levels for general text.
    • It is more efficient for analyze_counts.pl to batch up lines of text and send them to Elasticsearch together. Up to 100 lines can be grouped together, up to around 30,000 characters (over 50K seems to cause problems). If your input file has lots of individual lines with significantly more than 10K characters per line, you could have trouble.
  • The output file, which I call a "counts file", is pretty self-explanatory, but very long, so note that there are two sections: original tokens mapped to final tokens and final tokens mapped to original tokens.
    • The output is optimized for human readability, so there's lots of extra whitespace.
    • Obviously, if you had another source of pre- and post-analysis tokens, you could readily reformat them into the format output by analyze_counts.pl and then use compare_counts.pl to analyze them.
  • While the program is running, dots and numbers are output to STDERR as a progress indicator. Each dot represents 1000 lines of input and the numbers are running totals of lines of input processed. On the 1MB sample files, this isn't really necessary, but when processing bigger corpora, I like it.
  • The program does some hocus-pocus with 32-bit CJK characters that use high and low surrogates, because these have caused problems with character counts with the tokenizer I was using. I haven't run into any difficulties, but that bit of code has not been severely stress-tested.

Running compare_counts.pl[edit]

Here's the usage for reference:

usage: $0 -n <newfile> [-o <oldfile>] [-d <dir>] [-l <lang,lang,...>] 
[-x] [-1] [-h] [-f] [-s <#>] [-t <#>] 

-n <file>  "new" counts file 
-d <dir>   specify the dir for language data config; default: compare_counts/langdata/ 
-l <lang>  specify one or more language configs to load. 
See compare_counts/langdata/example.txt for config details 
-h         generate HTML output instead of text output 

Analyzer Self Analysis (new file only) 
-x         explore: automated prefix and suffix detection 
-1         give singleton output, showing one-member stemming groups, too 

Analyzer Comparison Analysis (old vs new file) 
-o <file>  "old" counts file, baseline to compare "new" file against 
-f         apply basic folding when comparing old and new analyzer output 
-s <#>     minimum stem length for recognizing regular prefix and suffix alternations 
-t <#>     terse output: 
1+ = skip new/old lists with the same words, but different counts 
2+ = skip new/old lists with the same words, after folding 

If you specify two counts files (-n & -o), then compare_counts.pl will run a comparison analysis. If you only specify one counts file (-n), then you get a self analysis.

A Few Useful Notes[edit]

On Types and Tokens[edit]

A lot of the analysis looks at types and tokens, and it’s important to distinguish the two. Tokens refer to individual words, counted each time they appear. Types count all instances of a word as one thing. It's also useful to keep in mind the distinction between pre-analysis and post-analysis types. Pre-analysis, we have the words (types) as they were split up by the tokenizer. Post-analysis, words that have also been regularized (by stemming, lowercasing, folding, etc.) are counted together as one type.

So, in the sentence The brown dog jumped, the black dog jumps, and the grey dogs are jumping. there are fourteen tokens (more or less what you would probably think of as normal words).

Pre-analysis, there are twelve types: The, brown, dog, jumped, the, black, jumps, and, grey, dogs, are, jumping—note that The and the are still distinct at this point, too!

With a plausible English analyzer, there might only be eight post-analysis types: the, brown, dog, jump, black, and, grey, be. A different English analyzer, that drops stop words, might give five post-analysis types: brown, dog, jump, black, grey.

RTL and You[edit]

When you analyze a big chunk of data from almost any Wikipedia, you get lots of words in other languages. Some of those languages are RTL (right-to-left)—with Arabic and Hebrew being the most common. Depending on how your operating system, editor, and browser are configured, you can get some ugly results.

If you aren't familiar with RTL languages, when they are displayed you can get flipped arrows, brackets, and braces, and sometimes in extra confusing ways like ]ℵℵℵ],* where ℵℵℵ is anything in Hebrew or Arabic. Reading right to left, that's an RTL opening bracket, ℵℵℵ, and an LTR closing bracket—because sometimes computers are stupid that way.

The HTML output is sometimes a little better, especially when it's possible to sneak in a helpful dir=auto, but not always. In MediaWiki, you can wrap the text in <div dir="rtl"> ... </div> for big swathes of mostly RTL output.

* To make sure this displays correctly, I had to cheat and use an LTR version of Hebrew alef, which is used in mathematics.

Comparison Analysis[edit]

The main purpose of a comparison analysis is to look at the impact of relatively localized changes to an analysis chain, including small to moderate changes to character filters, tokenizers, and token filters. For example, it's been used to look at the effects of turning on or off the word_break_helper and enabling ASCII-folding and ICU-folding.

Report Details[edit]

Among the details reported are:

