Frame Semantics for Text Understanding

Frame Semantics for Text Understanding

Frame Semantics for Text Understanding Charles J. Fillmore and Collin F. Baker International Computer Science Institute 1947 Center St.,Suite 600 Berk...

185KB Sizes 0 Downloads 0 Views

Recommend Documents

Using Frame Semantics for Knowledge Extraction from - CiteSeerX
Kenneth Anger. Rita Hayworth. Clonmacnoise. Skogskyrkogården NSA. General Motors. Man Ray. Elizabeth Taylor. Pompeii. C

Annotating Text to Deepen Understanding
Annotating. Text to Deepen. Understanding. Doug Fisher and. Nancy Frey. Page 2. Annotationis a noteof any form made whil

Kernel Sparse Feature Selection Based on Semantics in Text
training and 2428 documents for test. Sogou is provided for classifying Chinese by Sogou Lab. We select a subset which c

A Compositional Semantics for Turkish Correlatives - Cascadilla
In this paper, I propose a compositional semantics for the construction in Turkish that Iatridou (2013) identifies as th

A-Frame ?
called an A-Frame because it looks like the English let- ter "A". The A-Frame is ... of land is needed. The courtyard of

Text characters for cars
Oct 2, 2016 - This is a list of characters from the Pixar franchise Cars : the 2006 film ... List of The Incredibles cha

May 13, 2017 - Stephen Ilhardt. Audrey Iskandar. Dakota Jackson. Yasheen Jadidi ... Lauren Neja. Xiniu Nie. Kyuhyun Oh.

Reading for Understanding
Treaty of Paris (1763) treaty that ended the war between France and Britain. Proclamation of 1763 British declaration th

Prototype Semantics or Feature Semantics - an alternative? - CiteSeerX
le , spire, turret, or pinnacle must be used. As opposed to Leisi's ..... Chicago. Fillmore, C. 1975. "An alternative to

Logical Consequence: Semantics
Logical Consequence: Semantics. 1 Truth Assignments ..... Similar remarks hold about reading the material biconditional

Frame Semantics for Text Understanding Charles J. Fillmore and Collin F. Baker International Computer Science Institute 1947 Center St.,Suite 600 Berkeley, California, 94704 f llmore,[email protected]


An introduction to knowledge representation using Frame Semantics, as is being carried out in the FrameNet Project. A short news article is analyzed, providing examples of many of the questions being dealt with and the proposed solutions, including semantic composition, text coherence, polysemy and WSD, and evidentiality.

1 Introduction

The FrameNet(FN) research project (Lowe et al., 1997; Baker et al., 1998; Fillmore et al., 2001) is building a lexical resource that aims to provide, for a signi cant portion of the vocabulary of contemporary English, a body of semantically and syntactically annotated sentences from which reliable information can be reported on the valences or combinatorial possibilities of each item targeted for analysis. Key aspects to the work of the project are a commitment to a descriptive framework based on semantic frames containing frame elements (semantic roles)1, and a commitment to documenting its observations on the basis of carefully annotated attestations taken from large electronic corpora.2 ~The authors are Principal Investigator and Project Manager, respectively, of the National Science Foundation sponsored project, FrameNet++. We are grateful to the National Science Foundation for funding the work of the project through two grants, IRI #9618838 \Tools for Lexicon Building" March 1997{February 2000, and ITR/HCI #0086132 \FrameNet++: An On-Line Lexical Semantic Resource and its Application to Speech and Language Technology" September 2000{August 2003. The other Principal Investigators of FrameNet++ are Dan Jurafsky (University of Colorado at Boulder), Srini Narayanan (SRI International), and Mark Gawron (San Diego State University). 1 Semantic roles have played an important role in NLP for many years, from (Simmons, 1973) and (Schank, 1972), to (Rilo and Schmelzenbach, 1998), and in studies of human sentence processing, e.g. (Trueswell et al., 1994). 2 For the rst part of the project, the British National Corpus was used, courtesy of Oxford University Press; for continuing work, and especially for tasks of the kinds considered in this paper, FN are depending on both the BNC and the corpora of English news texts provided by the LDC (North American News Text Corpus, the Supplement, and AP Worldstream English). The project uses the Corpus Workbench 

