Semantic Role Labeling

In his Sanskrit grammar (7th–4th c. BCE), rules describing semantic relationships between a verb and its noun arguments (agent, instrument, destination…): the earliest known model of how events and their participants are realized linguistically.

“Who did what to whom” (and perhaps when and where): how participants relate to events.

The same purchase event and the same buyer appear with different verbs or nouns and in different syntactic positions (subject, indirect object, by-phrase, compound); semantic roles capture the shared shallow meaning: XYZ Corp is the buyer, the stock the thing bought.

Preferences a predicate expresses about its arguments, e.g. that the theme of eat is generally something edible.

Both are volitional, often animate actors with direct causal responsibility for their events: both are AGENTs. The objects, prototypically inanimate things affected by the action, are THEMEs.

Roles capturing semantic commonalities between arguments of different predicates (AGENT, THEME…); modern formulations are due to Fillmore (1968) and Gruber (1965). There is no agreed universal set; most have about a dozen.

The wind is a FORCE (non-volitional causer), debris the THEME, from the mall the SOURCE, into our yards the GOAL.

As a shallow meaning representation that supports simple inferences impossible from surface strings or parse trees, e.g. that “Company A acquired Company B” answers “Was Company B acquired?”; also as an intermediate language, e.g. for MT.

AGENT/subject, THEME/object, INSTRUMENT/with-PP (“John broke the window with a rock”); INSTRUMENT/subject (“The rock broke the window”); THEME/subject (“The window broke”, “The window was broken by John”).

The set of thematic role arguments a verb takes.

Multiple argument structure realizations of a verb’s roles. Dative alternation: “Doris gave the book to Cary” vs “Doris gave Cary the book”, the THEME and GOAL in either order.

Particular semantic classes: verbs of future having (offer, owe), send verbs (mail, hand), verbs of throwing (kick, throw)… Levin (1993) lists the classes and alternations of thousands of English verbs; VerbNet puts them online, linked to WordNet and FrameNet.

It is hard to agree on a standard role set or define roles formally; roles fragment into many subtypes; and finite discrete lists don’t support some generalizations. E.g. an AGENT is usually animate, volitional, sentient and causal, but a given NP may lack some of these.

Intermediary instruments can be subjects (“The knife cut the banana”), enabling instruments cannot (“Shelly ate the banana with a fork” but *“The fork ate the banana”).

Generalized semantic roles defined not by necessary and sufficient conditions but by heuristic features: the more agent-like properties an argument has (volitional, causing change, sentient, moving), the likelier it is PROTO-AGENT; the more patient-like (undergoing change, causally affected, stationary), PROTO-PATIENT.

PropBank uses proto-roles and verb-specific roles; FrameNet uses roles specific to a frame, a semantic situation shared by a group of words.

The Proposition Bank: sentences (e.g. the Penn Treebank) annotated with semantic roles defined per verb sense and named only by number: Arg0 is generally the PROTO-AGENT, Arg1 the PROTO-PATIENT, higher numbers verb-specific.

Less consistently: Arg2 often benefactive, instrument, attribute or end state; Arg3 start point, benefactive, instrument or attribute; Arg4 the end point.

A per-verb-sense list of roles with informal human-readable glosses. fall.01 “move downward”: Arg1 thing falling, Arg2 extent/amount fallen, Arg3 start point, Arg4 end point. “[Arg1 Sales] fell [Arg4 to $25 million] [Arg3 from $27 million].”

Its normal subject is a patient-like PROTO-PATIENT, not an agent.

In each, Big Fruit Co. is Arg0 (agent) and the price of bananas Arg1 (theme), despite different syntax; 5% is Arg2, the amount increased.

Non-numbered modifier/adjunct arguments stable across predicates: ArgM-TMP (time), ArgM-LOC (location), ArgM-DIR (direction), ArgM-MNR (manner), ArgM-PRP/CAU (purpose, cause), ArgM-NEG (negation), etc.

A PropBank-style annotation of noun predicates: in Apple’s agreement with IBM, Apple is Arg0 and IBM Arg2.

Commonality across different verbs and between verbs and nouns: in “the price of bananas increased 5%”, “…rose 5%” and “a 5% rise in the price of bananas”, the same thing went up by the same amount.

The holistic background knowledge that unites a set of words, e.g. air travel for reservation, flight, travel, buy, price, fare, meal, plane; related to models and scripts in AI and cognitive science.

A background knowledge structure defining frame-specific semantic roles (frame elements), the predicates (lexical units) that evoke it, and labeled example sentences. Each word evokes a frame and profiles some aspect of it.

Core roles are frame-specific; non-core roles express general properties like time and location, similar to PropBank ArgMs.

ITEM (entity with a position on the scale), ATTRIBUTE (the scalar property), DIFFERENCE (distance changed), INITIAL_VALUE, FINAL_VALUE, INITIAL_STATE, FINAL_STATE, VALUE_RANGE.

[ITEM Oil] rose [ATTRIBUTE in price] [DIFFERENCE by 2%].

Frames can inherit from each other and be linked by relations like causation: Cause_change_of_position_on_a_scale adds an AGENT (“[AGENT They] raised [ITEM the price of their soda] [DIFFERENCE by 2%]”).

FrameNet (frames, frame elements, lexical units; also for Spanish, German, Japanese, Portuguese, Italian, Chinese…), PropBank (English and Chinese), NomBank (noun predicates), VerbNet (verb classes), and the CoNLL-2005 and CoNLL-2012 SRL evaluation sets.

