{ "paper_id": "T75-2010", "header": { "generated_with": "S2ORC 1.0.0", "date_generated": "2023-01-19T07:43:15.860233Z" }, "title": "ORGANIZATION AND INFERENCE IN A FRAME-LIKE SYSTEM OF COMMON SENSE KNOWLEDGE", "authors": [ { "first": "Eugene", "middle": [], "last": "Charniak", "suffix": "", "affiliation": {}, "email": "" } ], "year": "", "venue": null, "identifiers": {}, "abstract": "", "pdf_parse": { "paper_id": "T75-2010", "_pdf_hash": "", "abstract": [], "body_text": [ { "text": "Rather than write two short papers for the two sessions I participate in (Memory: Organization, and Memory: Reasoning and Inferencing)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "INTRODUCTION", "sec_num": null }, { "text": "I have chosen to write one paper of twice the length.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "INTRODUCTION", "sec_num": null }, { "text": "My reason for doing so, beyond the fact that it takes less time to write one long paper than two short ones, is that the issues involved are so intertwined as to make separating them an unprofitable task. To reason with this knowledge requires that it be organized, by which I simply mean it must be structured so that the system can get at necessary knowledge when it is needed, but that unnecessary knowledge will not clog the system with the all too familiar \"combinatorial explosion\". I will start with my current thoughts on organization.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "INTRODUCTION", "sec_num": null }, { "text": "The scheme presented here has been clearly inspired by Minsky's \"frames paper\" (Minsky 74) so I have called the primary organizational grouping a \"frame\". It is perhaps indicative of the convergence of ideas reflected in (or perhaps inspired by) Minsky's paper that the overall organization proposed here (although not the details) is quite similar to the independently developed \"scripts\" of (Schank and Abelson 75).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "INTRODUCTION", "sec_num": null }, { "text": "I take a frame to be a static data structure about one stereotyped topic, such as shopping at the supermarket, taking a bath, or piggy banks. Each frame is primarily made up of many statements about the frame topic, called \"frame statements\" (henceforth abbreviated to FS). These statements are expressed in a suitable semantic representation, although I will simply express them in ordinary English in this paper.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "II. FRAMES, FRAME STATEMENTS, AND FRAME IMAGES", "sec_num": null }, { "text": "The primary mechanism of understanding line of a stork is to see it as instantiating one or more FS's. So, for example, a particular FS in the shopping at the supermarket frame would be:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "II. FRAMES, FRAME STATEMENTS, AND FRAME IMAGES", "sec_num": null }, { "text": "(I) SHOPPER obtain use of BASKET (SHOPPER, BASKET, and in general any part of an FS written in all capitals is a variable. These variables must be restricted so that SHOPPER is probably human, and certainly animate, while BASKET should only be bound to baskets, as opposed to, say, pockets.) This FS would be instantiated by the second line of story (2).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "II. FRAMES, FRAME STATEMENTS, AND FRAME IMAGES", "sec_num": null }, { "text": "(2) Jack was going to get some things at the supermarket. The basket he took was the last one left.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "II. FRAMES, FRAME STATEMENTS, AND FRAME IMAGES", "sec_num": null }, { "text": "Here we assume that part. of the second line will be represented by the story statement (SS):", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "II. FRAMES, FRAME STATEMENTS, AND FRAME IMAGES", "sec_num": null }, { "text": "(3) Jackl obtain use of basketl (Of course, really both (I) and (3) would be represented in some more abstract internal representation.) Naturally, (3) would be an instantiation of (I), and this fact would be recorded with a special pointer from 3to (I).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "II. FRAMES, FRAME STATEMENTS, AND FRAME IMAGES", "sec_num": null }, { "text": "I am, of course, making the common distinction between a data base which contains the particular story information, like", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "II. FRAMES, FRAME STATEMENTS, AND FRAME IMAGES", "sec_num": null }, { "text": "(3), and a \"knowledge base\" which contains our generalized real world knowledge, such as the supermarket frame.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "II. FRAMES, FRAME STATEMENTS, AND FRAME IMAGES", "sec_num": null }, { "text": "The supermarket frame will contain other FS's which refer to (I), such as:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "II. FRAMES, FRAME STATEMENTS, AND FRAME IMAGES", "sec_num": null }, { "text": "(4) (I) usually occurs before (5) (5) SHOPPER obtains PURCHASE-ITEMS Any modification (like (4)) of a particular FS (like (I)) will be assumed true of all SS's which instantiate that FS (like (3)), unless there is evidence to the contrary. Hence, using (4) we could conclude that Jack has not yet finished his shopping in (2). Other modifications of (I) would tell us that Jack was probably already in the supermarket when he obtained the basket, and that he got the basket to use during shopping. For example, failure to do so would cause the system to fail to detect the oddness in (6) and (7). (6) Jack went to the supermarket.