  • The number of tokens found by the tokenizers. If this changes, then the tokenizer is breaking up your input text differently.
  • "Pre" and "Post" types and tokens. "Pre" and "Post" refer to pre-analysis and post-analysis. The language analyzer drops some tokens (usually stop words, affecting type and token counts) and munges others (affecting type counts). So, if the tokenizer finds dog and dogs, we have two types. After analysis, though, they could both become dog, so now we only have one type.
    • We also try to categorize the number of types and tokens that are unchanged by the analysis. Three big categories (which overlap) are unchanged (token in, token out: dog stays dog), lc_unchanged (only difference is case, so DoGgO becomes doggo), and number (strings of digits—not including commas or periods—tend to remain unchanged).
  • Collisions and Splits: The original purpose of the distant progenitor of compare_counts.pl was to look for new collisions, that is, words that are grouped together that weren't before. For example, if we enable ASCII-folding, we would expect resumé and resume to both be indexed as resume, whereas before they would have been distinct. Splits are the opposite—words that used to be grouped together than no longer are.
    • We can currently only reliably detect collisions and splits that result from small to moderate changes to the analysis change; in particular, if the new analyzer significantly changes most post-analysis forms, the comparison analysis won't be able to reliably detect collisions and splits. Thus, it doesn't make sense to look for collisions and splits when you go from the default language analyzer (which doesn't know anything about any language or its morphology and grammar, just tokens and characters) to a language-specific analyzer that does know about the morphology of the language. It also doesn't work well when the tokenizer changes dramatically, such as switching from the CJK tokenizer (which breaks up CJK characters into overlapping bigrams) to a language-specific tokenizer for Chinese or Japanese (which tries to find actual word boundaries). It can still be useful to run the Comparison Analysis in these cases, but the collisions and splits are often numerous and useless.
    • Collisions and splits are reported as the number of types and tokens "involved". So, if a type with one token (i.e., a word that appeared exactly once) gets newly indexed with 9 other one-token types, we have 10 types and 10 tokens in a collision. If a 10-token type merges with a 100-token type, that's 2 types and 110 tokens. This doesn't tell you whether changes are good or not, but does indicate the relative impact of the changes.
  • Near Match Stats are reported as type and token counts for new collisions. The categories are folded (new type matches at an old type after default and language-specific folding, de-spacing, and de-hyphenating the new type), regulars (new type matches an old type after taking into account regular morphology, which is just removing regular prefixes and suffixes), and folded_regulars (new type matches an old type after both folding and regular morphology). These collectively are the "obvious" collisions. We aren't too surprised when resume, resumé, resumes, and resumed collide.
  • Lost and Found Tokens: One sneaky fact is that the normalization of a token can change without causing a split or collision because it doesn't group with any other types. If the Foo123bar was normalized as foo123bar before and foobar afterwards, we might never know because it was always on its own. So, we report lost and found tokens. "Lost" tokens are those that were present in the old analysis but are not in the new analysis, and "found" tokens are the other way around.
    • Lost and found tokens are listed both pre-analysis—where changes are usually due to the tokenizer—and post-analysis—where additional changes are usually due to stemming or folding.
    • Lost and found tokens are grouped into "categories", which are based on character set, the formatting of numbers, and some easily identified other types, like measurements, acronyms, etc. other is a catch-all for everything else; I sometimes mine the other category for new categories. Comparing lost categories to found categories sometimes makes it obvious what's changed.
      • Modifiers are added to the groups for certain invisible characters, like zero-width non-joiners (zwnj), zero-width joiners (zwj), zero-width spaces (zwsp), non-breaking spaces (nbsp), or bi-directional markers (bidi).
      • If you have a really big lost/found change, less -S can be your friend. All the changes in one category end up on one unwrapped line. This happens in particular when you enable a language analyzer and there wasn't one before. Many, many words are now stemmed, and their original unstemmed forms are "lost".
      • Tokens in each category that occur more than 500 times are re-listed with their counts on a separate line labeled "hi-freq tokens".
  • After the lost and found tokens, changed collisions are listed, divided up into new collisions (Net Gains) and new splits (Net Losses).
  • "Unexpected" collisions or splits are listed under Bad Collisions? and Bad Splits? What is expected or unexpected depends largely on the language and folding options selected.

Folding & Languages[edit]

For a comparison analysis, enabling folding (-f) applies the available folding to tokens for computing Near Match Stats and detecting potential bad collisions. Default folding (./compare_counts/langdata/default.txt) is always available, and additional folding can be specified in a language-specific config (e.g., ./compare_counts/langdata/russian.txt).

So, with folding enabled, léo being merged with Leo is no longer a potential bad collision because the difference is gone after default folding. Note that the folding specified in the default config is intentionally incomplete as all of it has been manually added. If ICU folding or some other change causes an unexpected collision I don't want to miss it because I was using the same or a similar library for folding here.

Additional language configs can be supplied. Use -d to specify the directory and -l to specify a comma-separated list, though usually only one language is needed. (NB: "default" is always loaded.) To specify that compare_counts/langdata/russian.txt should be loaded, use -d compare_counts/langdata -l russian. All language config files should end in .txt. Note that -d compare_counts/langdata is the default, so if you don't move that directory, you should only need -l.

Additional folding, such as Cyrillic Ё/ё to Е/е, or stress accent to nothing can be specified in a language config file.

Regular morphology used in a comparison analysis is specified in regular_suffixes and regular_prefixes. (See details under self analysis for info on the similar strip_prefixes and strip_suffixes.) Regular affixes are specified in groups that can alternate with each other, and can include the empty string (for affixes that can be removed). For a simplified version of English, we might specify "ing", "ed", and "" (the empty string) as alternates. So, if starting gets added to a group that already contains started or start, then the collision is not considered as a potentially bad collision, because it is explained by the regular morphology. Such collisions are not always correct—for example, work/worker would be a good collision, while heath/heather is not—but they are not unexpected. Only one affix alternation (i.e., one suffix or one prefix) is allowed per match.

While a lot of morphology in many languages is more complex than one simple suffix or prefix, they are much easier to handle than infixes, circumfixes, transfixes, strong inflections, polysynthesis, noun incorporation, and all the other really cool stuff languages do.

See compare_counts/langdata/example.txt for examples of how to format a language config file.

As noted, it is possible to specify multiple language configs in a comma-separated list, but it probably doesn't make sense to do so if the configs are actually for different languages. I've used it for conservative vs aggressive or computationally cheap vs expensive morphology and folding in the same language.

Stem Length[edit]

The minimum stem length (-s, default value: 5) is the minimum length of the string left after removing a prefix or suffix. So, even if -able and -ation were allowed to alternate, stable and station would count as a potentially bad collision (i.e., one that a human should review) because the apparent stem, st, is too short.

Terseness[edit]

Increasing the terseness further limits the groups shown under "Changed Collisions".