Automatically finding the semantic roles of each argument of each predicate in a sentence; usually supervised learning on FrameNet or PropBank.

Parse the sentence; find all predicates; for each predicate, classify every parse-tree node into a semantic role or NONE with a 1-of-N classifier trained on labeled data.

Pruning (heuristically discard unlikely constituents), identification (binary: argument or NONE), and classification (1-of-N role labeling of identified arguments).

Arguments aren’t independent: constituents must not overlap, and a verb can’t have two Arg0s. Local classifier probabilities are combined by Viterbi-style decoding, reranking or integer linear programming.

Governing predicate, phrase type, headword and its POS, the parse path from constituent to predicate (e.g. NP↑S↓VP↓VBD), voice (active/passive), linear position (before/after the predicate), subcategorization of the predicate (e.g. VP → VBD NP PP), named entity type, and first and last words of the constituent.

It compactly represents the grammatical function relation between constituent and predicate: NP↑S↓VP↓VBD is a typical subject path.

Passive clauses link semantic roles to surface positions very differently from active ones (the subject of a passive is usually Arg1).

Given the predicate, tag spans with BIO labels per role (B-ARG0, I-ARG0, B-ARG1…, O): encode the sentence plus the predicate (e.g. after [SEP]), concatenate each token’s encoding with a predicate indicator, and classify each token with a feedforward softmax (a CRF adds little).

Each argument label must be assigned to exactly the right word sequence or constituent; then precision, recall and F1 (identification and classification can be scored separately), e.g. on CoNLL-2005 or CoNLL-2012.

The THEME of eat should be edible; someplace nearby isn’t, so it must be a locative adjunct of intransitive eat, not the direct object (the “speaker as Godzilla” reading).

Senses: serve in “serves green-lipped mussels” restricts its THEME to food; in “Which airlines serve Denver?” to locations.

Very variable: imagine restricts its AGENT to animate beings but its THEME hardly at all; diagonalize requires a matrix; odorless requires something that could have an odor. The needed concept set is open-ended.

Add a type constraint on the role filler: \(\exists e, x, y\; Eating(e) \wedge Agent(e, x) \wedge Theme(e, y) \wedge EdibleThing(y)\). “ate a takeoff” becomes inconsistent.

FOL is overkill and requires a huge commonsense knowledge base. Instead, specify a WordNet synset per argument: a filler is acceptable if it is a hyponym of that synset (eat’s THEME = {food, nutrient}; hamburger’s hypernym chain reaches food).

hamburger, beefburger
 ⇒ sandwich
  ⇒ snack food
   ⇒ dish
    ⇒ nutriment, nourishment
     ⇒ food, nutrient
      ⇒ substance … ⇒ entity

Restrict them to {entity}, {physical entity} and {matrix}: this allows imagine/lift a hamburger but rules out diagonalize a hamburger.

Violations occur in well-formed sentences, e.g. under negation (“you can’t eat gold for lunch”) or when restrictions are overstated (“Mr. Kulkarni ate glass on an empty stomach”).

How much information a predicate gives about the semantic class of its arguments: the KL divergence between the class distribution for that verb’s arguments and the general class distribution, \(S_R(v) = D(P(c \mid v)\,\|\,P(c)) = \sum_c P(c \mid v)\log\frac{P(c \mid v)}{P(c)}\). Eat has high strength, be low.

The relative contribution of a class to a verb’s preference strength: \(A_R(v, c) = \frac{1}{S_R(v)}P(c \mid v)\log\frac{P(c \mid v)}{P(c)}\), estimated from parsed corpora treating each word as a partial observation of all WordNet classes containing it.

The conditional probability of an argument noun given the verb and relation, \(P(n \mid v, r)\), from co-occurrence counts in a huge parsed corpus; the inverse \(P(v \mid n, r)\); or log co-occurrence counts (better for subjects).

Replace each verb’s actual object (car) with a pseudoword combining it with a frequency-matched confounder (car/house); the model must choose which is the original object of drive. Alternatively, correlate with human plausibility ratings obtained by magnitude estimation.

Decomposing word meanings, especially predicates, into primitive semantic elements: KILL(x, y) ⇔ CAUSE(x, BECOME(NOT(ALIVE(y)))).

One state predicate OPEN combines with primitives: CAUSE(John, BECOME(OPEN(door))), BECOME(OPEN(door)), OPEN(door).

A more radical decomposition into about ten primitive predicates used in early NLP systems, e.g. ATRANS (abstract transfer of possession), PTRANS (physical transfer), MTRANS (mental transfer), MBUILD, PROPEL, MOVE, INGEST, EXPEL, SPEAK, ATTEND, each with fixed roles. “The waiter brought Mary the check” = ATRANS + PTRANS.

Studying Tesnière’s dependency grammar, Fillmore called argument roles actants, then cases, proposing a universal list (Agent, Patient, Instrument…); verbs were listed with their case frames.

Fillmore’s idea that words are understood against packages of connected expectations about situation types (frames, schemas, scenarios, scripts), presupposed as background.

Their consistent large annotations recast SRL from 1970s handwritten rules into supervised machine learning (Gildea & Jurafsky 2000 on FrameNet, then PropBank), later with dependency features and neural models (biLSTMs, then transformers).

Finding arguments that appear only in another sentence or must be inferred: in “This house has a new owner. The sale was finalized 10 days ago.”, the Arg1 of sale is the house.

Frames marking richer information about predicate arguments, such as the writer’s or reader’s sentiment toward them: “he survived a bombing” expresses sympathy for he and negative sentiment toward the bombing.