31. Speech acts

To say something is to do something: to assert, ask, request, promise, thank, warn. This chapter describes how English marks what a speaker is doing — explicit performative verbs, sentence types, modals, please, tags and punctuation — how often form and act part company in Mova's data, how indirect speech acts such as Can you open the window? are recognised, and how dialogue-act schemes (DAMSL, SWBD-DAMSL, ISO 24617-2) classify acts in conversation. It ends with what Mova's labeller, agreer, frames and dialogue layer do with acts.

Contents

1. Austin and Searle in brief
2. Explicit performatives in English
3. Illocutionary force indicators
4. Sentence types and acts in the data
5. Indirect speech acts
6. Recruitments: requests in interaction
7. Dialogue act schemes
8. Speech acts inside narrative
9. How Mova labels acts
10. Not yet

1. Austin and Searle in brief

Austin (1962, How to Do Things with Words) observed that some utterances do not describe but perform: I name this ship…, I promise to come. He split an utterance into three acts — the locutionary act of saying something meaningful, the illocutionary act performed in saying it (a warning, a promise), and the perlocutionary act performed by saying it (frightening, persuading) — and described the felicity conditions an act needs to succeed. Searle (1969, Speech Acts; 1975, "Indirect speech acts"; 1976, "A classification of illocutionary acts", Language in Society 5) made the conditions systematic and proposed five classes of illocutionary acts. None of these works is openly licensed; they are listed under Wanted in Sources: English syntax and pragmatics. Open accounts: Green, "Speech Acts", Stanford Encyclopedia of Philosophy (plato.stanford.edu/entries/speech-acts), and the background chapter of Floyd, Rossi & Enfield (eds.) (2020, Getting others to do things: A pragmatic typology of recruitments, Language Science Press, ch. 1, books/e/en/enfield-2020-getting-others-things-pragmatic), which summarises Searle's conditions for a request — the utterance refers to a future act of the hearer, and the speaker believes the hearer can do it, among others (citing Searle 1969: 66).

Searle's five classes, with Mova's Act values (the closed enum in prag/src/schema.rs):

Mova's set splits Searle's classes along lines that matter for reading tales and lessons: narrate (telling what happened in a story) is kept apart from assert (stating a general fact), evaluate (praise or blame) from both, and ask (a request for information) from request (a request for action). Warn is a directive in Searle's terms but is kept as its own act because its consequences differ: a warning predicts harm. Declarations are rare in the data and fall under other.

2. Explicit performatives in English

2.1 Form

An explicit performative names its own act with a verb in the first person, simple present, active, indicative, often with hereby: I promise…, I thank you…, I beg you…, I apologise…, I warn you…. The hereby test separates performative uses (I hereby promise) from descriptions (*I hereby promised, *She hereby promises).

Performative verbs take the content of the act as a clausal complement (ccomp: I promise that I will come; xcomp: I promise to come) or as an object; the syntax is that of chapter 23. The performative verb is the root, but the act is carried by the whole construction: the frame reader makes such a clause a Say frame with the act of its verb (§8).

2.2 Hedged and embedded performatives

English often embeds the performative under a modal or a verb of wanting, which softens it and makes it formally indirect: I would like to thank…, I must ask you…, There is one which I should like to present to your majesty (book:pg128:487, labelled offer, indirect), I should like to do something to show how grateful I am (book:pg3027:1170, labelled thank, indirect). These are conventional politeness forms (32).

2.3 Performative verbs in the data

Mova maps verbs of speaking to acts directly when they introduce speech (act_of in prag/src/frames.rs):

In addition, all verbs whose VerbNet classes are all communication classes (37.x: whisper, mumble, inquire, complain, gossip…, 152 verbs) count as verbs of speech (Pragmatics: frames — who did what to whom, why, and what came next, Speech).

