| { |
| "paper_id": "J00-2013", |
| "header": { |
| "generated_with": "S2ORC 1.0.0", |
| "date_generated": "2023-01-19T02:57:35.142531Z" |
| }, |
| "title": "Extended Finite State Models of Language AndrOs Kornai (editor)", |
| "authors": [ |
| { |
| "first": "Ed", |
| "middle": [], |
| "last": "Kaiser", |
| "suffix": "", |
| "affiliation": { |
| "laboratory": "", |
| "institution": "Cambridge University Press", |
| "location": {} |
| }, |
| "email": "" |
| } |
| ], |
| "year": "", |
| "venue": null, |
| "identifiers": {}, |
| "abstract": "In this volume, AndrOs Kornai has put together a collection of articles that strongly argue his contention that finite-state approaches to natural language processing (NLP) are now part of the mainstream, both theoretically and computationally. The papers included were first presented at a 1996 workshop, held in Budapest as part of the European Conference on Artificial Intelligence (ECAI '96), and have been chosen for inclusion in this volume to complement previous works on finite-state approaches. Articles referring to the Xerox regular expression calculus and the AT&T Bell Labs system of weighted finite-state transducers that appeared in Roche and Schabes's 1997 Finite State Language Processing are followed up by new articles in this book. In addition this volume includes a CD-ROM, which, although it does not contain the Xerox and AT&T Bell Labs toolkits, does have source code and executables for several of the systems described in the book, including an older version of Bruce Watson's FIRE Lite toolkit for constructing and minimizing finite automata. The attraction of finite-state approaches to NLP is their speed and efficiency. The question is their adequacy-just how powerful a formal system is needed to describe natural language? Consider the following sentence: A doctor (whom a doctor) m (hired) n hired another nurse.", |
| "pdf_parse": { |
| "paper_id": "J00-2013", |
| "_pdf_hash": "", |
| "abstract": [ |
| { |
| "text": "In this volume, AndrOs Kornai has put together a collection of articles that strongly argue his contention that finite-state approaches to natural language processing (NLP) are now part of the mainstream, both theoretically and computationally. The papers included were first presented at a 1996 workshop, held in Budapest as part of the European Conference on Artificial Intelligence (ECAI '96), and have been chosen for inclusion in this volume to complement previous works on finite-state approaches. Articles referring to the Xerox regular expression calculus and the AT&T Bell Labs system of weighted finite-state transducers that appeared in Roche and Schabes's 1997 Finite State Language Processing are followed up by new articles in this book. In addition this volume includes a CD-ROM, which, although it does not contain the Xerox and AT&T Bell Labs toolkits, does have source code and executables for several of the systems described in the book, including an older version of Bruce Watson's FIRE Lite toolkit for constructing and minimizing finite automata. The attraction of finite-state approaches to NLP is their speed and efficiency. The question is their adequacy-just how powerful a formal system is needed to describe natural language? Consider the following sentence: A doctor (whom a doctor) m (hired) n hired another nurse.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Abstract", |
| "sec_num": null |
| } |
| ], |
| "body_text": [ |
| { |
| "text": "As Gazdar and Mellish point out in their textbook, from which this example is taken, this is a legal English sentence only if m and n are equal (Gazdar and Mellish 1989) . Strings of the form anb n are not regular expressions and therefore cannot be represented by finite-state machines. They require a context-free grammar definition and a machine with an unbounded memory, such as the stack of a pushdown automaton. Natural language constructions can require even more complex forms such as anbnc n, which in turn must be defined using indexed grammars and parsed with a machine that has the equivalent of multiple stacks of memory. However \"suppose that we had access to hardware that would handle FSTNs [finite-state transition networks] ... ultrafast and observed that in actual occurrences of anb n constructions, the value of n never exceeded 3; then we might decide to compile our RTN [recursive transition network] ... descriptions down into FSTNs subject to an n = 3 upper bound (such a compilation is possible for any given finite upper bound on the value of n)\" (Gazdar and Mellish 1989) . Kornai points out that such finite upper bounds are indeed what is observed in parsing natural languages.", |
