STUnD

A GUI Search Tool for (bilingual) parallel Universal Dependencies treebanks that runs in your browser.

Usage

STUnD is available at demo.spraakbanken.gu.se/stund.

Specifying input files

The "Browse..." buttons are used to specify one or two (parallel) input files, which have to be in strict CoNNL-U format.

If you only specify the input file(s), leaving the other fields blank, clicking "search" will run a default query that returns the full treebank:

Running a query

Queries are specified in first text input field:

(note that double clicking on it will show the query history).

Monolingual queries

The example query in the picture is

TREE_ (FEATS_ "VerbForm=Sup") [AND [LEMMA "ha", FEATS_ "Tense=Pres"]]

This is a simple or monolingual query, looking for present perfect constructions in the Swedish treebank. It reads as

Look for (sub)trees (TREE_) where the root is a supinum ((FEATS_ "VerbForm=Sup")) and one of its direct dependents is the present of the verb "ha" (AND [LEMMA "ha", FEATS_ "Tense=Pres"]).

Now only the subtrees matching the query (often full trees in this case) are highlighted in bold (cf. last row).

With some knowledge of Swedish, this particular query can be rewritten more concisely as

TREE_ (FEATS_ "VerbForm=Sup") [FORM "har"]

It is then very easy to modify the query for other structurally similar tenses:

• TREE_ (FEATS_ "VerbForm=Sup") [FORM "hade"] (pluperfect)
• TREE_ (FEATS_ "VerbForm=Sup") [FORM "ha"] (perfect infinitive)

Compared to the original "present perfect" query, these make it easier to see how queries work: first, the program tries to align the two treebanks to identify semantically equivalent subtrees; then the query is run on the left (Swedish) treebank and matching subtrees are returned alongside their English counterpart. For this reason, it can find correspondences such as "att ha sagt"-"as saying" (row 4), even though the English construction is not at all similar to the Swedish one. Unlike query matching, the alignment step is not guaranteed to find all correspondences: in the picture below, you can see that sometimes a match is found in the Swedish treebank but nothing is highlighted in the corresponding English sentence (this is the case, for instance, in the last few rows). Some other times, a correspondence is found but is incorrect.

Of course, monolingual queries can also be run on single treebanks:

Parallel queries

Queries can also be parallel or bilingual. For instance, we can use the following pattern to serach for sentences where a Swedish present perfect corresponds to a passive present tense in English:

TREE_ (FEATS_ "VerbForm={Sup->Part}") [AND [LEMMA "{ha->be}", FEATS_ "Tense=Pres"]]

This produces the following results:

Note that the second hit here is a false positive, due to the fact that "are" in the clause "there are already..." is also a direct dependent of the main lexical verb "dropped". This is unfortunate, but difficult to avoid given how conjuncts are treated in UD.

The basic query language ("UD patterns") is described here, while the extended version for parallel (bilingual) queries ({X -> Y} syntax) is documented here.

Adding a replacement rule

The last input field can be used to specify a replacement rule to be applied to all matching subtrees in both languages. This can help highlighting the relevant parts of each query result and manipulate them.

Understanding replacement rules, which are described alongside the basic query language, can be slightly more challenging.

As a first example,

PRUNE (UPOS "VERB") 0

decreases the depth of trees rooted in a verb to 0, eliminating all dependents:

The more complex pattern

CHANGES [FILTER_SUBTREES (OR [UPOS "VERB", DEPREL_ "root", DEPREL_ "csubj", DEPREL_ "ccomp", DEPREL_ "xcomp", DEPREL_ "advcl", DEPREL_ "acl"]) (OR [DEPREL_ "aux", DEPREL_ "cop"]), PRUNE (DEPREL_ "aux") 0, PRUNE (DEPREL_ "cop") 0]

uses dependency labels to isolate verb constructions of maximum depth 2, thus producing, in conjunction with the first query, the following output:

CoNNL-U and tree mode

So far, we have seen how to use STUnD in plain text mode. Switching to CoNNL-U mode allows inspecting the CoNNL-U (sub)trees corresponding to bold text in the default text mode:

Tree mode renders them as SVG trees:

Saving the search results

Query results can be saved as plain text/TSV, CoNNL-U and HTML-embedded SVG trees. The output format depends on the mode (in the example below, for instance, results would be saved as graphical trees).

If two treebanks are being compared, results can be saved as two separate files, one per treebank, or as a single "parallel" file:

• in text mode, the T1 and T2 files are simple text (one sentence per line), while the parallel file is tab-separated. This makes it easy to import search results in any spreadsheet program
• in CoNNL-U mode, the output is always a new treebank, treebank that can, for instance, be used as input for more refined queries in StUnD, or simply imported into a CoNNL-U viewer for further inspection.[^1] In the parallel file, sentences from the two input treebanks are alternated
• similarly, parallel files in tree mode consist in alternating T1-T2 trees

Other use cases

So far, we have shown how to use STUnD on multilingual treebanks. Many of the tool's functionalities, however, are also relevant in other scenarios, such as comparing learner sentences with corrections:

In the image above, you can see STUnD in action on the VALICO treebank of L2 Italian, looking for feminine nouns incorrectly inflected as masculine.

By checking "highlight discrepancies", in addition, the tool will highlight all sentence pairs matching the query that present any difference:

In the case of a learner corpus, "discrepant" means "erroneous", but highlighting discrepancies can also be useful in other settings, such as when comparing different analyses of the same text to resolve disagreement in a linguistic annotation project. This functionality is, however, still very rudimentary. In the future, the plan is to refine it to only highlight sentences where the discrepancy occurs in the subtree matching the query.

What's happening under the hood?

STUnD is basically a GUI for L2-UD match, which allows running parallel queries on UD treebanks by combining UD-based subtree alignment with UD tree pattern matching.

While L2-UD (where "L2" stands for "second language") was originally developed for comparing learner texts and corrections, the current STUnD prototype uses a slightly modified, "bilingual" or "2L" version of L2-UD, optimized for comparisons between different languages. In the future, it will be possible for the user to choose what version to use, meaning that the tool will be usable on both types of treebanks.

STUnD is now a web application composed of a Haskell server and a JavaScript frontend. The code for its first locally-run prototype GUI is still available (but not actively maintained) on the threepenny branch.

Installation

If you want to compile and run STUnD directly on your Linux/Windows computer, you can use the Haskell Tool Stack or build and run it inside a Docker container.

Compile using Stack

If you do have Stack, you can run

stack build

or

stack install

which also installs the executable, should be all that is needed enough. On Windows, however, you might encounter problems related to curl, STUnD's only external dependency. If that's the case, see this.

Afterwards you can run STUnD using either stack run or run the installed executable directly.

Run using Docker

If you want to use Docker containers, the simplest way is to use docker compose. To build and run the image you can simply type:

docker compose up stund-gui

This will take a while for the first time because the image has to be built. Afterwards, running the container can be started directly with the same command.

Citing

If you use this tool, you are welcome to cite

@inproceedings{masciolini2024stund,
  title={{STUnD}: ett S{\"o}kverktyg f{\"o}r Tv{\aa}spr{\aa}kiga Universal Dependencies-tr{\"a}dbanker},
  author={Masciolini, Arianna and T{\'o}th, M{\'a}rton A},
  booktitle={Huminfra Conference},
  pages={95--109},
  year={2024}
}

A more extensive and up-to-date publication in English is currently in preparation.

[^1]: technical note: this works because all extracted subtrees are adjusted so that they have a root node and valid (sequential) IDs.