The basic FrameNet data are stored in a MySQL database a portion of which is shown in Fig. 1. Most signi cant for our purposes are the tables showing the relation between lemmata and frames (polysemy is a one-many relation between lemmata and the frames that express their meanings), and tables showing the relations between frames. Frame-toframe relations include (1) composition, by which a complex frame is shown to be decomposable as a temporal structure - often a structured procedural sequence of simpler frames - and (2) inheritance, by which a single frame can be seen as an elaboration of one or more other frames, with bindings between the elements of co-inherited frames. Lexical entries, including valence descriptions which summarize the attested combinability possibilities, are generated as reports derived algorithmically from the database. One of the means chosen for demonstrating the relevance of the database to NLP research is a planned pilot e ort at bringing FN data to bear on information extraction from newspaper accounts of crimes, criminal behavior, and instances of (lowlevel) criminal justice procedures. This commits us to selecting terminology that occurs frequently in such reports for detailed analysis touching on all aspects of FN research. Since FN is a lexicographic project, our concern in its application to research on text understanding has to be limited to its potential service in other sorts of activities. We have to take for granted the existence of independent resources providing syntactic parsing (including mini-grammars for dates, addresses, proper names, titles, institutions, etc., such as FASTUS (Appelt et al., 1993)), as well as anaphora resolution, real-world connections, and discourse coherence. It should be possible for FN data to be called on for assistance with word-sense disambiguation, semantic composition (the integration of information associated with semantic depensoftware from Institut fur Maschinelle Sprachverarbeitung of the University of Stuttgart for searching the corpora and selecting sentences for annotation. In the rst phase of the project, the Alembic Workbench from MITRE was used for annotation; for the second phase, FN sta have written a custom Java front-end to the MySQL database.

Figure 1: Entities and Relations for FN2 (Partial) dents into frames evoked by their semantic governors), valence-justi ed choices among competing parses, and activation of topic-related vocabulary (through shared frame membership) for recognition and sense-selection in successive parts of a text. There are also preliminary studies (Gildea and Jurafsky, 2000) which suggest that the annotation data from FN can also be used to seed automatic recognition of frame elements, both to assist the manual annotation process and to generate a much larger body of annotated sentences.

2 An Example Text

The text we wish to examine here is the following, taken from, dated 14 February 2001. In this text, as is common in journalistic writing, each paragraph contains one sentence. (Sentence numbering is ours.) 1. Washington (CNN)| Alleged White House gunman Robert Pickett was arraigned Wednesday at a federal court in Washington and ordered held without bond. 2. A federal magistrate informed Pickett of the charges against him - assaulting a federal ocer with a deadly weapon, which carries a maximum of ten years in prison. 3. The magistrate set a preliminary hearing for next Tuesday and ordered Pickett held without bond.

Index 8 7 6 2 1 3 9 10 13 12 11 14 15 4 5 16 17 18

Dependents none none 7,8 [6] 2,3,4,5 (2)9,10 none 11 none none 12,13,14 15 none none (2)16 (2)17 18 none

Form Alleged White House gunman Robert Pickett was arraigned Wednesday at a federal court in Washington and ordered held without bond.

POS adj n n n be ppt adv prp art adj n prp n cnj ppt ppt prp n

Table 1: Dependency Representation of Sentence 1 4. Pickett, who was shot in the knee by the Secret Service after allegedly ring two shots outside the White House, used crutches to walk into the court. 5. He did not enter a plea. If we assume a simple dependency parse, a rst pass in nding the ways in which semantically dependent elements are tted into the frame struc-

