This document gives a detailed description of a linguistic annotation scheme for multiword expressions (MWEs) in English sentences. A conference paper summarizing the scheme and a dataset created with it are available at http://www.ark.cs.cmu.edu/LexSem/.

The input to annotators is a tokenized sentence. The goal is to join tokens where appropriate; this is done with the special characters underscore (_) for strong multiword links, a tilde (~) for weak links:

• This is a highly~recommended fast_food restaurant .

Weak links can join strongly-linked expressions, such as fast_food + chain:

• This is a highly~recommended fast_food~chain .

Where an expression is interrupted by other tokens, use trailing and leading joiners:

• Do n't give_ Jon such _a_hard_time !

This even works when contiguous expression can fall within the gap:

• Do n't give_ Jonathan_Q._Arbuckle such _a_hard_time !

On rare occasion it may be necessary to use multiple gappy expressions, in which case indexing notation is available: a|1 b|2 c|1$3 d|2 e$3 implies two strong expressions—a_c and b_d—and one weak expression, a_c~e. An example: put_ whole$1 _heart_in$1 (amounts to: put_heart_in~whole). Also: make_a_ big$1 _to_do$1

Don't correct spelling mistakes in general, but interpret misspelled words according to their intended meaning. (E.g., if put it is clearly a typo of the particle verb put in, mark put_in.)

Do add _ if two separate tokens should really be one word. Include intermediate hyphens as part of an MWE:

anti - oil wells

or

anti - oil_wells

or

anti_-_oil_wells

NOT

anti_- oil_wells

or

anti -_oil_wells

If there’s a nonstandard variant of a term (more than just a misspelling), don’t join: craft_beer BUT handcraft beer

In general, don’t worry about the inflection of the word: come_in, came_in, have been coming_in all map to the same MWE lemma; likewise for grocery_store, grocery_stores.

If different pronouns can be used in a slot, do not join the pronoun. But if the slot requires a possessive, do join the possessive clitic if present: take_ her _time, take_ one _’s_time

While many idioms/MWEs are figurative, not all figurative language is used in a lexically specific way. For example, “my Sicilian family” referring to a pizza community may be nonliteral, but is not an MWE.

We do not annotate foreign sentences, but foreign names within an English sentence are MWEs.

A collocation is a pairing between two or more words that is unexpectedly frequent, but not syntactically or semantically unusual.

• eternally~grateful, can~not~wait, place~is~packed
• what s.o. has~to~say [is willing to say, brings to the conversation; not obligation (compare: I have to say good things to my boss to get promoted)]

A collocation may include components that are themselves MWEs:

• after~all~was_said_and_done

Drawing the line between free combination, collocation, and multiword is often difficult; annotators’ opinions will vary.

TODO (join with ~)

• Cj~and_company (ambiguous whether it is actually the name of a company or a guy and his crew)

Some semi-productive idioms are not well captured as lexicalized multiwords. These should not be joined:

• have + GOODNESS.ADJ + TIME.PERIOD: had a bad day, have a great year, etc.
• EVALUATIVE.ATTRIBUTE of s.o.: (real) Christian of you
• NUMERIC.QUANTITY PLURAL.TIME.NOUN running: two years running
• come + MENTAL.CHANGE.INFINITIVE: come to realize, believe, learn, adapt, have faith, …

Rarely, a token will seemingly participate in multiple MWEs, which cannot be represented in our annotation scheme. Use your best judgment in such cases.

• I recently threw a surprise birthday party for my wife at Fraiser 's .

Possible pairs:

surprise_party

birthday_party

threw_party

Decision:

threw~birthday_party

• triple_chocolate_chunk brownie: multiplier+chocolate, chocolate_chunk

Don’t worry if the parts of an expression are noncanonically ordered: gave_estimates, give_ an _estimate, the estimate_ that was _given

If one of the lexicalized parts is repeated due to coordination, attach the instance closest to the other lexicalized part: talked_to Bob and to Jill; North and South_America

