Bridging Text and Knowledge with Frames

Srini Narayanan Google Zurich / Brandschenkestrasse 110, 8002 Zurich, Switzerland [email protected]

Abstract evoke specific frames and establish a binding pat- tern to specific slots or roles (frame elements (FE)) FrameNet is the best currently operational within the frame. FrameNet describes the under- version of Chuck Fillmore’s Frame Se- lying frames for different lexical units, examines mantics. As FrameNet has evolved over sentences related to the frames using a very large the years, we have been building a se- corpus, and records (annotates) the ways in which ries of increasingly ambitious prototype information from the associated frames are ex- systems that exploit FrameNet as a se- pressed in these sentences. The result is a database mantic resource. Results from this work that contains a set of frames (related through hier- point to frames as a natural representation archy and composition), a set of frame elements for applications that require linking textual for each frame, and a set of frame annotated sen- meaning to world knowledge. tences that covers the different patterns of usage 1 Introduction for lexical units in the frame. Collin Baker’s pa- per in this conference has more details on the Frame Semantics (Fillmore, 1976) defines the FrameNet project including the current state of the meaning of a word with respect to the conceptual resource which is now available in multiple lan- structure (Frame) that it evokes. The promise of guages. This paper summarizes the results of ap- Frame Semantics is that it is a principled method plying FrameNet in a variety of NLP applications. to connect language analysis with concepts and knowledge. This paper summarizes over a decade 2.1 FrameNet data as seed patterns for of research at Berkeley1 on computational models Information Extraction bridging text and inference using Frame Seman- While FrameNet frames and FE tags are mean- tics. We will start with a brief description of the ingful to human interpreters, they are not suit- lexical resource, FrameNet2, designed with the ex- able for direct use in NLP applications. One plicit goal to capturing insights and findings from early project explored using the FrameNet anno- Frame Semantics in an on-line lexicon. We then tated dataset to automatically compile patterns and describe computational models that exploit the se- a lexicon for Information Extraction (IE) (Mohit mantic information in FrameNet for a variety of and Narayanan, 2003). A distinguishing feature NLP tasks. that made FrameNet attractive for this purpose was 2 FrameNet its explicit mandate to cover all the valence pat- terns for a target word, not just the frequent ones. The Berkeley FrameNet project (Fillmore, John- Thus, FrameNet annotations and valence alterna- son, & Petruck, 2003) is building a lexicon based tions were designed to capture the long tail for on the theory of Frame Semantics. In FrameNet, every target lexeme. We hypothesized that us- the meanings of lexical items (lexical units (LU)) ing a highly precise set of patterns and a lexicon are defined with respect to larger structured repre- automatically compiled from the FrameNet frame sentations, called Frames. Individual lexical units relational database and annotations should result 1All the work described was done while the author was good performance for the task. To increase cover- at the University of California, Berkeley and the Interna- age, we extended the frame lexicon with WordNet tional Computer Science Institute (ICSI) / 1947 Center Street, Berkeley CA 94704. synsets. As a first test, we culled a set of news 2http://framenet.icsi.berkeley.edu stories from Yahoo News Service with topics re- lated to the topic of crime. We also compiled a effects in perspective and framing choice, it re- set of IE patterns and lexicon from several crime mains a proof of concept demonstration and there related frames (such as Arrest, Detain, Arraign was a need to do an automatic translation to an in- and Verdict.) We were able to achieve an aver- ference formalism which would enable us to use age precision of 76.5% and an average recall to more robust reasoners (the trade-off was of course 66% for the stories in this domain. However, the that these off the shelf reasoners produced shal- relatively sparse and uneven domain coverage of lower inferences). FrameNet and the absence of high quality parsers (Scheffczyk, Baker, & Narayanan, 2010) auto- and named entity annotators (used for building ex- matically translated a crucial portion of FrameNet pressive and general patterns) at the time made the to the description logic based web ontology lan- pilot task difficult to repeat in an open domain set- guage OWL, and showed how state-of-the-art de- ting. While the coverage of FrameNet is still an scription logic reasoners can make inferences over issue, the enormous gains made in the quality and FrameNet-annotated sentences. Thus, annotated amount of parsed and named entity annotated data text becomes available to the Semantic Web and could make this early work attractive again where FrameNet itself can be linked to other ontolo- FrameNet can be used as a high precision seed pat- gies. While our OWL translation is limited to facts tern generator in a semi-supervised IE setting. included in FrameNet, links to ontologies make world knowledge available to reasoning about nat- 3 From Frames to Inference ural language. Therefore, are linked FrameNet to A fundamental aspect of Frame Semantics, one the Suggested Upper Merged Ontology (SUMO). that directly connected the linguistic insights of This ground work gives a clear motivation for the Chuck Fillmore to the early work in AI by Schank, design of further ontology bindings and defines the Abelson, Minsky, and others was the idea that baseline for measuring their benefits. Frames were central to how inferences were pack- Fillmore’s further