tures evoked by the key words in the text is to trace the connections between governors and their dependents, recognizing the di erence between syntactic heads and semantic heads when necessary. A sample of the kind of analysis we will need, for the rst sentence, is given in Table 1. The highest syntactic governor is the passive auxiliary was , whose dependents arraigned and ordered are the main frame-bearers in the sentence. The dependents of arraigned in this sentence give us the identity of the defendant (index 2), and the time and space coordinates of the arraignment (indices 9 and 10); the agencies behind the arraignment are unexpressed. The court's action (ordering Picket held without bond) is expressed by ordered (index 5) and its dependents. Syntactic interpretations associated with VP conjunction will make clear the role of the defendant in the ordering frame. Before exploring the possible contributions of FN data, we can make some initial observations about the text as a whole. We notice rst of all that the initial sentence informs us of an arraignment. From the meaning of this word we know that relevant activities within such a process involve informing the accused of the charges against him, setting a date for preliminary hearing, and making decisions to guarantee the defendant's future court appearances. Details of the time of the arraignment and the scheduled time of the preliminary hearing need to be calculated by using the date of the news report (Wednesday , next Tuesday ), together with the knowledge that a simple occurrence of Wednesday with past-time reference is understood as the most recent Wednesday, and next Tuesday is the Tuesday of the calendar week following the date of the report. Anaphoric links must be established between Robert Pickett (sentence 1), Pickett (sentences 2, 3, 4), and he (sentence 5) in this text, and between a federal magistrate (sentence 2) and the magistrate (sentence 3). The lexical items from the article for which FN needs to have analyses are alleged , gunman , arraign , Wednesday , federal , court , order , hold , bond , magistrate , inform , charge , assault , ocer , deadly , weapon , carry , maximum , year , prison , set , preliminary , hearing , next , Tuesday , shoot , knee , allegedly , re , shot , outside , used , crutches , walk , court , enter , and plea . A general way of applying FN valence information to the analysis of the words in a text is to choose a word (starting from the highest semanticallyrelevant predicate in a given sentence), determining the frames that it is capable of evoking, noticing the semantic roles of the props and participants in each such frame, trying to match the semantic needs associated with each such frame (and hence with each sense of the word) with phrases found in the sentence

at hand, choosing the one which makes the most coherent t, and entering the semantic structures associated with the dependent constituents into slots provided by the selected frame. Some of the structures in the text are nonproblematic. In Sentence 2, the verb inform evokes a frame of one person providing another person with some information, and the syntactic valence possibilities include the pattern by which the communicator is expressed as the verb's subject, a phrase designating the addressee of the informing act follows the verb and the phrase indicating the transmitted message is introduced with an of -PP. An understanding system that seeks coherence between portions of a text will have noted that the rst sentence evokes the Arraignment frame and that one of the functions of a court appearance for an arraignment is for the judge to inform the accused of the nature of the charges against him; to the extent that the evoked arraignment structure is retained for predicting sentence-to-sentence connections, it seems clear that the mention of a federal magistrate , inform , and charges against can be quickly incorporated into the growing semantic representation of the text. In particular, the meaning selected for magistrate (as `judicial ocer' rather than as, say, `the chief magistrate of the nation') and the meaning chosen for charges (as `accusation' as opposed to `agreements on delayed payment for merchandise'), will be determined by a coherence preference. It is not the work of FrameNet to line up such expectations, but it should be a service of the FN data to o er descriptions of the related frames and semantic connections among the words (magistrate as a synonym of judge , charge as a synonym of accusation , etc.) to facilitate the establishment of such coherence judgments. It should be noted that frame structures needed for establishing text coherence clearly go beyond mere argument structures. For the Arraignment frame, for example, the \slots" needing to be lled are not generally going to be available in the same sentence, let alone among the syntactic dependents of the verb arraign .

3 Multi-word Units and Parsing

The analysis cannot simply proceed on the basis of frame information built on the text's words taken one at a time. Many word sequences must be identi ed as xed phrases, the most obvious ones including White House , the White House , Robert Pickett , and the Secret Service . Tight collocations must be recognized for the following phrases: held without bond , assaulting a federal ocer with a deadly weapon , preliminary hearing , ring shots , and enter a plea . All of these phrases are themselves analyzable from their parts, but held without bond is one of the standard

phrases for reporting one of the court decisions in an arraignment hearing, assaulting a federal ocer with a deadly weapon is a named o ense in American law, preliminary hearing is a named step in the criminal justice process, and re and enter can be analyzed as standard support verbs for shot and plea respectively. All such information must take the form of separately listed lexical entries. Recognizing the ad hoc occasion-speci c compound White House gunman and its reference will have to depend on the understander's being in touch with current news. Sentence 3 speaks of set(ting) a preliminary hearing . Either the collocation between set and hearing or that between preliminary and hearing (or both) must be established in the lexicon or the understander must depend on knowledge of the steps in an arraignment hearing to choose among the possible senses of the highly polysemous word set. Since the thesaurus-like character of FrameNet is provided by linking words to the frames they evoke, all these words belong to a single (high-level) frame. (The FN lexicon will have to indicate that the valence possibilities for set in the Appointment frame include the phrasings set a date for the hearing , set the hearing for February 20 , and set a hearing date .) The text o ers a few analytic challenges. One of these is in the the appositional relation between the charges against him (plural) and assaulting a federal ocer (singular), and the association with the singular-agreement form carries , which violates some rather basic rules of English syntax. Another is the elliptical expression carries a maximum ten years in prison where the singular article a has to be connected with the grammatically anomalous maximum ten years (which appears to be plural). This could be seen as an ellipsis of a maximum sentence of ten years . And recognizing the correct interpretation of carries may not be straightforward, since its subject is not speci cally identi ed as a crime and its object is not speci cally identi ed as a punishment. The selectional needs for the correct sense of carry include a subject which names a crime and an object which can be construed as a penalty. Thus, the antecedent of which in this sentence, the assaulting phrase, has to be recognized as the name of a crime: the defendant is charged with committing a crime, and it is the crime which carries the sentence. The parsers we have tested seem to have diculty with was ordered held without bond , in which both the verb order and the secondary predicate that represents content of the order are in passive voice. The corresponding active matrix verbs do not strike one as rare (The judge ordered the defendant shot at dawn , I ordered the package redelivered ). It is interesting that of the more than 1,300 passive instances of the verb order in the British National Corpus, only two were found that have the syntactic struc-