• DO join Dr., Mr., etc. and other titles to a personal name: Dr._Lori_Levin, Henry_,_Prince_of_Wales, Captain_Jack_Sparrow
• DO join Ave., Rd., St., etc. in street names: Forbes_Ave.
• DO join city-state-region expressions: Bellevue~,~WA or Bellevue~WA (include the comma if there is one). Likewise: Ohiopyle_State_Park~,~Pennsylvania; Miami_University~,~Miami~,~Ohio; Amsterdam~,~The_Netherlands
• DON’T join normal dates/times together (but: Fourth_of_July for the holiday)
• Symbols
  • DON’T join normal % sign
  • DO join letter grade followed by plus or minus: A_+
  • DON’T join mathematical operators: 3 x the speed, 3 x 4 = 12 [x meaning “times”]
  • DO join # sign when it can be read as “number”: #_1
• DO join a number and “star(s)” in the sense of a rating: 5_-_star
• When in doubt, join cardinal directions: north_east, north_west, south_east, south_west, north_-_northeast, …
• DO attach ’s if part of the name of a retail establishment: Modell_’s
• DO join product expressions such as car Year/Make/Model or software Name/Version
  • excludes appositions that are not in a standard format (McDonald’s Dollar_Menu Chicken Sandwich)
• DO join names of foods/dishes if (a) the expression is noncompositional in some way, or (b) there is well-established cultural knowledge about the dish. Use ~ if unsure. For example:
  • General_Tso_’s_chicken, macaroni_and_cheese, green_tea, red_velvet cake, ice_cream_sandwich, chicken_salad salad
  • triple_chocolate_chunk brownie [multiplier+chocolate, chocolate_chunk]
  • pizza~roll, ham~and~cheese, cheese~and~crackers, spaghetti~with~meatballs
  • grilled BBQ chicken, pumpkin spice latte, green pepper, turkey sandwich, eggplant parmesan, strawberry banana milkshake
• DO join established varieties of animals/natural kinds: yellow_lab, desert_chameleon, Indian_elephant, furcifer_pardalias; BUT: brown dog
• DO join slogans: Trust_The_Midas_Touch, Just_Do_It, etc.

• pleased/happy/angry_with, mad_at
• good_for s.o. [healthy, desirable]

This is a special use of the preposition ‘on’, but it is does not generally join to form an MWE:

• drop_the_ball on s.o. [not literally!], die on s.o.
• hang_up~on s.o. [collocation]
• (‘step on s.o.’ is different: here it is the semantics of ‘step on’ that could convey negativity in certain contexts, not ‘on’ by itself)

• (appropriate) for_ {one's, a certain, ...} _age (of child)

3 years_old, month_old project. (Note that ago should NOT be joined because it is always postpositional.)

Join unless ‘all’ is paraphrasable as ‘completely’ or ‘entirely’:

• participle: all gone, all done, all_told [overall, in total]
• other adj: all ready, all_right [well, OK] (informal spelling: alright)

Do not join, even though the as PPs are correlated. Exceptions:

• as_long_as [while]

• one_by_one [one at a time]
• Don’t join if by indicates a product, as in a multidimensional measurement: three by five paper = 3 x 5 paper

A few English nouns take idiosyncratic measure words: 3 sheetsofpaper, 2 pairsofpants, a pieceofinformation

Do not attach the modifier if it has an ordinary meaning, e.g. go clear through the wall

• highly~recommended, highly~trained
• family~owned company

• capital_punishment
• big_rig [slang for truck]
• road_construction [the road isn't actually being constructed, but reconstructed!]
• silver_ Mariott _member [rewards program]
• electric_blanket
• last_minute
• price_range
• second_chance
• grocery_stores
• pizza_parlor, pizza place, burger joint (diagnostic: does “favorite X” occur on the web? [to filter out proper names])
• little~danger/risk
• public~welfare
• this place is a hidden~gem
• strike_one/two/three (unusual syntax!)

Cf. Quirk pp. 669–670

• out_of, in_between, in_front_of, next_to
• along_with
• as_well, as_well_as
• in_addition_to, in_light_of, on_account_of
• due_to, owing_to, because_of

From Quirk et al. 1972:

• but_that, in_that, in_order_that, insofar_that, in_the_event_that, save_that, so_that, such_that, except_that, for_all_that, now_that
• as_{far,long,soon}_as, inasmuch_as, insofar_as, as_if, as_though, in_case
• Do NOT mark the participial ones: assuming, considering, excepting, … that

• to_start_off_with
• that_said
• of_course

Though as a postmodifier it is a bit odd syntactically (anything else, who else, etc.), it does not generally participate in lexicalized idioms.