With terseness set to 1, groups with the same types, but different token counts are not shown. Changes in counts are generally caused by changes in tokenization (the most common being word breaking on underscores, hyphens, periods, and slashes), as no new tokens are being grouped together, but more or fewer instances of the tokens are being found. With terseness set to 1, the group below would not be shown as changed:

    xyz >> 
          o: [3 XYZ][1 Xyz][33 xyz] 
          n: [4 XYZ][1 Xyz][44 xyz] 

With terseness set to 2, groups with the same types after default + language-specific folding, and possibly with different token counts, are not shown. These are the "expected" changes attributable to commonly-seen folding. With terseness set to 2, the group below would not be shown as changed:

    xyz >> 
          o: [3 XYZ][1 Xyz][33 xyz] 
          n: [3 XYZ][1 Xyz][33 xyz][1 xÿž] 

HTML Output[edit]

HTML output (-h) for the comparison analysis doesn't add anything else other than a table of contents, but it's there if you want it.

Things to Investigate[edit]

  • Look for unexpectedly large changes in the number of tokens found (total and pre-/post-analysis). Some changes—particularly due to changes in whether stop words are being dropped, and changes in whether underscores, hyphens, periods, and slashes are word-breakers—might be expected if you've changed analyzers or tokenizers. Folding changes, though, normally shouldn't change the number of total tokens. (Though word breaking on stress accents, for example, might!)
  • New Collision/Split Stats should reflect a sensible impact. Enabling ASCII-folding in English should have minimal impact, since most English words don't have diacritics. For Vietnamese, the impact of diacritic folding could be huge! Going from the default analyzer to a language-specific analyzer could have a giant impact in a highly inflected language, since the stemmed form of most words is different from the original form.
  • There's often a lot going on in the Lost and Found tokens:
    • I like to run a sample of the lost tokens from each category through the analyzer directly on the command line to see what's happening to them to explain why they are lost. With folding, there are a handful of tokens in several different categories, because there are non-canonical forms of letters in several scripts that can get "regularized".
      • e.g., curl -s localhost:9200/wiki_content/_analyze?pretty -d '{"analyzer": "text", "text" : "xÿz" }'
    • Stop words can be lost if you've enabled a language-specific analyzer where there was none before.
    • IPA—which is common in Wikipedias and Wiktionaries—tends to get murderized by lots of analyzers, which often split on the more esoteric characters.
    • Rarely, something really drastic happens, like all tokens in "foreign" scripts are lost. That is double-plus-ungood.
  • Changed Collisions, Net Losses/Gains, Bad Splits/Collisions?—I like to skim through these and then re-run with folding enabled. Review of all of these or a sizable random sample of these by a fluent speaker is often helpful.
    • Even if you want to review all the changes, enabling folding and looking at the residue can highlight different kinds of changes from the basic folded changes.
  • I think it's good to take into account the "main" language you are working in when reviewing collisions. For example, English pretty much ignores accents. There are a few examples, like resumé, fiancée, and Zöe, where accents are relatively common or even important, but not many. So, when working on data from English Wikipedia, I don't really care that German words or Russian words are getting folded in a way that a German or Russian speaker wouldn't like. Lots of English speakers can't type the ö in German quadratförmig, so quadratformig should match—on English-language projects.
    • In general I expect and accept folding of characters that aren't part of the main language of the corpus being analyzed. In English—fold them all! In Swedish, folding å, ä, and ö is suspect—it might be acceptable, but probably not (it wasn't—see T160562)—but folding anything else is probably fine. On the other hand, the default German language analyzer folds common umlauts! I was shocked, and the jury is still out on whether it's good or not.

Self Analysis[edit]

The main purpose of a self analysis is to assess big changes to an analysis chain, such as implementing a language analyzer where there was none before.

Report Details[edit]

Among the details reported are:

  • Stemming Results: Words (types) that are analyzed the same are grouped together in a tab-separated table. The columns of the table are:
    • stem: the "stem" returned by the analyzer. Note that some analyzers are stemmers and some are lemmatizers. Lemmatizers return a "stem" that attempts to be the actual root/canonical form of the word being analyzed. Stemmers just return something reasonable, without worrying whether it is an actual word, or the root of the word being stemmed.
    • common: This shows the common beginning and common ending of all word types in the group. If all types in the group match a prefix or a suffix or both without modification (other than lowercasing), the middle bit is ".."; if a match is only found after affix-stripping (see below), the middle bit is "::"; if a match is only found after affix-stripping and default character folding (see below), the middle bit is "--". If no match is found, this column will have " -- " in it.
      • The terms prefix and suffix are used non-morphologically to refer to beginning and ending substrings, but I'm going to use beginning and ending because I'm already using prefix and suffix with their linguistic/grammar sense.
    • group total: This is the total number of tokens (instances of a word, more or less) in this group.
    • distinct types: This is the total number of pre-analysis types (distinct words, more or less) in this group.
    • types (counts): This is a list of the types and their frequencies in the group. The format is [<freq> <type>].
  • Case-Insensitive Type Group Counts: This is a small histogram of the number of case-folded pre-analysis types found in the groups in the Stemming Results. (This slightly blurs the lines between pre- and post-analysis types by case-folding—lowercasing—everything before counting types. Case rarely seems to be distinctive, while all the other kinds of character folding can be.) Depending on the size of the corpus and the grammar of the language, what to expect here can vary, but obvious outliers should be investigated. Some languages have many more forms of a word than others. Smaller corpora are less likely to have as many different forms of any particular word.
  • Stemmed Tokens Generated per Input Token: Most stemmers generate one stem per input token. Some, like HebMorph for Hebrew, deal with ambiguity by generating multiple possible stems. Very rarely, a token is stemmed to the empty string (giving a count of 0). Anything other than a count of 1 is probably unexpected.
  • Final Type Lengths: This is a histogram, with examples, of the length of tokens found by the analyzer. Up to ten examples are randomly selected from list of tokens of a given length.

Singletons[edit]

A lot of terms only show up once and have unique stems, particularly numbers, proper names, and foreign words (especially ones in a foreign script). By default these are left out of the Stemming Results, but you can include them by specifying -1 on the command line. I've only found this useful when something weird is going on with stemming and I want to look at more examples.

Folding & Languages[edit]

The -f option, which enables folding, doesn't do anything for a self analysis. The default folding config is loaded and used in the Stemming Results as a last-ditch attempt to find a common substring.