| "cite_spans": [ |
| { |
| "start": 144, |
| "end": 169, |
| "text": "(Gazdar and Mellish 1989)", |
| "ref_id": null |
| }, |
| { |
| "start": 1074, |
| "end": 1099, |
| "text": "(Gazdar and Mellish 1989)", |
| "ref_id": null |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": "Kornai argues in his introduction that while the surge of thought and development surrounding transformational models in the early 1960s threatened to remove all credibility from finite-state approaches to NLP, the \"extraordinary impact\" of Thomp-son's (1968) grep family of Unix tools lead to the pervasive belief that if you wanted to \"do something with text you needed to build finite automata.\" In this regard, Kornai's own chapter on vectorized finite-state automata describes an extremely efficient pattern-matching engine, around which the NewsMonitor system is built. This system extracts relational information, such as \"who is where\" or \"who bought what\", from issues of the Wall Street Journal (source code and sample data are included on the CD-ROM).", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": "Shortly after the publication of The Sound Pattern of English (Chomsky and Halle 1968) , Kornai points out, \"Johnson (1970) demonstrated that the context-sensitive machinery of SPE ... [could] be replaced by a much simpler one, based on finite-state transducers (FSTs); the same conclusion was reached independently by Kaplan and Kay, whose work remained an underground classic until it was finally published in Kaplan and Kay (1994) .\" These works inspired Koskenniemi's two-level system, and the Xerox rule compiler (Dalrymple et al. 1987) . Both are now dominant tools in the fields of computational phonology and morphology, as exemplified by Tateno et al. (Chapter 6), \"The Japanese lexical transducer based on stem-suffix style forms\" and Kim and Jang (Chapter 7), \"Acquiring rules for reducing morphological ambiguity from POS tagged corpus in Korean.\" The latter includes an algorithm for automatically inferring regular grammar rules for morphological relations directly from part-of-speech tagged corpora.", |
| "cite_spans": [ |
| { |
| "start": 62, |
| "end": 86, |
| "text": "(Chomsky and Halle 1968)", |
| "ref_id": "BIBREF1" |
| }, |
| { |
| "start": 89, |
| "end": 123, |
| "text": "Kornai points out, \"Johnson (1970)", |
| "ref_id": null |
| }, |
| { |
| "start": 185, |
| "end": 192, |
| "text": "[could]", |
| "ref_id": null |
| }, |
| { |
| "start": 412, |
| "end": 433, |
| "text": "Kaplan and Kay (1994)", |
| "ref_id": null |
| }, |
| { |
| "start": 518, |
| "end": 541, |
| "text": "(Dalrymple et al. 1987)", |
| "ref_id": "BIBREF2" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": "Although finite-state approaches to NLP were attempted as early as 1958, Kornai comments that finite-state syntax \"did not really come in from the cold until the nineties.\" Chapter 2 of this work, \"A parser from antiquity: An early application of finite state transducers to natural language parsing,\" by Joshi and Hopely, describes that 1958 parser. In a short commentary on that article, Lauri Karttunen points out that \"many of the currently popular methods for robust parsing are already present, fully articulated:", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": "\u2022 multiword tokens (because of, in front of);", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": "\u2022 tagging words by ambiguity classes (cool VA, [i.e., (V)erb (A)djective class]);", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": "\u2022 rule-based disambiguation;", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": "\u2022 syntactic markup by finite-state transduction;", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": "\u2022 depth-first strategy with backtracking and pushdown store for the analysis of recursive structures;", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": "\u2022 default selection of one structure among alternative analyses with option for later revision.