3. Illocutionary force indicators

Searle called the forms that signal force "illocutionary force indicating devices". In written English they are:

Mova's agreer takes the sentence form from the tree with these devices (rule P0, tree_form in prag/src/agree.rs): a final ? gives q, or wh when the first word is a wh-word (Penn tags WDT, WP, WP$, WRB); a root base-form verb (VB) with no subject and no auxiliary before it, opening the sentence after at most two softeners (please, just, never, always, do, don't, not, now, then, so, and, but, first, also), gives imp; a final ! gives excl; anything else is decl. For direct speech the form of the speech itself is used.

4. Sentence types and acts in the data

4.1 The default and the rest

Each sentence type has a default act: declaratives assert, interrogatives ask, imperatives request. The parse database shows how often the default holds (counts of prag.tsv, 311 101 rows, 2026-10-17; the full cross-table is in 30 §7.3):

Yes/no questions are the least faithful to their default: more than a third of them do something other than ask.

4.2 Examples by cell

All from the tale part of the parse database (public-domain texts); the act is the agreed label, ind. marks indirect=1.

Several rows show grammar that later chapters take up: the conditional fragment If you will hold the horse used as a request (an insubordinate if-clause, 24); the rhetorical questions Can you doubt any longer? and But who knows that? as assertions (26, 34); the negative concord of I won't never do so no mo' in dialect (27, 37).

5. Indirect speech acts

5.1 What makes an act indirect

An act is indirect when it is performed by means of another: the literal act (a question about ability) is not the point, and the hearer infers the intended act (a request). In Mova's schema, indirect=1 means "the form is not the usual one for the act" (Smart annotator — layers, context, memory). In the parse database 3 536 rows (1.1%) are indirect (2026-10-17); the most frequent cells:

The q / ask and wh / ask cells are indirect questions in a different sense: a question that asks something other than what it literally asks (Do you know it is a whole year's journey from here? — book:pg128:671, which asks whether the hearer realises the distance, not whether he knows it).

5.2 Conventional indirect requests in English

Requests in English are mostly made indirectly, by questioning or asserting one of the conditions a request needs. Gordon & Lakoff (1971, "Conversational postulates", CLS 7 — Wanted) and Searle (1975) observed that a speaker can request by asking about the hearer's ability or willingness, or by stating the speaker's want:

The forms are conventional: Can you…? works as a request while its paraphrase Are you able to…? does not, and please fits Can you please…? but not the paraphrase. Hinkelman & Allen (1989, "Two Constraints on Speech Act Ambiguity", aclanthology.org/P89-1026, hinkelman-1989-two-constraints-speech-act-ambiguity) argue that such conventions must be combined with plan-based reasoning, because pure idiom lists ignore context and pure inference cannot explain the conventions.

5.3 Computational accounts

The 44% human agreement figure is from the summary of Sarathy et al. in Pragmatics: algorithms from the literature, for Mova (2026-10-03).

5.4 Mova's indirect requests

Mova reasons about indirect requests in its dialogue layer (world/src/talk.rs, indirect_request), following Brown's conditions as used by Sarathy et al.:

1. The sentence starts with can/could/would/will you (please) or would you mind; the rest is the action.
2. Wanted: a modal question (or please) shows the speaker wants it.
3. Capable: the action is one the system can do (a fixed list of commands).
4. Willing: the action is not blocked right now (state-changing actions wait while training runs).
5. Obligated: the speaker is one the system works for.

If all hold, the question is read as a request and the reply says why ("you want it, I can do it, I am willing now, and I work for you"); if the action is not one it can do, the direct reading wins and the reply answers the ability question, listing what it can do. Tests: tests03::indirect_requests_need_their_preconditions (2026-10-03, Pragmatics: algorithms from the literature, for Mova (2026-10-03)). This layer works on cue words in the sentence, not on the tree; the sentence labeller (§9) learns indirect from silver labels and has no precondition reasoning.