As with Comparison Analysis, additional language configs can be supplied. Use -d to specify the directory and -l to specify a comma-separated list, though usually only one language is needed. (NB:default is always loaded.) To specify that compare_counts/langdata/polish.txt should be loaded, use -d compare_counts/langdata -l polish. All language config files should end in .txt. Note that -d compare_counts/langdata is the default, so if you don't move that directory, you should only need -l.

The config for strip_prefixes and strip_suffixes is similar to the regular morphology used in a Comparison Analysis (q.v.) specified by regular_suffixes and regular_prefixes.§

strip_prefixes and strip_suffixes are strippable affixes that can be stripped off tokens when looking for common substrings (the common column of the Stemming Results, above) in a self analysis. Single letter affixes should probably not be included here because the possibility for ambiguity is too high.

The list of strip_prefixes and strip_suffixes may also be determined in part by the analyzer being investigated rather than regular morphology. The Stempel Polish analyzer, for example, strips the prefixes nie-, naj-, anty-, nad-, and bez- (roughly English "-less", "-est", "anti-", "super-", and "un-"). Once these were identified (see the Explore option, below), it was helpful to ignore them when looking for stemming groups with no common ending or beginning substring.

§ The primary reason that strip_*fixes and regular_*fixes are distinct is that I wasn't sure whether or not they should be the same as I was implementing them. strip_*fixes, as they should not include very short affixes, are probably often a subset of regular_*fixes. I should probably refactor and consolidate them in a future version.

Explore (Automatic Prefix and Suffix Detection)[edit]

The "explore" option (-x) enables automatic prefix and suffix detection. For boring languages like English, Polish, and other European languages that use lots of prefixes and suffixes (rather than more interesting forms of morphology, see note † above), this is a good way to find the most common prefix and suffix alternations, and to see if your language-specific analyzer is doing anything particularly interesting or unexpected. For each stemming group, two tests for alternations—group and one-by-one—are done for both prefixes and suffixes.

  • In the group analysis of suffixes, the common beginning substring of all members of a stemming group is removed, and all pairs among the leftovers are considered potential suffix alternations.
  • In the one-by-one analysis of suffixes, for each pair of members of a stemming group, their common beginning substring is removed, and the leftovers are considered potential suffix alternations.
  • For prefixes, the process is essentially the same, but the common ending substrings are removed in each case, and the leftovers are considered potential prefix alternations.

Sometimes there's not a lot of difference between the group and one-by-one analyses. English, for example, doesn't use a lot of prefixes and doesn't have a lot of diacritics, so these are often the same. However, for example, among resume, resumé, résumé, resumes, resumés, résumés, the group common beginning substring is only r and there is no common ending substring. The one-by-one analysis finds three instances of the obvious -/s suffix alternation, though—along with some other dodgy alternations.

For languages with alternating diacritics (like German) or prefixes and suffixes (like Polish), the one-by-one analysis finds lots of alternations the group analysis would miss. On the other hand, for languages like English, the group analysis is less subject to detecting random junk.

The results of the analysis are presented in two additional sections, Common Prefix Alternations and Common Suffix Alternations. The each contain a table with the following columns:

  • 1x1 count: This is the number of times this alternation was seen in the one-by-one analysis.
  • group count: This is the number of times this alternation was seen in the group analysis.
  • alt 1 / alt 2: The two alternating affixes. Order is not important, so they are presented in alphabetical order. It is possible for alt 1 to be empty, which means something alternates with the empty string (as with English dog/dogs, which demonstrates the -/s suffix alternation.)
  • known: If a language config file has been specified (using -l), then this column indicates whether a given alternation is known, i.e., is specified in the config under regular_suffixes or regular_prefixes, or not. Known alternations are marked with an asterisk (*).

The results are sorted by combined group and one-by-one count. Anything with a combined count below 4 is ignored.

There's likely to be plenty of junk in the low counts, including the remnants of regular processes. For example, in English, stem/stemming, hum/humming, and program/programming all show an apparent -/ming alternation, which is really just the "sometimes double the final consonant before adding -ing" rule.

HTML Output[edit]

You can also generate HTML output for a self analysis with the -h option. The content is the same but the HTML version has additional links, a table of contents, and bolding & colors to highlight potentially interesting bits. The text version is nice if you are grooving on the command line and don't want to switch to a browser, and it's especially handy if you want to grep for a particular small subset of lines from the report.

In the HTML table of contents, you'll see Potential Problem Stems as a sub-item under Stemming Results. This links to the first group in the table with no common beginning or ending substring. The stem and common " -- " will be red, and the stem itself will be a link to the next group with no common beginning or ending substring. The stem in the last group with no common beginning or ending substring links to the bottom of the Stemming Results table, so you know you are done.

Under Common Prefix/Suffix Alternations, rows for unknown alternations are colored red, so you can more easily pick out the ones that are not in you language config file, should you care to consider adding them. (If no language config with regular_suffixes or regular_prefixes has been specified, no red highlighting occurs, since then everything would be red and that's just annoying.)

In the Histogram of Case-Insensitive Type Group Counts table, count values above a certain threshold (see Hidden Params below) have a link to the first group with that number of case-folded pre-analysis tokens in it. On that row of the table, the distinct types value (which is the same number as the one from the histogram section) links to the next one. The last one links back to the Histogram of Case-Insensitive Type Group Counts table.

And as noted above, Arabic and Hebrew are handled a little better in the Stemming Results table because each table cell gets its own dir=auto.