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "II. FRAMES, FRAME STATEMENTS, AND FRAME IMAGES", "sec_num": null }, { "text": "He got a cart and started up and down the aisles. Bill took the goods to the checkout counter and left. I I I I I I I I I I I I I I I I I (7) Jack went to the supermarket to get a bag of potatoes. After paying for the milk he left.", "cite_spans": [], "ref_spans": [ { "start": 104, "end": 153, "text": "I I I I I I I I I I I I I I I I I", "ref_id": null } ], "eq_spans": [], "section": "II. FRAMES, FRAME STATEMENTS, AND FRAME IMAGES", "sec_num": null }, { "text": "It is probably a bad idea to actually change the frame to keep track of such bindings . Instead I assume that the frame remains pure, and that the variable bindings are recorded in a separate data structure called a \"frame image\" (abbreviated FI) . For frames which describe some action, like our shopping at supermarket frame, we will create a separate FI for each instance~ of someone performing the action. So two different people shopping at the same time, or the same person shopping on two different occasions, would require two FI's to record those particulars which distinguish one instance of supermarket shopping from all others.", "cite_spans": [], "ref_spans": [ { "start": 86, "end": 262, "text": ". Instead I assume that the frame remains pure, and that the variable bindings are recorded in a separate data structure called a \"frame image\" (abbreviated FI)", "ref_id": null } ], "eq_spans": [], "section": "I I", "sec_num": null }, { "text": "Much of this information will be stored in the variable bindings (shopper, purchase items, store, shopping cart used, etc.) However the variable bindings do not exhaust the information we wish to store in the FI, for example it will probably prove necessary to have pointers from the FI to some, if not all, of the SS's which instantiate FS's of the frame in question. Of slightly more interest is that the FI of a frame describing an action will keep track of how far the activity has progressed. So, for example, we would find the following story odd:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "I I", "sec_num": null }, { "text": "(8) Jack drove to the supermarket . He got what he needed, and took it to his car. He then got a shopping cart.", "cite_spans": [], "ref_spans": [ { "start": 34, "end": 50, "text": ". He got what", "ref_id": null } ], "eq_spans": [], "section": "I I", "sec_num": null }, { "text": "We have already said that FS's are modified by time ordering statements, so to note the oddity of (8) it is only necessary to have one or more progress nointers in the FI to the most time-wise advanced FS yet mentioned in the story. Then when new statements are found in the story which instantiate FS's in the frame the program will automatically check to see if these FS's are consistent with the current progress pointer(s).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "I I", "sec_num": null }, { "text": "If It seems to me one of the great advantages of frames that they seem capable of being used in multiple ways, something which is not obviously true, for example, of \"demons\" (Charniak 72).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "I I", "sec_num": null }, { "text": "In this and the next two sections we will take a closer look at the internal structure of a frame.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "III. FRAMES AND SUB-FRAMES", "sec_num": null }, { "text": "In However, not only is the idea unsatisfying, since it would require the duplication of the many facts the two activities have in common, but it would also lead to problems in the comprehension of certain types of stories. For example, one could easily imagine a story which starts out with Jack using a cart, the wheel of the cart sticking, and rather than going to get a second cart Jack finishes his shopping without a cart.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "III. FRAMES AND SUB-FRAMES", "sec_num": null }, { "text": "It seems a priori that such combinations would be extraordinarily hard to account for with two completely separate frames for the two forms of supermarket shopping.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "III. FRAMES AND SUB-FRAMES", "sec_num": null }, { "text": "(Alternatively, it is common practice in crowded situations to park one's cart in a general vicinity of several items and pick up the individual items without the cart, only to bring them all to the cart at a later point and resume shopping with the cart.)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "III. FRAMES AND SUB-FRAMES", "sec_num": null }, { "text": "One possibility for a single frame which handles both kinds of supermarket shopping is:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "III. FRAMES AND SUB-FRAMES", "sec_num": null }, { "text": "(10) a) Goal: SHOPPER owns PURCHASE-ITEMS b) SHOPPER decide if to use a basket.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "III. FRAMES AND SUB-FRAMES", "sec_num": null }, { "text": "If so, set CART to T c) If CART then SHOPPER obtain BASKET d) SHOPPER obtain PURCHASE-ITEMS method 44 e) Do for all ITEM PURCHASE-ITEMS f) SHOPPER decide on next ITEM g) SHOPPER at ITEM h)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "III. FRAMES AND SUB-FRAMES", "sec_num": null }, { "text": "If CART then BASKET also at ITEM i) SHOPPER hold ITEM j)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "III. FRAMES AND SUB-FRAMES", "sec_num": null }, { "text": "If CART then ITEM in BASKET k) End i) SHOPPER at CHECK-OUT COUNTER m) If CART then BASKET at CHECK-OUT COUNTER n) SHOPPER pay for PURCHASE-ITEMS o) SHOPPER leave SUPERMARKET", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "III. FRAMES AND SUB-FRAMES", "sec_num": null }, { "text": "Now one problem with 10 The story does not say that Jack transferred his groceries, and to infer that he did require essentially the same reasoning process required to understand why Jack put the groceries back in the cart in (11).