6. Recruitments: requests in interaction

Conversation analysis treats requests as one way among many to recruit others: Floyd, Rossi & Enfield (eds.) (2020) define a recruitment sequence as one where a person gets another to do something practical, verbally or not, and compare eight languages with one coding scheme; they find cross-linguistic diversity in this domain "relatively low" (ch. 1, books/e/en/enfield-2020-getting-others-things-pragmatic). In the English chapter, Kendrick (2020, "Recruitment in English: A quantitative study", ch. 4 of the same volume) finds that most English recruitment sequences in his video data are minimal — a single recruiting move — in 65.9% of cases (n = 139), organised as adjacency pairs where an imperative (move over a little) or an interrogative request makes compliance the expected response.

The point for a text reader is that a request is a sequence, not a sentence: it may be pre-empted, granted by an action, or refused with a reason. The tales show this: "Can you not protect me, then?" inquired the sister (book:pg11027:1158) is followed by the brother's response, and the frame reader links the two through Say frames and next (38).

7. Dialogue act schemes

7.1 DAMSL and SWBD-DAMSL

Dialogue-act schemes classify utterances in conversation by their function, including functions Searle's classes do not cover: backchannels (uh-huh), acknowledgements, answers, turn management. SWBD-DAMSL adapted the DAMSL scheme to the Switchboard telephone corpus: Stolcke et al. (2000, "Dialogue act modeling for automatic tagging and recognition of conversational speech", Computational Linguistics 26(3), aclanthology.org/J00-3003 our card, stolcke-2000-dialogue-act-modeling-automatic-tagging) annotated 1 155 conversations (about 198 000 utterances) with 42 labels; the distribution is very skewed — statements about 36%, backchannel/acknowledge 19%, opinion 13%. Their HMM tagger reached 71% accuracy on reference transcripts and 65% on recognised speech with prosody, against a 35% baseline and 84% human agreement. Jurafsky et al. (1998, "Lexical, Prosodic, and Syntactic Cues for Dialog Acts", aclanthology.org/W98-0319, jurafsky-1998-lexical-prosodic-syntactic-cues-dialog) show how short feedback words (yeah, uh-huh, okay) split into continuers, agreements, yes-answers and assessments by their words, prosody and the grammar of assessments.

7.2 ISO 24617-2

The ISO standard for dialogue act annotation (Bunt et al. 2010, aclanthology.org/L10-1385, bunt-2010-iso-standard-dialogue-act-annotation; revised in Bunt, Petukhova & Fang 2017, aclanthology.org/W17-7404 our card; second edition Bunt et al. 2020, aclanthology.org/2020.lrec-1.69 our card, bunt-2020-iso-standard-dialogue-act-annotation) is multidimensional: one segment can carry several functions in different dimensions. Summary after Pragmatics: algorithms from the literature, for Mova (2026-10-03):

Qualifiers add conditionality (Can you tell me… as a conditional request), certainty and sentiment; an answer depends on its question, and feedback on a segment.

7.3 Mova's acts against the schemes

Mova does not annotate feedback, turn or time management, because its texts are written narratives and lessons, where those functions are mostly absent; the dialogue layer handles greetings, thanks, apologies and goodbyes as social obligations (38).

8. Speech acts inside narrative

In tales most acts are performed by characters inside quotations. Mova's frame reader builds a Say frame for a verb of saying or thinking and puts what was said into its quote (prag/src/frames.rs; Pragmatics: frames — who did what to whom, why, and what came next):

• the act of a Say frame comes from its verb (§2.3);
• for the plain verbs of saying (say, answer, reply, add, call, call out, whisper, mutter, remark) the act is refined by what is said: a quote with ? makes it ask, a quote with ! makes it express (switch act-by-quote);
• say + sorry becomes apologize (switch say-sorry);
• the speaker after a quote ("…" cries the Badger), the person told or asked as the recipient, and inside the quote I = the speaker, you = the addressee;
• a quote with no verb of saying is given to the last speaker of the discourse.

(1) world frames on the whole sentence of Aesop's Fables book:pg11339:1145 (Mova 0.4, CPU, 2026-10-17; the frames of the narration before the quote abridged):

By and by he began to revive, and, feeling very hungry, called out to a passing Sheep and said,
"Would you kindly bring me some water from the stream close by?"