\"", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": "Contemporary systems that incorporate these features in finite-state approaches to parsing and modeling syntax are described in the following papers in the book:", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": "\u2022 Chanod and Tapanainen, Chapter 8, \"Finite state based reductionist parsing for French\";", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": "\u2022 Grefenstette, Chapter 9, \"Light parsing as finite state filtering\";", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": "\u2022 Kornai, Chapter 10, \"Vectorized finite state automata\";", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": "\u2022 Roche, Chapter 11, \"Finite state transducers: Parsing free and frozen sentences\".", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": "The papers by Vilares et al. (Chapter 12), \"Text and speech translation by means of subsequential transducers,\" and Ejerhed (Chapter 13), \"Finite state segmentation of discourse into clauses,\" move the application of finite-state techniques into exciting new areas of machine translation and discourse segmentation. In Chapter 11, Roche expands on the part-of-speech-tagging and parsing articles of Roche and Schabes (1997) . Specifically, he details a finite-state method for syntactically parsing light verbs. Such verbs have complements that allow only a certain degree of variability:", |
| "cite_spans": [ |
| { |
| "start": 399, |
| "end": 423, |
| "text": "Roche and Schabes (1997)", |
| "ref_id": null |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": "(1) 2(3)", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": "John makes concessions to his friend.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": "John makes a right turn.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": ",John makes a right turn to his friend.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": "In this example the predicative noun concessions is the real head of the sentence, governing the number and nature of arguments. The light verb makes is in a support role. Roche's paper shows that such constructions, for which \"rewriting mechanisms such as context free parsing are at best unnatural,\" can be processed by finite-state transducer parsing. Kornai points out that such light-verb constructions were seen in the tradition of Chomsky (1970) as \"core cases of transformational grammar.\" Roche's article argues persuasively that finite-state models are not just \"an efficient but somewhat inaccurate tool,\" as might be imagined from a transformational grammar perspective, \"but rather one of the best formalisms at hand to represent accurately complex linguistic phenomena.\"", |
| "cite_spans": [ |
| { |
| "start": 438, |
| "end": 452, |
| "text": "Chomsky (1970)", |
| "ref_id": "BIBREF0" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": "Kornai suggests also that another \"important way in which mainstream syntax is impacted by finite-state techniques can be called 'finite-state to the rescue'. The paper by Schulz and Mikolajewski [Chapter 14, \"Between finite state and Prolog: Constraintbased automata for efficient recognition of phrases\"] ... describes how constraintbased grammars can be speeded up by finite-state methods, and the paper by Srinivas [Chapter 15, \"Explanation-based learning and finite state transducers: Applications to parsing lexicalized tree adjoining grammars\"] ... shows how corpus-based acquisition of LTAGs is facilitated by finite-state techniques.\" Speech recognition is dominated by statistical techniques, in particular HMMs (which are finite-state mechanisms) and statistical n-gram language models. In regards to syntax parsing, Kornai points out that \"an important step in bringing rule-based and statistical work closer is the framework of weighted finite-state transducers developed at AT&T Bell Labs, represented in this volume by the Tzoukermann and Radev paper\" (Chapter 16, \"Use of weighted finite state transducers in part of speech tagging\"). This article describes a process of classifying words into \"genotypes\" through the use of weighted, negative-constraint transducers. All members of a particular genotype share the same set of possible part-of-speech tags. After tagging, genotype n-gram models are built. Tzoukermann and Radev state that their system \"correctly disambiguates 96% of words in unrestricted texts.