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "I I I I", "sec_num": null }, { "text": "To handle stories like (11) and (12) we must therefore put two pieces of information into (10) which are not there at present. First, that in those cases like lines (-1Oh) and (10m) where we have the basket going along with the person, the reason is to keep the previously collected items with one. Second, that to carry something with a cart requires that it be in the cart. We will indicate the first of these by:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "I I I I", "sec_num": null }, { "text": "(13) SHOPPER at ITEM side condition -DONE at ITEM also Secondly, it assumes the existence of a separate cart-carry frame in which we store information about using carts to carry things.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "I I I I", "sec_num": null }, { "text": "We will see other advantages of this move later, but at the moment we can at least note that if one were to ask \"Why does Jack use a basket?\" two answers (at least) would be possible -\"to do shopping\", or \"to carry his groceries\". (13) allows for both of these answers, whereas (10g) and (1Oh) only allow for the former, since there is no separate \"carry\" level. car. By stating this as suggesting cart-carry (SHOPPER, BASKET, PURCHASE-ITEMS, CAR) we no longer need an instruction to put the groceries into the basket again before setting out.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "I I I I", "sec_num": null }, { "text": "So far then I have argued that frames must be able to reference sub-frames, and in particular, the supermarket frame needs some sub-frame like cart-carry.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "IV. SHARING FRAME STATEMENTS BETWEEN FRAMES", "sec_num": null }, { "text": "There is nothing exceedingly strange in this, but the next step will perhaps be a bit more interesting. Here I will suggest that some of the frame statements in our supermarket frame be shared with the cart-carry frame. To see the reasons for this, let us start by noting that the failure to allow for common FS's will lead to some curious redundancies in our frame.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "IV. SHARING FRAME STATEMENTS BETWEEN FRAMES", "sec_num": null }, { "text": "One of these occurs in the DO loop of (10) which handles the collection of the PURCHASE-ITEMS. With our latest changes, this portion of (10) (lines (e) through (k)) looks like:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "IV. SHARING FRAME STATEMENTS BETWEEN FRAMES", "sec_num": null }, { "text": "(14) a) Do for all ITEM PURCHASE-ITEMS b) SHOPPER choose next ITEM PURCHASE-ITEMS -DONE c) SHOPPER at ITEM d) side-condition DONE at ITEM also e) |method -suggested f) k-~cart-carry (SHOPPER,BASKET,DONE,ITEM) g) SHOPPER hold ITEM h) If CART then ITEM in BASKET i) DONE <-DONE + ITEM j) End", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "IV. SHARING FRAME STATEMENTS BETWEEN FRAMES", "sec_num": null }, { "text": "The redundancy is this: we have already stated that cart-carry has a strict sub-state that the things to be carried must be in the basket. But in (14) we further specify in line (h) that ITEM is to be in BASKET, which is simply a special case.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "IV. SHARING FRAME STATEMENTS BETWEEN FRAMES", "sec_num": null }, { "text": "Once again my initial reason for being concerned with this is that I find such redundancy unappealing, but again there seem to be more solid reasons for doing away with it.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "IV. SHARING FRAME STATEMENTS BETWEEN FRAMES", "sec_num": null }, { "text": "In particular consider the following story: Furthermore, we can now remove the \"If CART then\" portion of (I0c) by assuming the eminently reasonable convention that a shared node in frame-1 which has a pointer to frame-2 is only applicable to the action in rame-1 if frame-2 is activated in the sense that we have created a frame image for frame-2.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "IV. SHARING FRAME STATEMENTS BETWEEN FRAMES", "sec_num": null }, { "text": "When performing the action in real life this means that upon deciding to use a cart one sets up a cart-carry image and this in turn makes the vaious FS's in the supermarket frame dealing with carts relevant to one's activities.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "IV. SHARING FRAME STATEMENTS BETWEEN FRAMES", "sec_num": null }, { "text": "While reading a story the general rule will be that any SS instantiating an FS which is shared with cart-carry will be sufficient to create an FI for cart-carry. place using the time sequence information in the frame plus the progress pointer in the FI. This is, of course, as it should be. One cannot, or at least should not, discuss structure independently of use and the use of frames is to allow us to make inferences about the stories we read. Nevertheless, there remain many issues of inference left untouched by the previous discussion and I will cover one or two of them here.