• “What does ‘else’ even mean?!” - Henrietta

• do_ X _a_favor_and Y
do_ X _a_favor_, Y
  vs. plain do_favor
• you_get_what_you_pay_for (NOT: ‘you get what you purchase’)
• get_ the_hell _out
• why in_the_hell [can be any WH word]
• do_n’t_forget, forget_it !, never_mind
• i havetosay, gotta~say, etc.: semantics = obligation on self to say something, pragmatics = can’t restrain self from expressing an opinion
• Who_knows [rhetorical question]
• no_way
• Phatic expressions: I_’m_sorry, Thank_you

We do not mark ordinary there be existentials as multiwords.

get_upgraded, get_ cat _neutered, get_ a bill _passed

In the sense of ‘arrive’, not really a multiword:

• get back home, got to the school

• get_ready, get_done, get_busy, get_older
• get_a_flat
• get_correct

If a verb, adjective, or noun typically takes an infinitival complement, and the infinitive verb is not fixed, don't join to:

• little to say
• important to determine his fate
• able/ability to find information
• chose/choice to do nothing
• willing(ness) to sail

But if it is a special construction, the to should be joined:

• in_order_to VP
• at_liberty_to VP
• ready_to_rumble
• special modal/tense constructions: ought_to, have_to (obligation or necessity), going_to, about_to (but want to, need to, try to)

• a long~{day, week, year} (long in the sense of bad/difficult; cannot be referring to an actual duration because the noun is a time unit with precise duration)

Join these (including n’t and a, but not do) if sufficiently conventional: did n't_sleep_a_wink, did n't_charge_a_cent/penny/dime, did n't_eat_a_morsel/scrap/bite/crumb (of food)

These include the so-called determiner-less PPs (in_town vs. in the town).

• in_a_ nice/good/... _way
• out_of_site
• on_staff
• at_all
• at_liberty_to
• for_sure
• mediocre at_best
• to_boot
• in_town
• on_earth, ontheplanet, intheworld, in the country/universe

• in_cash/quarters/Euros

• to_ her amusement, to our chagrin, to the _surprise of all present
• to_ my _satisfaction

• capacity_for love
• his problem_with the decision
• extensive damage_to the furniture

Sometimes these participate in verb-noun constructions:

• have_a_problem_with [be annoyed with], have_ a _problem_with [have trouble with something not working]
• do_damage_to the furniture

• TODO: explain principles
• talk_with/to, speak_with/to, filled_with
• NOT: learn about, booked at (hotel)
• wait_for
• look_for
• test_for
• (a)rising_from
• disposed_of
• take_care_of
• trust_with
• listen_to, pay_attention_to
• compare_ X _to Y, X compared_to Y
• been_to LOCATION/doctor etc.
• trip_over, trip_on
• (do_)damage_to the furniture
• take_in [bring to an establishment for some purpose, e.g. a car for service]
• focus_on
• nibble/nosh/snack/munch_on
• kept_up_with [keep pace, manage not to fall behind]
• looking_for my friend [seeking out my friend] vs. looking for my friend [on behalf of]
• not multiword:
  • stay at hotel
  • supply with, fit_out with [‘with’ marks the thing transferred/given]

Join: close_to, close_by(_to) far_from, far_away(_from)

• Join article if not followed by a noun (paraphrasable with ‘identical’): his objective was the_same each time / each time he had the same objective
• exact_same

• walks_around (path covering an area)
• stay~away = keep_away
• run_out [get used up] vs. run out_of the filter [leak]
• back
  • Generally do not include literal uses: go back [motion], came/headed back [returned to a location]
  • money_back, cash_back [change of medium: overpaying with credit card so as to receive the difference in cash]
  • s.o.’s money_back, CONDITION or_your_money_back [refund]
  • pay_ s.o. _back, get money back [returning a loan; get money back is possible but not really idiomatic with this meaning]
  • brought back [taking a car to the shop again for further repairs] vs. brought_back [returning a purchase for a refund]
  • turned_back [turned around to travel back]
  • get_back_to s.o. [return to communicate information to s.o.]

• TODO: explain principles
• rent_out
  • with ‘out’, disambiguates the landlord/permanent owner vs. tenant/temporary user
  • (BUT: rent out an entire place?)
• turn_on, turn_off
• pick_up [retrieve from store]

• If prenominal, don’t join of: a_lot/little/bit/couple/few (of), some/plenty of
  • EXCEPTION: a_number_of (TODO: why?) check what H did in xxxx5x
• Join ‘square’ or ‘cubic’ within a unit of measurement: square_miles/yards/..., cubic_centimeter/...
• Join half_a when modifying a quantity: half_a day’s work, half_a million
• cf. classifiers

• less than, more than: Join if functioning as a relational “operator.” Heuristic: can ‘<’ or ‘>’ be substituted?
  • less_than a week later (‘< a week’)
  • more happy than sad (NOT: ‘> happy than sad’)
  • I agree with him more than with you (NOT: ‘> with you’)