\"", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": "Since the main algorithms and model-building techniques associated with the Xerox and AT&T Bell Labs approaches to finite-state NLP are well described elsewhere, this book examines ways of extending the scope of finite-state machinery. For example, the \"more complex formal systems discussed by Csuhaj-Varjfl [Chapter 17, \"Colonies: Nederhof and Bertsch's article defines a new subclass of context-free languages that combines deterministic languages (also known as the LR(k) languages) within the sequential framework of a finite-state automaton. The result is a class of languages that can be recognized and parsed in linear time, which also allows for deterministic center-self-embedding (i.e., rule descriptions that define strings of the form anb n for unbounded values of n). This approach is in some respects similar to lexicalized tree adjoining grammars (Srinivas, Chapter 15), which also have a regular level (i.e., initial trees) that contain a context-free lower-level (i.e., auxiliary trees--for center-selfembedding). The notion of combining regular grammars with context-free or restricted context-free grammars (like the LR(k) grammars) to create new subclasses of grammars and languages that lend themselves to treatment by finite-state approaches is an exciting area of research that is well represented in this book.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": "Overall, 1 believe that this book will be a valued addition to the library of anyone interested in emerging finite-state approaches to NLP.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| } |
| ], |
| "back_matter": [ |
| { |
| "text": "Ed Kaiser is interested in integrating structural language models directly into the speech recognition process. He has developed a robust, finite-state parser (inspired by CMU's Phoenix system) as a first step towards that integration, and published two papers on the compilation techniques involved. Kaiser's address is Computer Science and Engineering Department, Oregon Graduate Institute, P.O. Box 91000, Portland, OR, 97291-1000; e-mail: kaiser@cse.ogi.edu", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "annex", |
| "sec_num": null |
| } |
| ], |
| "bib_entries": { |
| "BIBREF0": { |
| "ref_id": "b0", |
| "title": "Remarks on nominalization", |
| "authors": [ |
| { |
| "first": "Noam", |
| "middle": [], |
| "last": "Chomsky", |
| "suffix": "" |
| } |
| ], |
| "year": 1970, |
| "venue": "Readings in English Transformational Grammar", |
| "volume": "", |
| "issue": "", |
| "pages": "184--221", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Chomsky, Noam. 1970. Remarks on nominalization. In R. Jacobs and P. Rosenbaum, editors, Readings in English Transformational Grammar. Blaisdell, Waltham, MA, pages 184--221.", |
| "links": null |
| }, |
| "BIBREF1": { |
| "ref_id": "b1", |
| "title": "The Sound Pattern of English", |
| "authors": [ |
| { |
| "first": "Noam", |
| "middle": [], |
| "last": "Chomsky", |
| "suffix": "" |
| }, |
| { |
| "first": "Morris", |
| "middle": [], |
| "last": "Halle", |
| "suffix": "" |
| } |
| ], |
| "year": 1968, |
| "venue": "", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Chomsky, Noam and Morris Halle. 1968. The Sound Pattern of English. Harper and Row, New York.", |
| "links": null |
| }, |
| "BIBREF2": { |
| "ref_id": "b2", |
| "title": "Tools for morphological analysis. Center for the Study of Language and Information Report", |
| "authors": [ |
| { |
| "first": "Mary", |
| "middle": [], |
| "last": "Dalrymple", |
| "suffix": "" |
| }, |
| { |
| "first": "Ronald", |
| "middle": [ |
| "M" |
| ], |
| "last": "Kaplan", |
| "suffix": "" |
| }, |
| { |
| "first": "Lauri", |
| "middle": [], |
| "last": "Karttunen", |
| "suffix": "" |
| }, |
| { |
| "first": "Kimmo", |
| "middle": [], |
| "last": "Koskenniemi", |
| "suffix": "" |
| }, |
| { |
| "first": "Sami", |
| "middle": [], |
| "last": "Shaio", |
| "suffix": "" |
| }, |
| { |
| "first": "Michael", |
| "middle": [], |
| "last": "Wescoat", |
| "suffix": "" |
| } |
| ], |
| "year": 1987, |
| "venue": "", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Dalrymple, Mary, Ronald M. Kaplan, Lauri Karttunen, Kimmo Koskenniemi, Sami Shaio, and Michael Wescoat. 1987. Tools for morphological analysis. Center for the Study of Language and Information Report CSLI-87-108. CSLI, Stanford, CA.", |
| "links": null |
| } |
| }, |
| "ref_entries": { |
| "FIGREF0": { |
| "uris": null, |
| "type_str": "figure", |
| "num": null, |
| "text": "A multi-agent approach to language generation\"] .... Nederhof and Bertsch [Chapter 18, \"An innovative finite state concept for recognition and parsing of context free languages\"] .... and Ristad [Chapter 19, \"Hidden Markov models with finite state supervision\"] ... are likely to provide a fertile ground for further experimentation with extended finite state models of language.\"" |
| } |
| } |
| } |
| } |