#3 Say(assert) _ call out to=sheep
     roles by vn manner_speaking-37.3 «NP V PP.recipient»: agent=Agent to sheep=Recipient
#5 Say(ask) _ say | says→#6 cause→#0
     roles by vn say-37.7-1-1 «NP V that S»: agent=Agent
#6 Fact(does) you bring patient=water to=i place=stream attrs=kindly | quoted
     roles by vn bring-11.3-1 «NP V NP NP»: agent=Agent i=Destination water=Theme

The Say frame of said gets the literal act ask from the ? of its quote, and call out to a passing Sheep makes the sheep a recipient of speech. Three things are missing or wrong, and they are typical: the speaker is unresolved (_, the wolf of the earlier sentences is out of the one-sentence context given here), the sheep is not carried over as the addressee of said, and the quoted request is a general Fact(does) rather than an event the hearer is asked to perform. The intended act (request) is in the sentence label (q / request / indirect=1), not in the frame. Bringing the two together is listed under Not yet.

Weakest on the frame bench are exactly these speech parts: speaker F1 .32 and says F1 .38 on dev (2026-10-10), because dialogue often has no verb of saying and the parser splits quotes (Pragmatics: frames — who did what to whom, why, and what came next, Not yet).

9. How Mova labels acts

The act field is labelled per sentence by the pipeline of 30 §8: teacher silver → three transparent classifiers on the same features → the agreer → the tiered teacher. Specific to acts:

• Features (prag/src/feats.rs): the first words and the last word of the utterance, end punctuation, ? and !, length; from the tree the root's UPOS, form and lemma, tense, the first auxiliary, inversion, a wh-word, a modal, negation, the subject and its person, you, first person, a vocative; the act of the previous sentence and whether the sentence is direct speech.
• Agreer rules (prag/src/agree.rs): P1 an imperative whose act is narrate/assert/evaluate/other becomes request (or warn when both NLI witnesses say warn); P2 outside the books course, narrate becomes assert when both witnesses agree — the labeller learned on tales and called lesson facts narration.
• Teacher changes (teacher column, 2026-10-17): the act was changed in 53 023 of 311 101 rows (17.0%). The new values, most often: assert 23 371, narrate 13 316, warn 3 474, other 3 399, evaluate 1 910, request 1 844, suggest 1 387, express 1 171, ask 1 051. The assert/narrate border (a general statement vs telling what happened) accounts for most changes; among the rarer acts, warnings are the ones the labeller misses most.

10. Not yet

• Indirectness is labelled, not inferred. The precondition reasoning of §5.4 runs in the dialogue layer only, on cue words; the sentence labeller predicts indirect from features.
• Frames keep the literal act. A Say frame of "Would you kindly bring me some water?" is ask; the request it performs is not propagated into the frame.
• No multidimensional acts. One act per sentence; ISO 24617-2's feedback, turn and time dimensions are not annotated.
• Declarations have no act of their own (other).
• Prosody is absent: rising declaratives and intonation-marked acts are invisible in text.
• Evaluation is weakly checked: the agreer's evaluation rule (P3) was switched off after the witnesses agreed on wrong answers (prag/src/agree.rs).

In Mova

• The act enum and labels: prag/src/schema.rs; features: prag/src/feats.rs; classifiers: prag/src/model.rs; agreer (P0 form from the tree, P1, P2): prag/src/agree.rs; silver: prag/src/silver.rs.
• Speech in frames (act_of, Say frames, act by the quote's punctuation): prag/src/frames.rs.
• Dialogue acts, social obligations and indirect requests with preconditions: world/src/talk.rs.
• Docs: Smart annotator — layers, context, memory (Pragmatics: what was meant), Pragmatics: algorithms from the literature, for Mova (2026-10-03) (Speech acts and dialogue acts), Pragmatics: frames — who did what to whom, why, and what came next.

