Syntax: Constituents, Dependencies, and Parsing

In “The inspector questioned the engineer near the turbine,” the phrase “near the turbine” may describe the engineer, the questioning event, or the inspector. All readings use the same words.

Syntax represents possible structural relationships that constrain interpretation. A parser selects one analysis, often with uncertainty and domain-dependent errors. That selection is also where parsers fail most. In two independent error taxonomies, five years and two formalisms apart, prepositional attachment was the largest single error category in every parser examined. The ambiguity in the sentence above is not a teaching device. It is the measured top of the error list.

Both can express useful organization while emphasizing different objects.

Constituency groups words into nested phrases such as noun and verb phrases. It represents phrase boundaries, is natural for hierarchical composition, gives trees with nonterminal categories, and is useful when the thing you need is a span. Its canonical English resource is the Penn Treebank, and the next section reads that corpus as a budget.

Dependency connects words through directed head–dependent relations. It centers lexical heads, keeps a compact relation structure, works well across flexible word order, and supports relation and feature extraction. Its cross-lingual resource is Universal Dependencies, whose spread across languages is examined further down.

Shallow syntax marks chunks or local patterns without a complete tree. It is a cheaper representation, useful for noun phrases, and avoids some global commitments — at the cost of missing long-range structure.

The exact label inventory depends on the annotation framework. But the reasoning is broadly reusable, and it is a sequence of questions rather than a diagram.

Which word is the predicate or head that organizes the central event or phrase? Which are the core arguments — who acts, what is affected, which complements are selected? Which words are modifiers, adding time, location, manner, quantity or description? How are the function words and features expressed: case, auxiliaries, negation, coordination, determiners? And how do the clause connections attach, for subordinate, relative and coordinated clauses?

Every one of those questions has an answer the parser had to choose. The modifier question is the one it gets wrong most often.

A useful error set targets constructions rather than random sentences. The ordering below is an empirical result, not an author's hunch.

Somebody counted. On Wall Street Journal section 23, the self-trained reranked Charniak parser produced 846 prepositional-phrase attachment errors involving 1,455 nodes. That was the top category for every parser in the study. Kummerfeld and colleagues state it flatly: “As expected, PP attachment is the largest contributor to errors, across all parsers.” Five years and a change of formalism later, Qi and Manning's error breakdown on PTB-SD dependencies put the same category top again: 545 of 3,884 total errors for arc-swift, 569 of 4,100 for arc-eager, 571 of 4,106 for arc-standard.

The decision is not easy for people either. Three Penn Treebank annotators judged 300 randomly selected Wall Street Journal test events. Shown only the four head words, they were 88.2% correct. Shown the whole sentence, 93.2%. The maximum-entropy model tested on those same 300 events scored 78.0%, and 81.6% on the full 3,097-event test set. The five points the humans gained by being allowed to read the rest of the sentence is the part worth keeping. Ratnaparkhi and colleagues put it this way: “Secondly, the ME model does not use information contained in the rest of the sentence, although it is apparently useful in predicting the attachment, as evidenced by a 5% average gain in the treebankers' accuracy.”

The parser introduces its own errors, so downstream value must be measured end to end. Five steps, in order.

First, name the structural need: subject–object roles, clause scope, modifier attachment, or phrase spans.

Second, establish a no-parse baseline. The attachment literature shows exactly what that discipline is worth. The standard benchmark extracted from the Penn Treebank Wall Street Journal material holds 20,801 training, 4,039 development and 3,097 test items. Collins and Brooks describe it in one line: “The data consisted of training and test files of 20801 and 3097 quintuples respectively.” Two baselines on that data involve no parsing at all. Always attach to the noun scores 59.0%. Using the attachment most often seen for that preposition scores 72.2%. Their backed-off model reaches 84.5%, dropping to 81.6% once training events occurring fewer than five times are discarded. A one-line rule gets 72.2%; the published model buys roughly twelve points on top of it. Whether twelve points justify a parser in your pipeline is a product question, and it is the same shape of question every time.

Third, select suitable language resources, matching annotation scheme, language, genre and license. The CoNLL 2018 spread above is what happens when that match is assumed rather than checked.

Fourth, inspect parser error propagation: review the downstream failures caused by incorrect heads, boundaries or labels. Attachment errors will dominate them.

Fifth, retain uncertainty or alternatives, rather than forcing brittle rules when several analyses are plausible.

Dependency parsing often reports labeled or unlabeled attachment. Constituency parsing reports span-based scores. Aggregate metrics can hide weak negation scope, coordination, long dependencies, or a target language.

It is worth seeing once what a headline LAS number is an average of. The CoNLL 2018 shared task on multilingual parsing from raw text to Universal Dependencies ranked 25 participating systems plus a UDPipe 1.2 baseline on 82 test treebanks in 57 languages. HIT-SCIR won, and its own abstract reports the result: “In the final evaluation, our system was ranked first according to LAS (75.84%) and outperformed the other systems by a large margin.” That 75.84 is macro-averaged over the 82 test sets. The UDPipe 1.2 baseline reached 65.80, which was rank 18. Only 17 of the 25 systems beat it.

So the number is not a score on a language. It is one figure standing in for 82 treebanks whose individual results, as the previous section showed, run from 94.86 to 13.70. Create phenomenon and domain slices. Then measure the downstream operation that motivated parsing. A small parse-score gain may have no value if the required relation remains unreliable.

Choose fifty sentences describing equipment failures. Build one baseline from local token windows and another that follows predicted dependency paths between entities. Score the dumb baseline before the clever one. On the classic attachment benchmark, simply doing whatever that preposition usually does already reached 72.2% before anyone parsed anything. Any structural machinery has to be measured against a number like that, not against zero.

Review cases involving negation, coordination, passive voice, and attachment. Attribute gains and failures to the parser, the relation classifier, or the annotation scheme.