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "IV. SHARING FRAME STATEMENTS BETWEEN FRAMES", "sec_num": null }, { "text": "In Charniak 72I argued that some inferencing had to be done as the story was read (i.e. at \"read time\") rather than after a question was asked (\"question time\"). I think it is fair to say that this is now generally accepted, and the question now is how much inference is done at read time, and of what sort. I will not argue the point here (for a more up-to-date presentation of the issue see (Charniak 75b)), but rather make the rather strong assumption that one has not \"understood\" the text unless one has made a certain, as yet undefined, class of inferences which enable one to tie the text together into a coherent whole.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "IV. SHARING FRAME STATEMENTS BETWEEN FRAMES", "sec_num": null }, { "text": "this assumption we are immediately confronted with a pressing problem.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Given", "sec_num": null }, { "text": "In principle one can draw an infinite number of inferences from a given body of text, and even in practice the total number would be prohibitively large. Some way then is needed to distinguish those inferences which need be made from those which do not.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Given", "sec_num": null }, { "text": "It is my impression that the system outlined previously in this paper has several nice properties in this regard.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Given", "sec_num": null }, { "text": "For one thing, many inferences which would have to be made in a \"demon\" system (Charniak 72) need not be made in the system we have just outlined.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Given", "sec_num": null }, { "text": "To again take the example of Jack getting a shopping cart, I pointed out that the supermarket frame allowed the system to make several inferences about the statement, like why he did it, and that he had yet to start the shopping, but I left it vague as to whether these inferences should actually be made at read time, or only if a qustion was asked. In fact, there seems to be little reason for actually making most of these inferences at read time.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Given", "sec_num": null }, { "text": "Since Jack then went to the store.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Given", "sec_num": null }, { "text": "In (22) we should expect Jack to be buying a kite at the store. But consider the statement (23) Jack1 own kitel On one hand we can only predict that this is likely to occur using information from the birthday frame (perhaps in conjunction with the given subframe). On the other hand, this statement serves as the goal statement in the store frame. To enable (23) to act as the link between frames in this fashion it must appear explicitly in the data base with frame pointers to two diferent frame statements in two different frames.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Given", "sec_num": null }, { "text": "That is to say, (23) appears in two active frames, and hence is a justified instantiation. On one hand this will enable me to explain why I have given up on the latter, while at the same time pointing out what I see as the advantages of the former. For those of you not familiar with (Charniak 72) a paragraph of summary is in order.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Given", "sec_num": null }, { "text": "The point of the demon model was that many lines which had little significance in themselves took on greater significance in context, a prime example being \"There was no sound\" in the context of a child shaking his piggy-bank.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Given", "sec_num": null }, { "text": "To account for such situations th model associated with each \"topic concept\"", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Given", "sec_num": null }, { "text": "(e.g. piggy-bank, or supermarket) a \"base routine\" which was a program which set up \"demons\" which would lie in wait for lines like \"There was no sound\" (which would be the \"pattern\" of the demon). Should the line occur, the demon would in effect say, \"I know what this line means\", and proceed to put statements in the data base which explicated the significance of the line (e.g. a statement like \"This line implies there is nothing in the piggy bank\"). We can draw the parallel between base routines and frames on one hand, and demons and frame statements on the other.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Given", "sec_num": null }, { "text": "I will come back to the significance of this parallel later, for the moment Just take it as a helpful analogy.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Given", "sec_num": null }, { "text": "One problem with the demon model is that in some cases one is forced into ad hoc formulations of facts due to the theoretical machinery seemingly not being suited to the problem. To put this in terms of pointers, the fact in (30) only allows a pointer from \"rain\" to \"umbrella\", it does not allow a pointer from \"umbrella\" to \"rain\" and hence cannot be used to help us conclude in (33) that the problem is rain.