insights into the connections aged. In this view, framing provided preferential between textual inference and world knowledge access to specific expected inferences. These in- led us to ask the question of how a linguistic ferences were said to be in the frame. Schankian analysis of a written document can contribute to scripts (such as the famous restaurant script) identifying, tracking and populating the eventu- (Schank and Abelson, 1977) are a good example alities that are presented in the document, either of such inferential packaging in terms of expected directly or indirectly, and representing degrees of sequences of events, participants, and outcomes. belief concerning them. This work, reported in In addition to providing such general inferences, (Fillmore, Narayanan, & Baker, 2006), attempts Chuck Fillmore observed that linguistic framing to clarify the boundary between on the one hand also provided a way to delineate multiple perspec- the information that can be derived on the basis tives on an event (including foregrounding, back- of linguistic knowledge alone (composed of lex- grounding, and participant perspective). An exam- ical meanings and the meanings of grammatical ple can be found in the perspective difference pro- constructions) and on the other hand, reasoning vided by the lexical items sell, buy, or pay, which based on beliefs about the source of a document, all evoke the commercial transaction frame. world knowledge, and common sense. In particu- (Chang, Narayanan, & Petruck, 2002), built a lar, we show that the kind of information produced computational formalism that captured structural by FrameNet can have a special role in contribut- frame relationships among participants in a dy- ing to text understanding, starting from the ba- namic scenario. This representation was used to sic facts of the combinatorial properties of frame- describe the internal structure and relationships bearing words (verbs, nouns, adjectives and prepo- between FrameNet frames in terms of parameters sitions) and arriving at the means of recognizing for active event simulations for inference. We ap- the anaphoric properties of specific unexpressed plied our formalism to the commerce domain and event participants. Framenet defines a new layer of showed how it provides a flexible means of han- anaphora resolution and text cohesion based on the dling linguistic perspective and other challenges annotations of the different types of null instantia- of semantic representation. While this work was tions (Definite Null Instantiation (DNI), Indefinite able to computationally model subtle inferential Null Instantiation (INI), and Constructional Null Instantiation (CNI)). A full exploitation of these and meaning, aka a construction. Constructions linguistic signals in a coreference resolver is still exist at lexical (and sub-lexical) levels as well as pending. at larger granularities (phrases, discourse) play- ing a crucial role in the compositionally of lan- 4 Frame Semantics in Question guage. This proposal, entitled construction gram- Answering mar, has gained considerable empirical support in large part due to the investigations of Fillmore, As FrameNet matured, we started asking if it could his colleagues and students, reported in a series of be used for semantically based question answering beautiful papers on the grammatical and composi- for questions that went beyond factoids and re- tional properties of constructions.3 quired deeper semantic information. (Narayanan Research on construction grammar has played and Harabagiu, 2004; Sinha and Narayanan, 2005; a fundamental role within our Berkeley interdis- Sinha, 2008) report on a prototype question an- ciplinary project, NTL4, which is attempting to swering system that attempted to answer questions build cognitively plausible computational mod- related to causality, event structure, and tempo- els of language acquisition and use. Specifically, rality in specific domains. The project on Ques- NTL research has resulted in the grammar for- tion Answering (QA) was a joint effort with Sanda malism called Embodied Construction Grammar Harabagiu’s group at UT Dallas. (ECG), where the meaning pole of a construction The QA work was based on the fact that events, is expressed in terms of bindings between bodily while independent of language themselves, are schemas (also called Image Schemas) and frames. frequently discussed in natural language, yield- ECG allows constraints of all kinds (phonological, ing copious data in that form. To reason about syntactic, semantic, etc.) in a unification based complex events requires an interface from event probabilistic framework, where the best fitting in- models to data sources. We sought to exploit se- terpretation in context is selected as the analysis mantic frames as an intermediate structure and in- of the input. ECG is formally defined and com- terface between event descriptions in natural lan- putationally implemented, and has been used for guage and event models that produce inferences linguistic analysis, in models of language com- to answer questions. In the course of this project, prehension, and for cognitive models of language we came up with the basic framework and algo- learning5. rithms combining a variety of NLP techniques in- cluding Parsing, Topic Modeling, Named Entity 6 Frame Semantics and Metaphor Recognition, and Semantic Role Labeling with deep event structure inference in multiple do- FrameNet has long held the goal of including in- mains. The frame structure in language provides formation about metaphorical usage in language. a bi-directional mapping from language to event The most recent project on Frame Semantics is the models, enabling us to link information found in ICSI MetaNet project, where the goal is to build text about an event of interest to models that repre- a system that extracts linguistic manifestations of sent that event. The