• trying_to_say (with implication of dishonesty or manipulativeness)
• went_out_of_ their _way (went to extra effort)
• go_on_and_on [= talk endlessly] (cf. go_on by itself meaning ‘continue’)
• went_so_far_as_to_say: include ‘say’ because it has a connotation of negativity (beyond ‘went_so_far_as_to (do something)’)
• have_a_gift_for [= possess talent]
• love_ Boston _to_death
• go_to_the_bathroom [use the potty]
• give_ X _a_try [test it out]
• dropped_the_issue, drop_the_subject, drop_it !
• say/tell/lie_ (s.t.) to_ s.o. _’s_face
• made_to_order
• took~forever

A support verb is a semantically “light” (mostly contentless) verb whose object is a noun that by itself denotes a state, event, or relation; the noun can be thought of as providing the bulk of the meaning/determining the meaning of the verb [FN Book 2010, p. 31]. We join the verb to the noun in question:

• make_ a _decision/statement/request/speech/lecture
• take_ a(n) _test/exam
• take_ a _picture/photo
• give_speeches/lectures/interviews
• undergo/have/receive/get_ an _operation
• do/perform_surgery

Some useful properties:

1. Most commonly, support verbs are extremely frequent verbs that exhibit a certain degree of grammaticalization: have, get, give, make, take, etc.

2. One indication of lightness is when the noun cannot felicitously be omitted in a question (She made a decision. / #What did she make?; She had an operation. / ?#What did she have?; They perform surgery on newborns. / #What do they perform on newborns?)

3. Support verb constructions can often be paraphrased with a semantically “heavy” verb, which may be derivationally related to the noun: make_ a _decision = decide, give_ an _interview = be interviewed, undergo_ an _operation = be operated_on. (The noun surgery has no verb in English, but we could imagine “surgure” as a word! In other cases it would be not unreasonable to incorporate the noun into a verb: take_ a _test = test-take.)

4. Caution is required: some expressions are not support verbs, though they appear to be at first blush:

   • get a donation: donation refers to the money donated, not the act of donating. (What did she get in the mail? A donation.)
   • have a barbecue: here have has the sense of hold (an organized event). (What did she have at her house? A barbecue.)
   • have a disease/an illness
   • witnessed an operation: the verb and the noun refer to distinct events.

5. NOTE: We exclude the copula from our definition of support, though on rare occasions an idiom lexicalizes a copula: be_the_case.

Following [Calzolari et al. 2002], we distinguish “Type II” support verbs which do contribute some meaning, though it is understood in the context of the event/scenario established by the noun:

• start~ a ~race
  • most aspectual verbs—begin/end/start/stop/continue/finish/repeat/interrupt etc.—would qualify when used with an eventive noun
• pass~ an ~exam
• keep~ a ~promise
• answer~ a ~question
• execute~ a ~program, evaluate~ a ~function

Type II support verbs are lower priority for us than “core” support verbs.

verb-noun idioms

Some verb-object combinations are idiomatic, though they do not qualify as support verb constructions. We count these as multiwords as well.

• pay_attention
• take...time: There are several related idioms involving use of one’s time for some purpose. Include ‘the’ for the “extra effort” sense: take_the_time to help. Include a preposition for took_time_out_of (sacrifice), took_time_out/off (scheduled vacation).
• waste/spend/save/have~time/money
  • CHANGE: was waste_time
• give_ an _estimate, give_ a _quote on something [typically includes the process of estimation as well as offering that estimate to the customer]

Typically, don’t join:

• well done/made/oiled

Exceptions:

• my hamburger is well_done
• that was a_job_well_done
• a well_-_oiled_machine
• he is well_-_read
• well_fed

• a_lot
• N_star
• in_cash
• possibly: get/have_done (hair, etc.)
• highly~recommended, highly~trained
• city, state, etc. locations: _ => ~
• waste_time, spend_time => ~

TODO

• good job, great job, look good
• good job, good work, hard work [I’d be OK with ~ for these but we decided previously that good/great work should be left alone]
• ‘look at it’: include ‘it’? could be specific or not
• Short of that, One more thing -- ?
• fix problem -- I'd say this is a collocation, so fix~problem
• best restaurant out_there
• fast and friendly [sounds slightly better than “friendly and fast”, but that probably reflects a preference for the word with fewer syllables to come first]
• walk_in_the_door: entering a room or establishment
• have/get_ nails/hair/tattoo/etc. _done (grooming)
• ?? have/get done [work/repairs]
• ?? do~work/job (cf. surgery)
• ?? do~dishes