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Given", "sec_num": null }, { "text": "It These same things must be checked in a frame, but since the scope of the variable is the entire frame, rather than a single FS, the overhead, so to speak, is shared. Furthermore, the inferences about a given FS are stored implicitly in the structure of the frame, whereas they had to be stated explicitly in the demon. So a second advantage of the frames approach over demons is the conceptual economy one obtains in the expression of facts.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Given", "sec_num": null }, { "text": "The analogy between FS's and demons also points to a third way in which the frames approach seems superior.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Given", "sec_num": null }, { "text": "One problem which bothers many people (including myself) about the demon approach is that it seemingly calls for large numbers of demons to be activated every time a given topic is mentioned in the story, although it is unlikely that more than a small fraction of the demons will ever be used.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Given", "sec_num": null }, { "text": "There are two possible reasons why people feel this is a problem.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Given", "sec_num": null }, { "text": "One is that so many active demons might make it hard to locate those demons which really should apply.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Given", "sec_num": null }, { "text": "Frames do not help with this problem since there will be equal numbers of FS's.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Given", "sec_num": null }, { "text": "To see the second reason why activating large numbers of demons is problematic, note that if it took no time at all to set up a demon, setting up many of them would seem less bad.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Given", "sec_num": null }, { "text": "But of course it does take time to set up a demon, and it becomes a problem to Justify this computation in light of the unlikeliness of the demon ever being used. Frames do offer a potential solution to this second problem because with frames, rather than supermarket activating many demons, we need only create a frame image for one frame (i.e. supermarket).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Given", "sec_num": null }, { "text": "This would take much less time, and hence would be better, but it should be noted that we pay a price. In particular most of the work involved in setting up a demon is to index our storage of active demons so that retrieving the ones needed will be reasonably easy . By comparison looking through frames to find matching FS's promises to be a time consuming task unless we do something similar. This is what I meant earlier when I said that perhaps each frame would have its own index to its contents.", "cite_spans": [], "ref_spans": [ { "start": 265, "end": 413, "text": ". By comparison looking through frames to find matching FS's promises to be a time consuming task unless we do something similar.", "ref_id": null } ], "eq_spans": [], "section": "Given", "sec_num": null }, { "text": "On the other hand, the approach presented here allows one to trade more time for locating an FS in return for less time to set up a new topic (frame), and the spectre of all those never to be used demons makes me inclined to accept this trade.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Given", "sec_num": null }, { "text": "Finally, the frames presented here have no problem handing time relations between FS's as we saw earlier in the paper. The same cannot be said of demons.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Given", "sec_num": null }, { "text": "We saw earlier how we might use a progress pointer to allow the program to notice actions which were out of sequence. information is much more complex than a simple string, or even lattice which indicates the time orderings of actions. For example some time orderings are \"strict\" in the sense that one cannot possibly do things any other way, while others are \"suggested\" in the sense that it is a good idea to do the actions in a given order, but possible to do them some other way, while yet others are \"regulatory\" in the sense that it is possible, but illegal to do the actions in the opposite order (Charniak 75a).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Given", "sec_num": null }, { "text": "In the frames model one can store time ordering statements in the frame along with the rest. ", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Given", "sec_num": null } ], "back_matter": [ { "text": "(32) As Jack was leaving the house he heard on the radio that it might rain. He went to the closet.If asked why he did this we would respond that he was probably getting an umbrella. (It would be also possible to answer that he was getting his raincoat, but this is not important since \"raincoat\" would also have information connecting it to rain.) The extra mechanism which is needed here is the ability to put together the \"expectation\" of getting an umbrella, with our knowledge of where umbrellas are normally kept to conclude that he is going to get his umbrella in spite of the fact that the word \"umbrella\" was never mentioned in (32).The trouble with this solution is that it would not account for the following story:(33) Jack began to worry when he realized that everyone on the street was carrying an umbrella.Question: What was Jack worrying about?Answer: That it might rain, and he was without an umbrella.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "annex", "sec_num": null } ], "bib_entries": {}, "ref_entries": {} } }