Things to Investigate[edit]

There is a lot of information here. What to do with it all? The things I generally look for include:

  • Big stemming groups. Surprisingly large groups are more likely to have something strange in them. I always investigate the largest groups—definitely any outliers, and usually any count with a frequency less than 10 (e.g., if there were only 6 groups with 12 distinct word types in them, we should investigate all 6.)
    • Sometimes big groups are readily explained. There could be an ambiguous word that is, say, both a noun and a verb, and the stemmer deals with this by just lumping all the noun forms and all the verb forms together. It's not great, but it is understandable. Other times, things are just incorrectly grouped together.
    • I include the largest groups in my data for speaker review.
  • Stemming groups with no common beginning/ending substring. This isn't always a problem; for example it's reasonable to group English good/better/best together, or is/am/are/be/been/was/were—but it's best to take a look at such cases.
    • I usually include all the groups with no common substring in my data for speaker review, unless there's something about the language that makes this very common.
  • Random stemming group sample. It's better when the speakers come back and say that the largest groups and the groups with no common substring are fine, but even if they don't, it's not the end of the world. A random sample of groups with >1 member (i.e., excluding singletons) will be most representative of the accuracy of the analyzer over all.
    • I include 50 or 100 random groups in my data for speaker review, depending on how many big groups and groups with no common substring there are.
  • Tokens with ≠1 stem. Unless the stemmer regularly spits out multiple stems, it's best to check into tokens that generate multiple stems, or no stem!
  • Really long tokens. Any corpus could include a long German compound noun, but it's good to check on the unexpectedly long words. Is there a tokenization problem? That's particularly likely for spaceless languages!
  • Token category by length. The examples under Final Type Lengths are randomly selected, and so give a very rough idea of the proportions of the categories found among tokens of a particular length. For a language that uses the Latin alphabet, for example, I expect tokens of length 1 to include a few numbers, a few Latin characters, and, for Wikipedia samples, a few random Unicode characters (which could have been included as individual letters, or have been separated out by not-so-great Unicode tokenization). Longer tokens tend to include website domains and other non-words (abcdefghijklmnopqrstuvwxyz pops up from time to time). If anything looks unexpected—like all example tokens of length 8 are in foreign alphabet—investigate!
  • Unexpectedly common prefix/suffix alternations. If you are using -x for automated prefix and suffix detection (see above), not all of the automatically detected prefix and suffix alternations are going to be valid. In English, if better and best are stemmed together, then we have a "suffix" alternation between -tter and -st, which is obviously a fluke in English. If the flukey-looking alternations are very high frequency, then you should track down the source and see what's going on.

Hidden Params[edit]

There are currently a few parameters that have been reified into variables (magic numbers bad!) but not exposed to the command line. I'm pretty sure the default values are reasonable, but every now and then I change them in the code. You might want to, too. I should expose them on the command line, and eventually I probably will.

  • $min_histogram_link = 10; This is the minimum count value in the Histogram of Case-Insensitive Type Group Counts that gets a link. 10 is usually great, except when nothing goes that high, then 5 is a good value.
  • $token_length_examples = 10; This is the number of examples in the Histogram of Final Type Lengths. Sometimes when there are only 15, and I want to see them all.
  • $min_alternation_freq = 4; This is the minimum combined group/one-by-one count in the Common Prefix/Suffix Alternations. Alternations that occur less frequently are not displayed. Sometimes you really want to lower it to zero and see eatre/éâtre,ebec/ébec and friends.

Analysis Example: English[edit]

Below I will walk through sample analyses of the English corpus found in the samples/data/ directory.

My Setup[edit]

Your setup doesn't have to be exactly the same as mine, but this works for me without a lot of customization beyond re-configuring mediawiki for the analyzer I'm working with. You could set up specific analyzers, etc., directly within Elasticsearch, but I do all my work in the context of MediaWiki and the various wiki projects, so I just work in vagrant. You will need Perl and Elasticsearch to run analyze_counts.pl, and just Perl to run compare_counts.pl.

I have a reasonably up-to-date install of MediaWiki vagrant with the cirrussearch role enabled. At the time of this writing, vagrant comes with the plugins analysis-icu, experimental-highlighter, extra installed.

My copy of vagrant doesn't have a lot of documents indexed so I can quickly and easily re-config/re-index into another language (with updateSearchIndexConfig.php, see below).

Generate Count Files for English With and Without Folding[edit]

The primary purpose of this example is to do a comparison analysis between configs with and without character folding. However, we'll take a look at a self analysis on the folded config, too.

Since generating the counts files requires re-configuring Elasticsearch and/or MediaWiki, you can skip these first few steps and just jump down to Comparison Analysis of the Unfolded and Folded Count Files and use the provided English count files, samples/output/en.counts.unfolded.txt and samples/output/en.counts.folded.txt.

To generate the count files, analyze_counts.pl needs access to the English language analyzer; it assumes Elasticsearch is running on localhost at port 9200, with an index called wiki_content and a properly configured analyzer called text.

On the one hand, that's why using vagrant is easy—everything is all set up. On the other hand, all those assumptions are built into one curl command inside analyze_counts.pl, and can easily be changed:

curl -s localhost:9200/wiki_content/_analyze?pretty -d '{"analyzer": "text", "text" : "$escline" }'

Re-Configure MediaWiki/Elasticsearch for English Without Folding Enabled[edit]

  • in LocalSettings.php set $wgLanguageCode = "en";
  • in AnalysisConfigBuilder.php comment out the $filters[] = 'asciifolding'; line from function customize()
  • inside vagrant (i.e., do vagrant ssh first), run:
    • mwscript extensions/CirrusSearch/maintenance/updateSearchIndexConfig.php --reindexAndRemoveOk --indexIdentifier now
    • for a faster re-index on really small document sets, change MONITOR_SLEEP_SECONDS in Reindexer.php to 1 first.

Note: If you have the analysis-icu plugin installed, the lowercase filter will be replaced with icu_normalizer filter. This is the setup I'm running with.

You can check on your config after running updateSearchIndexConfig.php above and going to your Cirrus Settings Dump. Search the page for "text" (with quotes) to see the current analyzer config.

Generate the "Before"/Unfolded Counts File[edit]

  • inside vagrant, run: ./analyze_counts.pl samples/data/en.txt > samples/output/en.counts.unfolded.txt
    • STDERR output: ..... 5385

samples/output/en.counts.unfolded.txt now contains the counts for normalized tokens found by the English analyzer without ASCII/ICU folding. It maps both original tokens (found in the text) to final tokens (those that would be indexed).

