Systematic Mismatches Across Annotations

1. Systematic Mismatches Across Annotations Alan Lee and Aravind Joshi Institute for Research in Cognitive Science & Department of Computer and Information Science, University of Pennsylvania ULA Workshop, U of Colorado, Boulder March 2008

2. Preliminaries… • We observe that certain annotated features of the Penn Discourse Treebank 2.0 (PDTB) do not match up neatly with annotations at the syntactic level. • What do certain mismatches suggest for linguistic theory? How do we get from syntax to discourse? • How does this affect NLP applications?

3. Outline • Attribution spans • Parallel Connectives • AltLex • Polarity and Determinacy

4. Outline • Attribution spans • Parallel Connectives • AltLex • Polarity and Determinacy

5. The company says it is talking with several prospects. attribution Attribution Spans • Relation between agents and abstract objects (discourse relations or their arguments) • Annotation: Text Spans and Four features (source, type, polarity, determinacy). More on the features later.

6. S NP VP SBAR-ADV VP There IN S have been no Orders for the Cray-3 NP VP though S the company V it is talking With several prospects says Discourse arguments Syntactic arguments • There have been no orders for the Cray-3 so far, thoughthe company says it is talking with several prospects. • Discourse semantics: contrary-to-expectation relation between “there being no orders for the Cray-3” and “there being a possibility of some prospects”. • Sentence semantics: contrary-to-expectation relation between “there being no orders for the Cray-3” and “the company saying something”.

7. S SBAR-ADV NP-SBJ VP S IN MD VP the application by his RGH Inc. Although NP-SBJ VP could VB NP takeover experts VBD SBAR signal said his interest in helping revive a failed labor- management bid NP-SBJ VP VBD SBAR they Mr. Steinberg will make a bid by himself doubted • Althoughtakeover experts said they doubtedMr. Steinberg will make a bid by himself, the application by his Reliance Group Holdings Inc. could signal his interest in helping revive a failed labor-management bid. • Discourse semantics: contrary-to-expectation relation between “Mr. Steinberg not making a bid by himself” and “the RGH application signaling his bidding interest”. • Sentence semantics: contrary-to-expectation relation between “experts saying something” and “the RGH application signaling Mr. Steinberg’s bidding interest”.

8. Mismatches occur with other relations as well, such as causal relations: • Investors are nervous about the issuebecausethey saythe company's ability to meet debt payments is dependent on too many variables, including the sale of assets and the need to mortgage property to retire some existing debt. • Discourse semantics: causal relation between “investors being nervous” and “problems with the company’s ability to meet debt payments” • Sentence semantics: causal relation between “investors being nervous” and “investors saying something”!

9. How to address mismatch? • One possibility - treat attribution as a different layer of structure in discourse. (and also in syntax?) • This has the effect of reducing the complexity of the discourse structure.

10. Discourse Graphbank (Wolf & Gibson 2005) • Farm prices in October edged up 0.7% from September • as raw milk prices continued their rise, • the Agriculture Department said. • Milk sold to the nation's dairy plants and dealers averaged $14.50 for each hundred pounds, • up 50 cents from September and up $1.50 from October 1988, • the department said.

11. sim elab attr attr 1-2 4-5 1 2 3 4 5 6 ce elab ce - cause/effect; elab - elaboration; sim - similiarity; attr - atribution

12. elab 1-2 4-5 1 2 3,attr 4 5 6,attr ce elab ce - cause/effect; elab - elaboration; [ sim - similiarity; attr - atribution ]

13. Residual issues Does attribution scope over the entire relation, or just Arg1? Even if B.A.T receives approval for the restructuring, the company will remain in play, say shareholders and analysts, thoughthe situation may unfold over the next 12 months, rather than six. Arg1: attributed to shareholders and analysts Rel and Arg2: attributed to Writer Guideline: in case of doubt, attribute to the Writer

14. Residual issues • Attribution cannot always be excluded by default • Advocates said the 90-cent-an-hour rise, to $4.25 an hour by April 1991, is too small for the working poor, whileopponents argued that the increase will still hurt small business and cost many thousands of jobs. What implications does this have for the approach of treating attribution as an independent layer of discourse?

15. Outline • Attribution spans • Parallel Connectives • AltLex • Polarity and Determinacy

16. Parallel Connectives Either he wasn’t being real in the past orhe isn’t being real right now. (1549) You’ve either got a chair oryou don’t. (2428) Ifthe answers to these questions are affirmative, thenthese institutional investors are likely to be favorably disposed toward a specific poison pill. (0275) • Parallel connectives are annotated discontinuously • In the PDTB, both parts of a parallel connective are treated as equally prominent (no hierarchical relationship)

17. In Penn Treebank, the treatment of a parallel connective depends on its position within sentence. When “Either” is sentence-initial, both “either” and “or” are annotated as CC. • Either he wasn’t being real in the past orhe isn’t being real right now. (wsj_1549) S CC S CC S Either he wasn’t being real in the past or he isn’t being real right now

18. This is not possible when “either” is sentence-medial. Here, “either” is treated as an RB and “or” is as a CC. • You’ve either got a chair oryou don’t. (wsj_2428) S S CC S or NP-SBJ VP ADVP VP you don’t RB You ‘ve got a chair either

19. How to represent parallel connective? DL-TAG approach: elementary discourse tree with two lexical anchors (DC = discourse clause) DC DC Either DC because DC DC or DC But question remains: how to transition from syntactic structure to discourse structure?