Sources

• Austin, J. L. 1962. How to Do Things with Words. Oxford: Clarendon. Not open — Wanted.
• Searle, J. R. 1969. Speech Acts. 1975. Indirect speech acts. 1976. A classification of illocutionary acts, Language in Society 5. Not open — Wanted.
• Gordon, D., Lakoff, G. 1971. Conversational postulates. CLS 7. Not open — Wanted.
• Green, M. "Speech Acts". Stanford Encyclopedia of Philosophy. plato.stanford.edu/entries/speech-acts
• Floyd, S., Rossi, G., Enfield, N. J. (eds.) 2020. Getting others to do things: A pragmatic typology of recruitments. Language Science Press. Ch. 1 (introduction), ch. 4 (Kendrick, K. H., Recruitment in English: A quantitative study). books/e/en/enfield-2020-getting-others-things-pragmatic
• Allen, J. F. 1979. Plans, Inference, and Indirect Speech Acts. ACL. aclanthology.org/P79-1021 our card; allen-1979-plans-inference-indirect-speech-acts
• Perrault, C. R., Allen, J. F. 1980. A Plan-Based Analysis of Indirect Speech Acts. AJCL 6(3–4). aclanthology.org/J80-3003; perrault-1980-plan-based-analysis-indirect-speech
• Brown, G. P. 1980. Characterizing Indirect Speech Acts. AJCL 6(3–4). aclanthology.org/J80-3002; brown-1980-characterizing-indirect-speech-acts
• Hinkelman, E. A., Allen, J. F. 1989. Two Constraints on Speech Act Ambiguity. ACL. aclanthology.org/P89-1026; hinkelman-1989-two-constraints-speech-act-ambiguity
• Sarathy, V., Tsuetaki, A., Roque, A., Scheutz, M. 2020. Reasoning Requirements for Indirect Speech Act Interpretation. COLING. aclanthology.org/2020.coling-main.433 our card; sarathy-2020-reasoning-requirements-indirect-speech-act
• Roque, A., Tsuetaki, A., Sarathy, V., Scheutz, M. 2020. Developing a Corpus of Indirect Speech Act Schemas. LREC. aclanthology.org/2020.lrec-1.28 our card; roque-2020-developing-corpus-indirect-speech-act
• Louis, A., Roth, D., Radlinski, F. 2020. "I'd rather just go to bed": Understanding Indirect Answers. EMNLP. aclanthology.org/2020.emnlp-main.601 our card; louis-2020-i-d-rather-just-go
• Stolcke, A. et al. 2000. Dialogue act modeling for automatic tagging and recognition of conversational speech. Computational Linguistics 26(3). aclanthology.org/J00-3003 our card; stolcke-2000-dialogue-act-modeling-automatic-tagging
• Jurafsky, D., Shriberg, E., Fox, B., Curl, T. 1998. Lexical, Prosodic, and Syntactic Cues for Dialog Acts. aclanthology.org/W98-0319; jurafsky-1998-lexical-prosodic-syntactic-cues-dialog
• Bunt, H. et al. 2010. Towards an ISO Standard for Dialogue Act Annotation. LREC. aclanthology.org/L10-1385; bunt-2010-iso-standard-dialogue-act-annotation
• Bunt, H., Petukhova, V., Fang, A. C. 2017. Revisiting the ISO standard for dialogue act annotation. ISA-13. aclanthology.org/W17-7404 our card; bunt-2017-revisiting-iso-standard-dialogue-act
• Bunt, H. et al. 2020. The ISO Standard for Dialogue Act Annotation, Second Edition. LREC. aclanthology.org/2020.lrec-1.69 our card; bunt-2020-iso-standard-dialogue-act-annotation
• Mova's parse database prag.tsv and trees, counted 2026-10-17; tales identified by Project Gutenberg number (titles from the tales registry).
• Repo docs listed under In Mova.

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/languages/en/31-speech-acts/