Re-Configure MediaWiki/Elasticsearch for English With Folding Enabled[edit]

Some of the steps and info are the same as above, but are repeated here for completeness.

  • in LocalSettings.php, you should still have $wgLanguageCode = "en";
  • in AnalysisConfigBuilder.php un-comment out the $filters[] = 'asciifolding'; line from function customize()
  • inside vagrant (i.e., do vagrant ssh first), run:
    • mwscript extensions/CirrusSearch/maintenance/updateSearchIndexConfig.php --reindexAndRemoveOk --indexIdentifier now
    • for a faster re-index on really small document sets, change MONITOR_SLEEP_SECONDS in Reindexer.php to 1 first.

Note: If you have the analysis-icu plugin installed, the lowercase filter will be replaced with icu_normalizer and the asciifolding filter will be replaced with icu_folding. This is the setup I'm running with.

You can check on your config after running updateSearchIndexConfig.php above and going to your Cirrus Settings Dump. Search the page for "text" (with quotes) to see the current analyzer config.

Generate the "After"/Folded Counts File[edit]

  • inside vagrant, run: ./analyze_counts.pl samples/data/pl.txt > samples/output/en.counts.folded.txt
    • STDERR output: ..... 5385

The output to STDERR is the same—it's just a progress indicator on the number of lines in the input file.

samples/output/en.counts.folded.txt now contains the counts for normalized tokens with ASCII/ICU folding.

Comparison Analysis of the Unfolded and Folded Count Files[edit]

(NB: I have icu-analysis installed, so I'm using icu_normalizer/icu_folding rather than lowercase/asciifolding, so if you regenerated the counts files without them, your results will vary.)

Baseline Analysis[edit]

The simplest analysis is straightforward, with the unfolded counts file as the "old" file and the folded counts file as the "new" file:

  • ./compare_counts.pl -o samples/output/en.counts.unfolded.txt -n samples/output/en.counts.folded.txt > samples/output/en.comp.unfolded_vs_folded.txt

Open samples/output/en.comp.unfolded_vs_folded.txt in your favorite text editor or just use less on the command line, and let's see what we see:

  • The Total old/new tokens are the same, which is what we'd expect. The additional folding here takes place after tokenization and shouldn't affect pre-analysis token count.
  • Under Pre/Post Type & Token Stats the new file's post-analysis types have decreased. As we expect, some words are being folded together.
  • New Collision Stats show no new splits, and only a very few new collisions—less than 1%. The actual impact of folding somewhat is larger; some tokens are singletons both befor and after folding.
  • There are no Lost/Found pre-analysis tokens because we didn't change the pre-analysis processing, specifically tokenizing.
  • Looking at the Lost post-analysis tokens by category we see:
    • Under other we have ˈzbraxlin. While that looks like an apostrophe, it's really an IPA stress mark. One of my favorite things about English Wikipedia is that you can search for almost any character and land on a page that tells you what that character is. Since I wouldn't expect an apostrophe to result in something being in the other category, searching Wikipedia for the character tells us what it is.
    • For several of the non-Latin scripts, you can compare the lost tokens to the found tokens and see minor variations that are the result of folding. (Pssst: I suggest zooming in a lot in your browser or editor when looking at scripts you aren't familiar with. It makes the details pop!)
      • Arabic أبو loses the hamza to become ابو after folding.
      • Cyrillic барсукоў loses the breve to become барсукоу.
      • Hebrew יִשְׂרְאֵלִית loses the niqqud to become ישראליות. Note that not all lost tokens have counterparts in the found tokens. Unfolded לִירָה becomes folded לירה, but since לירה is already in the corpus elsewhere, it is not "found". This unfolded/folded pair will show up again later under collisions.
      • Devanagari conjuncts are expanded (e.g., भट्टराई becomes भटटराई).
      • Kannada ashkara are converted from diacritics to full forms (e.g., ಸಂಕೇಶ್ವರ becomes ಸಂಕೇಶವರ).
      • Katakana dakuten are removed (e.g., テレ becomes テレ).
      • IPA doesn't show up in the found tokens, because IPA symbols are either dropped if they are diacritic-like or converted to plain Latin letters if they are letter-like. e.g., ˈeːijaˌfjœrðʏr̥ becomes eijafjoerdyr, which is in the found tokens under Basic Latin.
      • There aren't any "Latin (Basic)" lost tokens, and there aren't any "Latin (Extended)" found tokens because the Extended Latin characters were all converted to Basic Latin characters by folding.
  • A quick skim of the Changed Collisions and Bad Collisions? doesn't show anything untoward. Notice that there are two Hebrew items at the end of both lists. And if you aren't RTL-enabled, they are probably formatted poorly.

Simple Folded Analysis[edit]

Let's try it again, but this time enable folding (adding -f):

  • ./compare_counts.pl -o samples/output/en.counts.unfolded.txt -n samples/output/en.counts.folded.txt -f > samples/output/en.comp.unfolded_vs_folded._f.txt

Taking a look, the main difference is in the Bad Collisions? section. There are now only 11 collisions instead of 71. The other 60 were the "expected" ones, based on the folding rules specified in compare_counts/langdata/default.txt. (The New Collision Near Match Stats also shows 60 folded collisions now.)

Two of the eleven are the Hebrew words with niqqud that we saw earlier. Seven of the remaining nine, slightly reformatted, are:

d     [1 ḍ]      ->  [48 D][16 d] 
g     [1 ɢ]      ->  [23 G][35 g] 
ku    [3 ḵu]     ->  [1 ku] 
kun   [1 ḵun]    ->  [1 kun] 
t     [1 ṭ]      ->  [29 T][32 t] 
re    [1 ʁɛ̃]     ->  [3 Re][26 re] 
vila  [1 ˈvilɐ]  ->  [3 Vila] 

Those are all quote reasonable. It seems that compare_counts/langdata/default.txt doesn't include certain combining diacritics and variants with rare diacritics. Nothing much to see here.