ture we see here, and one of them is: He was ordered held for a bail hearing on Tuesday . We suspect that this family of phrases should be treated as exhibiting a xed pattern related to the criminal process frame.

4 Evidentiality

An important representation problem presents itself with the words alleged and allegedly . When the adjective alleged precedes a nominal constituent it is associated with the category signi ed by the nominal, i.e. the NP represents something that has been claimed to be an instance of the category, and this is not a problem. And when the adverb allegedly belongs to the highest predicate in an event predication, it means that the event in question is a report and protects the author from being accused of making the claim himself. But when the adverb is embedded in a fact-reporting phrase that identi es a sub-event, as in who was shot in the knee after allegedly ring two shots outside the White House , there arises a representational problem involving adverb scope. The shot in the defendant's knee did not occur after someone alleged that he red shots outside the White House. We wish to represent the kind of meaning shown here as evidential, by which the author of a report is adding some sort of evidential quali cation to a part of the description. The frame in question, Evidentiality, is a metalinguistic frame that has as frame elements the Ascriber (of an evidential label) and the Description (of an event or a category). The word list for this frame includes alleged , allegedly , reported , reportedly , known , certi ed certi ably , authentic , suspected , self-described , admitted , and perhaps a few others.

5 Slot-Fillers

Most of the work in FrameNet to date has focussed on the verbs (and some nouns and adjectives) which we call frame-bearing or frame-evoking|those lexical heads which evoke a frame whose frame elements are typically expressed by the dependent NPs, PPs, VPs and Ss. But now the FN database also has the means of targeting frame-relevant dependent nouns for a separate kind of annotation. This would include the names of artifacts when they appear in sentences exhibiting information about the functions for which they have been manufactured. In our text we have examples of instruments in a with -phrase, as in assaulting with a deadly weapon , and as the object of the verb use , as in use crutches to walk into the court . The FN project intends to provide \reverse" information about these lexical items, characterizing the slots they ll in the frames in which they most naturally occur and identifying the manner of their syntactic marking.

The FN treatment of weapons and weapon use will include subframes involving the discharging of rearms, where distinctions for direct object roles include the weapon, the projectile and the target (shoot a gun , shoot a bullet , shoot a person , cf. shoot at a person ); other verbs that speci cally deal with using a rearm are discharge and re ( re a gun , * re a person , re a bullet ). (In the case of re a shot we can treat the verb as a support verb for the event noun shot .)

6 Conclusion

Table 2 shows a summary of some of the frames discussed in this paper and the relations among them. In addition to inheritance (elaboration) relations among frames, we de ne corresponding relations between elements of frames which are so related; some of these are shown by means of equals signs in the \Elements" column. For example, Courtdate-setting elaborates Appointment, which has a subframe Action. The Judge of Court-datesetting is the Prot1 of Appointment, the Defense and Prosecution (collectively) are the Prot2 of Appointment, the Action of Appointment is constrained to be one of the subframes of the Criminal Process frame, and the location of this action is a courtroom. These frame descriptions are still very preliminary, and have been greatly compressed to t into a readable chart. The 16 frames shown here represent the minimum needed to understand the criminal proceedings described in this news story, including a few high-level, abstract frames such as Action and Event. Although we are just beginning work in this content domain, we estimate that less than 200 frames will suce to represent most of the semantics of the vast majority of newspaper crime stories, covering hundreds of frame-evoking lexical units and their valences. (We are not attempting to represent expert knowledge of law, fraudulent accounting, stock market manipulation, etc.) The FrameNet project is not itself dedicated to NLP e orts as such, but but we hope the information it makes available to the research community is of the sort which suitably ts any of several kinds of NLP activities. Evidence on the syntactic and collocational environments of polysemous words in di erent senses should be an aid to Word Sense Disambiguation. The association of lemmata with the frames they evoke, and hence with other lemmata belonging to the same frame, should aid in topic recognition and hence coherence establishment. The provision of a large body of semantically annotated sentences (that is, annotated with respect to single key words within them) should amount to a training corpus for automatic semantic tagging. In the extent to which the FN database is capable of exhibiting all of the major valence possibilities for each