20. Outline • Attribution spans • Parallel Connectives • AltLex • Polarity and Determinacy

21. Alternative Lexicalization(AltLex) A discourse relation is inferred between two sentences which do not contain an Explicit connective, but insertion of an Implicit connective leads to redundancy. This is because the relation is alternatively lexicalized by some non-connective expression: • Under a post-1987 crash reform, the Chicago Mercantile Exchange wouldn’t permit the December S&P futures to fall further than 12 points for a half hour. AltLex = (consequence)That caused a brief period of panic seeling of stocks on the Big Board.

22. Discourse Connectives and Syntactic Constituency • Most explicit connectives correspond to syntactic constituencies. E.g. (“because” IN, “but” CC, “as a result” PP, etc.) • Some small exceptions with parallel connectives, as we have seen.

23. AltLex expressions often do not correspond to syntactic constituencies. Under a post-1987 crash reform, the Chicago Mercantile Exchange wouldn’t permit the December S&P futures to fall further than 12 points for a half hour. AltLex = (consequence)That caused a brief period of panic selling of stocks on the Big Board. S NP-SBJ VP DT VBD DT PP-LOC That caused a brief period of panic selling…..

24. For a list of AltLex expressions annotated in the PDTB: http://www.seas.upenn.edu/~pdtb/altlex-strings.txt Or search using PDTB Browser (shameless plug) : http://www.seas.upenn.edu/~pdtb/PDTBAPI/pdtbbrowser.jnlp

25. Outline • Attribution spans • Parallel Connectives • AltLex • Polarity and Determinacy

26. Attribution Features Attribution is annotated on relations and arguments, with FOUR Features. Source: encodes the different agents to whom proposition is attributed • Wr: Writer agent • Ot: Other non-writer agent • Arb: Generic/Atbitrary non-writer agent • Inh: Used only for arguments; attribution inherited from relation Type: encodes different types of Abstract Objects • Comm: Verbs of communication • PAtt: Verbs of propositional attitude • Ftv: Factive verbs • Ctrl: Control verbs • Null: Used only for arguments with no explicit attribution

27. Polarity vs Determinacy Polarity: Indicates narrow scope of surface negated attributions. (Neg-raising, Klima 1964). Marked as “Neg” when neg-raising occurs. “Null” otherwise. John doesn’t think the book fell (> John thinks the book didn’t fall) Determinacy: Attributions rendered indeterminate in certain contexts. Marked as “Indet”, or “Null” otherwise. John didn’t say the book fell (> no lowering of negation) Only a certain class of verbs can have negative polarity, i.e. induce neg-raising. Verbs of Propositional Attitude (PAtt) have this behavior, but not others.

28. Polarity vs Determinacy… I don’t believe they have the culture to adequately service high-net-worth individuals. (0927) Discourse semantics: I believe they DO NOT have the culture to adequately service high-net-worth individuals. (0927) Negation of “expect” is lowered onto the argument. The attribution is marked as negative polarity. • Note that the attribution event of “expecting” did occur (is determinate).

29. Polarity vs Determinacy… It didn’t sayif it’s earlier results were influenced significantly by nonrecurring elements.(1711) • Negation of “say” is NOT lowered onto the argument. The attribution is marked as indeterminate. • The attribution event (of “saying”) did not actually occur.

30. At Syntactic Level… At which level should discrepancy in the “polarity” vs “determinacy” type of negation be captured? - In PropBank, negations of attribution verbs are uniformly marked as a negative feature for the adjunct feature “ARGM”. - In TimeML, they contain a polarity feature of “Neg”. I don’tBELIEVEthey have the culture to adequately service high-net-worth individuals. ARG1 I ARG2 they have the culture… ARGM Neg (PropBank) No Neg for lower predicate “have” POLARITY Neg (TimeML) Should the negation be marked as ARGM for the lower predicate (“have”) instead?

31. At Syntactic Level… It didn’t SAYif it’s earlier results were influenced significantly by nonrecurring elements. ARG1 It ARG2 if it’s earlier results were influenced significantly by nonrecurring elements ARGM Neg (PropBank) POLARITY Neg (TimeML) “Saying” event is indeterminate. Does this still count as an event? How to order this temporally?

32. Some questions… • How much of discourse is “projected” from syntax? • Is there a need for a different architecture, different building blocks? • How are these issues manifested cross-linguistically? Currently, discourse annotation work being done for Hindi, Turkish, Czech and Finnish (possibly).