The remaining two, though, show something interesting:

gab   [1 gábor]  ->  [1 Gabes][1 gabions] 
worn  [2 wörner] ->  [6 worn] 

If you look in AnalysisConfigBuilder.php you'll see that asciifolding is configured to happen right before kstem, which does the actual stemming for English. With diacritics, gábor and wörner are untouched. Without the diacritics, their final -er and -or are seen as suffixes (like work-er and direct-or) and stripped off.

Unless you try to maintain a complete lexicon of every word in a language—which is impossible!—this kind of thing is going to happen, because heuristics are never perfect. (If they were, they'd be rules, and English has never met a spelling "rule" it couldn't break.) So, while this isn't particularly good, it is understandable, and more or less expected.

Whether or not ASCII folding should happen before or after stemming in English is something I looked at with a much earlier (and more confusing) version of these tools. My detailed write up is on MediaWiki.org.

Tersely[edit]

Setting terseness to 2 (-t 2) would remove those same 60 folded items from the Changed Collisions/Net Gains list.

With more complex changes, terseness and folding do highlight different elements of the changes, particularly when there are multiple new collisions into the same group. In this small English corpus, only two pre-analysis types, í and ɨ, collided into the same group.

I also use the two sections for different things: I like to skim Changed Collisions to see what's happening and use lines from Bad Collisions? for speaker review. Sometimes a handy regex will let you grep exactly the right subset you want, too.

Folded English Self Analysis[edit]

We'll use the same folded English count file (with -n) that we used in the comparison analysis for this self analysis, enabling the "explore" option (-x) for automatic prefix/suffix detection, including the English language config (-l english) to see what it knows, and requesting the HTML report (-h) for its extra linky goodness.

  • ./compare_counts.pl -n samples/output/en.counts.folded.txt -x -l english -h > samples/output/en.comp.folded_self.html

Open samples/output/en.comp.folded_self.html in your favorite browser. In the table of contents, click on Potential Problem Stems, which will take you to the first group that doesn't have either a common beginning or ending substring.

  1. The first group is stemmed to d, and includes —which is just something the default folding happens not to cover. Click on the red d for the next group...
  2. The stem dutch includes tokens Dutch and Holland. Neat! Click on red dutch to continue...
  3. Stem g and IPA token ɢ. Nothing to see here...
  4. Stem philippines includes tokens Filipino and Philippines. Neat again...
  5. Stem re includes IPA token ʁɛ̃, which includes a combining ~ that is not covered by the default folding...
  6. Stem t and token ṭ, similar to the situation with d in (1) above.
  7. Stem vila and IPA token ˈvilɐ. I don't usually folding characters to nothing, so that stress accent is still there...

... and that's it. Clicking on the last one takes you to the end of the Stemming Results. If you scroll up a bit, you can see the last two items, which are in Hebrew. The table plus our friend dir=auto helps a lot with the formatting.

Nothing terrible, and we learned that there are some specific lexical exceptions in the English tokenizer. And so searching for holland east india company on English Wikipedia gets you the right thing.

Next up, English doesn't have any Common Prefix Alternations, but it has lots of Common Suffix Alternations. Many of them are in the English language config, but some are not yet. Silent e, as in hope/hoped/hopes is responsible for the -/d and -/s alternations (the first two in red). Pluralizing words ending in -y often does indeed give -ies. There's some good stuff at the top of the list. Towards the bottom we see a small number of semi-spurious alternations that are caused by doubling final letters before adding -ing or -ed. And some weird put understandable stuff, like iness/y (from emptiness/empty and loneliness/lonely, but fortunately notbusiness/busy!).

Mining these isn't going to do much for English because the stemmer here doesn't seem to know any prefixes, so common beginning substrings are usually easy to find. But for languages with both prefixes and suffixes (like Polish) this helps us find regular alternations that will decrease the number of items with no common beginning or ending substrings—or odd alternations that might indicate something weird or unexpected is going on.

The Histogram of Case-Insensitive Type Group Counts shows that we skipped 14,580 singletons (which you can get see with the -1 option if you want them). A quick skim of the groups under Stemming Results shows a lot of variation in the two-item groups. They aren't all, say, singulars and plurals. A larger corpus might show a larger typical non-singleton group.

The overly-verbosely titled Histogram of Stemmed Tokens Generated per Input Token shows that the English analyzer is generally one-token-in, one-token-out.

The Histogram of Final Type Lengths shows that there are lots of 1-character CJK types, and that the longest types are generally technical and/or German. More or less as expected.

Highlights from other Analyses[edit]

Since the English example ended up being longer than I expected, I'm not going to work another example right now. Instead I'll point out some interesting stuff this kind of analysis has let me find, pointing back to my notes on MediaWiki.org.