sense of each word, it should provide material for enhanced statistical surveys of word sense frequencies for polysemous words, and preferred subcategorization frames for given word senses.


Douglas E. Appelt, Jerry R. Hobbs, John Bear, David Israel, and Mabry Tyson. 1993. Fastus: A nite-state processor for information extraction from real-world text. In Proceedings. IJCAI-93, Chambery, France. Collin F. Baker, Charles J. Fillmore, and John B. Lowe. 1998. The Berkeley FrameNet Project. In COLING-ACL '98: Proceedings of the Conference, held at the University of Montreal, pages 86{90. Association for Computational Linguistics. Charles J. Fillmore, Charles Wooters, and Collin F. Baker. 2001. Building a large lexical databank which provides deep semantics. In Benjamin Tsou and Olivia Kwong, editors, Proceedings of the 15th Paci c Asia Conference on Language, Information and Computation, Hong Kong. Daniel Gildea and Daniel Jurafsky. 2000. Automatic labeling of semantic roles. In ACL 2000: Proceedings of ACL 2000, Hong Kong. Association for Computational Linguistics. John B. Lowe, Collin F. Baker, and Charles J. Fillmore. 1997. A frame-semantic approach to semantic annotation. In Marc Light, editor, Tagging Text with Lexical Semantics: Why, What and How? Special Interest Group on the Lexicon, Association for Computational Linguistics. Ellen Rilo and Mark Schmelzenbach. 1998. An empirical approach to conceptual case frame acquisition. In Proceedings of the Sixth Workshop on Very Large Corpora, pages 49{56, Montreal, Canada. Roger C. Schank. 1972. Conceptual dependency: A theory of natural language processing. Cognitive Psychology, 3:552{631. Robert F. Simmons. 1973. Semantic networks: Their computation and use for understanding English sentences. In Roger C. Schank and Kenneth Mark Colby, editors, Computer Models of Thought and Language, pages 61{113. W.H. Freeman and Co., San Francisco. John C. Trueswell, Michael K. Tanenhaus, and Susan M. Garnsey. 1994. Semantic in uences on parsing: Use of of thematic role information in syntactic ambiguity resolution. Journal of Memory and Language, 33:285{318.




Criminal Process Arraignment Con nement Pre-trial Con nement

Con nement

Appointment Court-date-setting


Event Action




Action on Bodily target




Crime-law Assault Assault with deadly weapon Shooting

Crimeagainstpeople Action on bodily target Assault Action


Judge(s) Ocer(s) Courtroom Court Prosecution (group) Prosecutor Defense (group) Defendant(s) Defense Attorney(s) Charges Law Defendant Court Jailer Prisoner Bail Prot1 Prot2 Judge = Prot1 Defense & Prosecution = Prot2 Action = subframe of criminal process Action.Place = courtroom Theme (\a ected object") (Cause) (Result) Place Time Actor (Means) (Manner) Actor.type = Sentient Prot1 = Actor Prot2 Prot2 body part = Theme Perpetrator = Actor Victim = Theme Means/Weapon = Means De nition Penalty Jurisdiction Assailant = Perpetrator Intention = \bodily harm" (Threat) = (of Action) (Action) = Crime-against-people.Action = Action-on-Bodily-Target Weapon.type = \deadly" Shooter = Actor Gun = Means Projectile = Theme (Point of impact)

Sub-Frames Court Appearance Arrest Accusation Arraignment Preliminary hearing Trial Verdict Stating of Charges Entering of Plea Con nement Posting of bail Release on bail Return to court Flight Promise Action

Action = Event Forming of Intention Action = Action Result = Result Malice = Intention Action = \harming victim"

Table 2: Summary of Frames and Relations