THE PROBLEM OF UNREALIZED COMPLEMENTS AND ITS RELATION TO FRAMES AND SCRIPTS ============================================================================ Working Note 29, Peter Clark, Boeing Research and Technology (2008) peter.e.clark@boeing.com 1. The Problem -------------- Complements and adjuncts are often unrealized in language, e.g.,: (1) A man sold a book [to someone] [for a price] (2) A man sold a book for $10 [to someone] (3) A woman received a book [from someone] (4) Proposals [for funds] [from people/institutions] [to a funding agency] must be 15 pages. (5) Members [of an organization] objected to the proposal. (6) There was violence [by someone] [to someone/something] in Budapest last night Their absence causes problems for a simple deductive rule engine trying to apply rules, as the rules' conditions may not be satisfied by the explicitly stated information, e.g., the below rules are (undesirably) not triggered by the above sentences: (r1) IF X sells Y to Z THEN Z buys Y from X (r2) IF X sells Y to Z for Cash THEN Z gives Cash to X (r3) IF X receives Y from Z THEN Z gives Y to X (r4) IF X is violent to Y THEN X damages Y As a result, entailments may be missed, e.g.: T1: A book was sold. H1: A book was bought / received / paid for. In the above example, although DIRT contains rules suggesting T1 -> H1, these rules don't fire when applied in the normal deductive fashion as their conditions are not explicitly satisfied, and hence the desired implications are missed by the system. This phenomenon has also been referred to as "implicit role reference" in the literature (see Appendix). 2. Solutions ------------ (a) Unrealized complements -------------------------- If the rule's condition "X sells Y" does not fire on text "Y was sold", the system is essentially assuming that: if no seller (X) is mentioned, then we cannot be sure that there is a seller. However, in language, often unstated complements DO exist in the real world, implied by other elements (e.g., the verb) of the sentence. To handle this, one can relax the inference engine's assumption in the rule application, so a rule triple (X r Y), e.g., ("sell" agent "person"), will match any text containing X', e.g., "sell", providing there is no clashing text triple (X' r Y'), e.g., ("sell" agent "company"). Now rule condition testing is starting to look a little more like matching. We could take this one step further and also allow rule conditions like "X sells Y for Money" to match text "X sells Y", thus abducing the existence of Money. The existence of the "for" complement may be a certain thing, or just a probable thing, depending on whether the "for" complement is always or usually present in selling (if it's only sometimes present, it's really an adjunct, not a complement). Similarly whether that complement is of type Money may be certain or just probable, depending on the semantics of selling; this is real world knowledge, not syntactic knowledge. It is knowledge of the expectations that the frame (verb) should invoke. Alternatively, a simple way of filling in complements would be to use a resource like FrameNet (say) to find the complements of a verb and their types, and assert them. For example, FrameNet tells us that "sell" has a buyer, seller, goods, and money (although doesn't specify their types). We can match the frame with the text and fill in the complements. In fact, this is really very similar to the matching of a rule's condition described above, the difference being that FrameNet frames are geared towards describing the complements (players, Frame Elements) of a verb without further implication, while rules are primarily geared towards implication (i.e., saying what facts follow from the given elements). Of course, we cannot universally assume that an unmentioned preposition or complement exists. Consider the rule: IF a Covering is over a Thing THEN the Covering covers the Thing This rule should *not* fire on a covering (e.g., a blanket), abducing that the covering is implicitly over something. Similarly: IF a Person walks under a Thing THEN the Thing is over the Person we shouldn't abduce from the existence of a Person walking (or just a Person) that the Person is under something. The bottom line is that we need to distinguish essential arguments (complements) from optional arguments (adjuncts), and only allow abduction on the essential arguments. (A trivial rule, implemented in our software, is that only subjects and objects can be abduced). To make matters more complicated, it may be that the complement/adjunct distinction depends on the sense of the verb in play, e.g., the syntactic object of break#v1 as in "the engine broke" is an adjunct (optional), while the object of break #v2 as in "the engine broke the axle" is a complement (obligatory). We earlier said if the abduced complement "clashes" with what is known, then the complement should not be abduced. But the notion of "clash" is tricky, as often multiple objects can be in the same semantic relation with a verb. Consider Rule: "A bandage covers a wound" Text: "A bandage covers stitches" Although the objects of "cover" are distinct, they do not clash as more than one thing can be covered; in this case it is perfectly sensible for a bandage to cover both a wound AND stitches, and for the existence of the wound to be abduced. (b) Coordinated completion of unrealized complements ---------------------------------------------------- Often a verb's complements can be filled in in multiple, coordinated ways, eg, people are often violent to people, storms are often violent to property, etc. Thus sometimes if we know one complement, it might help us determine the likely types of other complements. If, as in FrameNet, we only have a single frame encompassing all uses of the verb sense, then it may have such general type restrictions on its roles (frame elements) as to be largely uninformative. Instead, it may be better to have a collection of different, alternative specialized uses of a verb sense. These essenntially correspond to having multiple, distinct FrameNet-style frames for a given verb sense. For example, consider again the earlier sentence: (5) Proposals for funds [to a funding agency] must be 15 pages. Here the complement "for funds" to "proposal" suggests the recipient ("to") of the proposals is a funding agency. We'd like one instantiation of the "proposal" frame to include the complements "funds" and "funding agency" (and maybe "institution wanting funds"). Alternatively, if we had a sentence about a "proposal for marriage", we would expect the "to" complement to be different. [Aside: note here we're working with noun modifiers, rather than verb complements, but the problem is essentially the same]. Similarly if the algorithm suggests a set of complements, but one clashes with a complement that *is* given, then this suggests the whole set may be inapplicable. There should be a coordinated application of the whole set. (c) Anaphoric unrealized complements ------------------------------------ The appropriate complement may not be attached to the verb in the sentence, but still be mentioned earlier in the text. For example: (7) Members of the Redding Fire Department brought their ladder truck to campus and raised the 45-foot ladder. Students took turns climbing to the top [of the ladder]. (Source: news article) The complement of "climb" is the earlier-mentioned "ladder". We'd like a system to realize this, not just introduce a Thing (the "climbee") as the object complement of "climb". The fact the complement is realized in a different sentence adds another level of complication; we cannot just match frames/rule conditions with text on a sentence by sentence basis, but need to take into account the whole paragraph. Similarly, consider: (8) Jeff decided to go surfing. There were sightings of Great Whites off Newport, but Jeff wasn't concerned [about himself being eaten by the Great Whites.] (Source: Lange and Wharton, 1999). (8) is particularly interesting, because the unrealized complement (namely the proposition "Great Whites eat Jeff") isn't mentioned in the text, but an element of that proposition ("Great Whites") is. In fact, the process of filling in complements with more specific values (e.g., "funds"-"funding agency") again looks a lot like matching scripts with text, where the "scripts" are a single event frame. If there are multiple words with missing complements, then each can help disambiguate the other as to the correct complements to fill in. 3. From Unrealized Complements to Scripts ----------------------------------------- 3.1 Scripts ----------- It is not a huge leap to go from the unrealized complement problem to the whole script-matching problem. Consider my two favorite script examples: T2: The bomb attack destroyed the shrine H2: The bomb exploded T3: A soldier was killed in a gun battle H3: The soldier was shot In T2 there is something akin to a "frame" or "complements" associated with "bomb" in which it explodes and destroys things, and matching this against T2 will suggest H2. Strictly, this frame is not attached to/part of the semantic knowledge about just "bomb"; rather it's an independent structure associated with several terms including "bomb" and "destroy". Similarly in T3, there are expectations associated with "kill" and "gun" (and "soldier") in which someone is shot, which should apply in this case. The bottom line is that it's all really matching against expectations at various levels of granularity. 3.2 The Matching and Knowledge Problems --------------------------------------- A concept in context often suggests other concepts in relation to it (e.g., each complement), even if those other concepts are not stated in text. Matching consists of hypothesizing these additional concepts. This process may be as simple as a word suggesting a single triple (e.g., "sell" -> "sell to person"), or as complex as a whole structure suggesting a full-blown script. Implications, both small and large, will reinforce or contradict each other in various ways. We might say that there are two key problems: - The MATCHING PROBLEM is to find the most coherent set of implications from a given set of assertions (e.g., stated in text). - The KNOWLEDGE PROBLEM is to construct a database of plausible implications in the first place, including a degree of confidence on what constitutes evidence to trigger them, and a degree of confidence in their various implications. Tuples seem like they should play a role in solving the knowledge problem. Similarly the DIRT paraphrase database seems like a possible knowledge source for this. It might be one can construct larger tuple-structure expectations, but with lower confidence -- even spanning multiple sentences, perhaps. It is not clear whether such expectations should be encoded as rules or scripts. The good thing about rules is that there are well-defined mechanisms for using them. The problem with rules is that they impose a somewhat artificial directionality on knowledge. For example, for "bomb" we might need rules expressing different permutations such as (informally) bomb & destroy -> explode bomb & explode -> destroy explode & destroy -> bomb These are all different ways of saying "some evidence should suggest the whole lot". On the other hand, matching against the script "bomb & destroy & explode" is poorly specified: What degree of match is needed to consider the script applicable? Clearly some features are more significant than others, and some are essential; some weighting scheme would be needed. REFERENCES ========== T. Lange and C. Wharton (1999). "Retrieval from Episodic Memory by Inferencing and Disambiguation", in Understanding Language Understanding, pp 107-180, Ed. A. Ram and K. Moorman. MA: MIT Press. POSTSCRIPT 1/2/15 ================= I later discovered this issue has also been explored by Joel Tetreault (Univ Rochester) under the title of "Implicit Role Reference" ([1,2]). Here, he treats these as the complement of indirect anaphora: Indirect anaphora: Explicit reference back to an implicit entity. Implicit Role Reference: Implicit reference back to an explicit entity. e.g., (1) Take engine E1 from Avon to Dansville (2a) Pick up the boxcar and take it to Broxburn [from ?] here, the (implicit) "from" of "take" = (previously mentioned) Dansville. The computational task is to "resolve" this implicit reference. In Joel's statistics, about half (= a lot!) of a verb's semantic roles are implicit and need to be filled in. The fillers are typically mentioned earlier. He did specific experiments with missing to/from fillers in transportation events (TRAINS domain). His solution was to search backwards in the text, rather than do a world simulation (<- I'd have preferred this approach). (Papers and PPT available at http://www.cs.rochester.edu/~tetreaul/academic.html) [1] Joel R. Tetreault. Tense and Implicit Role Reference. Annotation Standards for Temporal Information in Natural Language, Workshop in LREC 2002 Las Palmas de Gran Canaria, May 27, 2002, p.61-64. [2] Joel R. Tetreault. Implicit Role Reference. 2002 International Symposium on Reference Resolution for Natural Language Processing. Alicante, Spain, June 3 - 4, 2002, p.109-115.