  • Unexpected folding in French: When looking at whether to add ASCII folding to to the French analyzer, I discovered that some unexpected folding was already being done by the analyzer! In order to get more comprehensive ASCII folding, we were using the the guide that Elasticsearch provides for unpacking the default analyzers into their constituent parts for customization. We expected the unpacked analyzer to be identical; I tested it primarily to make sure I'd configured the unpacked version correctly. Turns out, there are some minor differences that come out in large corpus testing!
  • Upgrading to a real Polish analyzer: The analysis revealed that the Stempel Polish analyzer was unexpectedly stripping non-grammatical prefixesnie-, naj-, anty-, nad-, and bez- (roughly "-less", "-est", "anti-", "super-", and "un-"), as previously mentioned above. This isn't necessarily a bad thing, just unexpected, since most stemmers only strip grammatical affixes (like verb conjugations and noun declensions). More surprisingly, the statistical nature of the stemmer was causing it to glitch, especially on numbers and foreign words. This is what lead me to several of my heuristics, particularly looking at very big groups and groups with no common beginning or ending substrings. Despite the problems, Stempel is still very cool and seems to be an improvement overall.
    • If you want to do some Polish analysis for yourself, I've included samples/output/pl.counts.default.txt and samples/output/pl.counts.stempel.txt, which are the default and Stempel count files for the 1MB Polish sample in samples/data/pl.txt.
  • Doing exactly the wrong thing in Swedish: Nothing too exciting came from my initial Swedish analysis, except that it made it very clear during speaker review that we were doing exactly the wrong thing. Once we did it right, there were still some sub-optimal collisions, but the impact was much smaller and much less violent to Swedish words.
  • Ukrainian is very similar to Russian: For historical reasons, Ukrainian-language projects used to use Russian morphological processing. After discussing it with a Ukrainian speaker, we decided it was, surprisingly, better than nothing, especially once we made the Russian config aware that it was used for Ukrainian. Analysis showed lots of new collisions (the Ukrainian analyzer stemming stuff that the Russian analyzer missed) and lots of splits (the Russian analyzer stemming stuff it shouldn't have). The analysis of Ukrainian also revealed a small number of fairly dodgy-looking groups. Speaker review revealed that they were generally okay—which restored confidence. And, really, should I, as an English speaker, be surprised by anything in any other language, given that English has be / being / been / am / is / are / was / were?
  • Chinese, Traditional and Simplified: When I went to implement a new analysis chain for Chinese, it included a plugin for Traditional to Simplified conversion (T2S), which was needed because the available tokenizers only work well on Simplified characters. Unfortunately, the Elasticsearch T2S conversion is separate from the MediaWiki display T2S conversion, so mismatches were possible. Fortunately, I was able to extract the MediaWiki display T2S conversion and convert it to folding info in compare_counts/langdata/chinese.txt, which allowed me to ignore any collisions caused by mappings that the two independent converters agreed on. Whew! That reduced thousands of conversion-related collisions down to a manageable 99.
  • Hebrew likes affixes and dislikes vowels: The HebMorph analyzer made me pay attention to how many tokens an analyzer generates. There was a huge increase in total tokens in my corpus, from 2.5M to 6.1M. Because of the ambiguity in Hebrew—vowels are mostly omitted and there are lots of common one- and two-letter prefixes and suffixes—the analyzer doesn't give just one best guess at the root form, but typically two and often three. Using different options for the analyzer got the count as low as 3.8M.
  • Loooooooong Japanese Tokens: The switch from the default CJK analyzer (which uses overlapping bigrams for CJK characters) to the Japanese-specific Kuromoji decreased the number of tokens found from 5.5M to 2.4M in my corpus—that's a lot fewer overlapping tokens, and many words longer than 2 characters. However, the lost/found token analysis showed that it was ignoring tokens in at least a dozen other scripts, and doing weird things with fullwidth numbers. The Kuromoji analyzer also revealed the need to look at really long tokens. In addition to some long Latin-script and Thai tokens, there were some really long Japanese tokens, which turned out to be tokenizing errors; these phrases and sentences just weren't broken up.
  • Custom tools: Sometimes you need custom tools because there is no end to the weird and wonderful variation that is the awesomeness of language.
    • Another problem with Japanese was that speakers doing review asked for samples of where the tokens had come from. I was able to hack analyze_counts.pl to generate samples for a specific list of tokens, but it was too messy to keep in the code. A proper implementation might be useful.
    • For Chinese and Japanese—both spaceless languages—I used an external tool to evaluate segmentation done by the analyzer. For Chinese, the tool came with a segmented corpus. For Japanese I had to find an annotated corpus and munge it into a useful form. Fun times! (Everything is described in more detail, with links, in my write ups.)
  • Using multiple text editors/viewers is sometimes helpful: I sometimes use less on the command line to look at compare_counts.pl reports. Sometimes I use vi, sometimes TextWrangler. And for HTML output, I use often use Chrome, unless it is giving me trouble with particular character sets, then I use Safari, which often renders complex text and non-Latin scripts best. A few interesting things have come up:
    • A doubly accented Russian word, Ли́́вшиц, can show up with two accents on top of each other, with one accent (presumably actually two overlayed), or with the accents as separate characters. When overlayed, they'd be impossible to find!
    • Some text processors hide bi-directional marks and various zero-width Unicode characters. I try to highlight them now, but I never would've found them in certain editors.

Stuff Yet To Do[edit]

Here are a bunch of things I should probably do, but may never get around to:

  • analyze_counts.pl
    • Add config options for host, port, index, and analyzer.
    • Line batching:
      • Add config for number of lines and total size of input.
      • Check for max input size before tacking on the new line rather than after (but it's so much more complicated!).
    • Proper implementation of context sampling for specified tokens.
  • compare_counts.pl
    • Add checks that language files and other config files exist and/or open properly.
    • Refactor / consolidate strip_*fixes and regular_*fixes.
    • Better error checking for input format.
    • Make the "hi-freq tokens" cutoff (currently 500) configurable.
    • Make the combined group/one-by-one cutoff (currently 4) configurable.
    • Add samples to Histogram of Stemmed Tokens Generated per Input Token.
    • Add samples to Common Prefix/Suffix Alternations.
    • Expose the hidden parameters.
    • Add a random seed so Histogram of Final Type Lengths is consistent between runs (useful during dev).
  • Why Not Both!
    • Use proper JSON parsing.
    • Explicitly specify output files.
    • Optionally disable progress indicator.
    • Add some unit tests.

Contact me if you want to encourage me to prioritize any of these improvements over the others.

Disclaimer: Believe it or not, these tools started out as various Unix utilities and Perl one-liners piped together on the command line. Their development has been a process of discovery and evolution, rather than top-down design, because I didn't know what I needed to know until I found that I didn't know it. ;)

As a result, there's some pretty crappy code in here, and all of it needs a top-down refactor, which will likely never happen. But I was told—quite correctly—that I should share my tools to up the bus number on this whole "analysis analysis" thing.

I've tried to do some basic clean-up with perlcritic and Perl Best Practices, but like many developers, the horror of ugly-but-working code is often not enough to overcome inertia and/or other priorities.

Hopefully the extensive documentation makes up for the sub-optimal code.