proof of concept system used metaphor (words and phrases that are based on frame parsed input with a set of hand built domain metaphor) from text and interprets them automati- ontologies for specific domains. The system was cally in four different languages. able to answer domain questions involving causal, The MetaNet project, is a multi-lingual, multi- diagnostic, and hypothetical reasoning. While the university, multi-disciplinary effort that incorpo- results clearly showed the utility of FrameNet as rates FN methodology as well as corpus and ma- a resource supporting deep semantic inference, it chine learning techniques, deep cognitive linguis- also delineated the necessity and role of domain tics, and behavioral and imaging experiments. specific ontologies and inference required to real- MetaNet models metaphor as a mapping be- ize an end-to-end system using FrameNet. tween two different frames. Such mappings 5 Frames, Constructions and Grammar 3http://www.constructiongrammar.org/ is a currently active resource on the topic with contributions from Yet another of Fillmore’s profound insights was a variety of international scholars. 4http://ntl.icsi.berkeley.edu/ntl the observation that every unit of grammar is most 5http://ecgweb.pbworks.com/w/page/ effectively described as a pairing between form 15044343/FrontPage project information from a source frame to a tar- bridging text, speech, and vision. get frame. The information projected is par- tial and can include the frame, its slots, and Acknowledgments filler constraints. An initial repository of map- All the projects described here were collaborations pings that draws on FrameNet frames as sources with the FrameNet and NTL groups at ICSI and and targets of the mappings is used as base in- Berkeley. This line of inquiry will continue to be formation by a system that extracts additional dedicated to and guided by Chuck Fillmore’s in- metaphors using machine learning. The sys- valuable insights, vision, and body of work. tem uses what it has learned about the relation- ships between the frame elements of conceptual metaphors to find new metaphors in text. The References 6 MetaNet Wiki is a database of such mappings, Nancy Chang, Srini Narayanan, and Miriam R. L. drawing on FrameNet’s inventory of Frames. The Petruck. 2002. Putting Frames in Perspective. Pro- mappings currently exist in four different lan- ceedings of 19th International Conference on Com- guages. FrameNet frames and mappings constrain putational Linguistics. Taipei: COLING, 2002. the search for new metaphors, and the discovery of Charles Fillmore. 1976. Frame semantics and the na- new metaphors by a corpus based machine learn- ture of language. Annals of the New York Academy ing algorithm both a) provides empirical support of Sciences: Conference on the Origin and Develop- ment of Language and Speech, 280, 1,20-32, 1976. for the existing frames and mappings and b) more importantly potentially extends the set by identi- Charles Fillmore, Christopher R. Johnson, and Miriam fying gaps and inconsistencies in coverage. This R. L. Petruck. 2003. Background to FrameNet. In- ternational Journal of Lexicography interaction facilitates an iterative design process . 16.3 (2003): 235-250. in MetaNet which is empirically driven and usage based, just as Fillmore would have insisted.7 Charles Fillmore, Srini Narayanan, Collin Baker. 2006. What Linguistics can contribute to Event 7 Conclusion and Future Directions Extraction. AAAI Workshop on Event Extraction, AAAI 2006, Boston. Frame semantics in general and FrameNet in par- Behrang Mohit and Srini Narayanan. 2003. Semantic ticular show considerable promise for use in deep Extraction from Wide-Coverage Lexical Resources, semantic analysis. FrameNet frames are intended HLT-NAACL, Short Papers, 2003. to capture crucial generalizations not available in Srini Narayanan and Sanda Harabagiu 2004. Ques- other lexical resources. Various prototype sys- tion Answering based on Semantic Structures. In- tems have clearly demonstrated the potential of ternational Conference on Computational Linguis- FrameNet to make a qualitative difference in se- tics (COLING 2004), Geneva, Switzerland, August mantic NLP. There remain two crucial gaps that 22-29, 2004. have to be bridged. One is the issue of cover- Roger Schank and Robert P. Abelson. 1977 Scripts, age. The second is the lack of a formal repre- plans, goals and understanding: An inquiry into hu- sentation covering the more subtle inferential as- man knowledge structures, Erlbaum, 1977. ISBN 0- 470-99033-3. pects of FrameNet. Progress is being made of both fronts as is evidenced in some of the papers in this Jan Scheffczyk, Collin F. Baker, Srini Narayanan. workshop. If successful, the next few years should 2010. Reasoning over Natural Language Text by Means of FrameNet and Ontologies. Ontology and see an increasing use and acceptance of FrameNet the Lexicon: A Natural Language Processing Per- as a crucial semantic resource bridging language spective. Eds. Chu-ren Huang, et al. Studies in Nat- analysis with inference. This will lead to scalable ural Language Processing. Cambridge, UK: Cam- versions of the systems described in this paper, bridge University Press, 2010. 53-71. but will also give rise to new applications. One Steve Sinha and Srini Narayanan. 2005. Model Based particularly intriguing area of research is the use Answer Selection. Textual Inference in Question of frames for cross-modal semantic representation Answering Workshop, AAAI 2005, Pittsburgh, July 2005. 6http://metaphor.icsi.berkeley.edu 7Even at 83, Chuck’s brilliant attention to detail and in- Steven Sinha. 2008. Answering Questions about Com- fectious enthusiasm fundamentally shaped the early MetaNet plex Events PhD Dissertation, Department of EECS, project on a day-to-day basis, till his illness sadly made direct Berkeley, CA, 2008. participation impossible after 2012.