Introduction to NLP Tools 09/23/2003 1 Motivation • Machine Translation – From English to French • What’s needed? 2 Motivation Cont’d (1) • Syntactic parser • Part-Of-Speech Tagger – Example: NP -> adj noun • Morphological Analyzer – Example: “tools” -> “tool” “Who is he?” -> “Who is he ?” • Semantic Analyzer – Word sense disambiguate (“wash dishes”) – Choose the correct translation 3 Motivation Cont’d (2) • Lexicons – The information of the word How many senses? What’s the possible translations of the word? • Corpus – Useful for learning a tool – Useful for evaluation 4 Outline • • • • • • Lexicons Text corpora Morphological tools Part-Of-Speech(POS) taggers Syntactic parsers Semantic knowledge bases and semantic parser • Speech tools 5 Lexicons • Definition – A repository for words • Lexicons in LDC(Linguistic Data Consortium) – creating and sharing linguistic resources: data, tools and standards. • CELEX • WordNet 6 CELEX • • • • Dutch Center for Lexical Information Lexical databases of English , Dutch and German 21,000 nouns, 8,000 adjectives and 6,000 verbs English: – – – – – – – – – – – English Orthography, Lemmas English Phonology, Lemmas English Morphology, Lemmas English Syntax, Lemmas English Frequency, Lemmas English Orthography, Wordforms English Phonology, Wordforms English Morphology, Wordforms English Frequency, Wordforms English Corpus Types English Frequency, Syllables 7 WordNet • A database of lexical relations • Inspired by current psycholinguistic theories of human lexical memory • Synset: a set of synonyms, representing one underlying lexical concept – Example: • fool {chump, fish, fool, gull, mark, patsy, fall guy, sucker, schlemiel, shlemiel, soft touch, mug} • Relations link the synsets: hypernym, HasMember, Member-Of, Antonym, etc. 8 WordNet Cont’d • Example pu-erh.cs.utexas.edu$ wn bike -partn Part Meronyms of noun bike 2 senses of bike Sense 1 motorcycle, bike HAS PART: mudguard, splashguard Sense 2 bicycle, bike, wheel HAS PART: bicycle seat, saddle HAS PART: bicycle wheel HAS PART: chain HAS PART: coaster brake HAS PART: handlebar HAS PART: mudguard, splashguard HAS PART: pedal, treadle, foot lever HAS PART: sprocket, sprocket wheel • Example Pu-erh.cs.utexas.edu$wn bike Information available for noun bike -hypen Hypernyms -hypon, -treen Hyponyms & Hyponym Tree -synsn Synonyms (ordered by frequency) -partn Has Part Meronyms -meron All Meronyms -famln Familiarity & Polysemy Count -coorn Coordinate Sisters -simsn Synonyms (grouped by similarity of meaning) -hmern Hierarchical Meronyms -grepn List of Compound Words -over Overview of Senses Information available for verb bike -hypev Hypernyms -hypov, -treev Hyponyms & Hyponym Tree -synsv Synonyms (ordered by frequency) -famlv Familiarity & Polysemy Count -framv Verb Frames -simsv Synonyms (grouped by similarity of meaning) -grepv List of Compound Words -over Overview of Senses 9 Corpus • Definition – Collections of text and speech • • • • LDC Penn Treebank DSO Hansard 10 Some of the Top Corpus from LDC • TIPSTER – Information Retrieval, Data Extrraction datasets – TIPSTER project, TREC project • TIMIT Acoustic-Phonetic Continuous Speech Corpus – A corpus of read speech designed to – Provide speech data for the acquisition of acousticphonetic knowledge – Useful for the development and evaluation of automatic speech recognition systems • ECI(European Corpus Initiative Multilingual Corpus) multilingual electronic text corpus • NTIMIT – A phonetically – balanced, continuous speech, telephone bandwidth speech database 11 Penn Treebank • A collection of corpora • Tagged with POS, Syntactic roles, predicate/argument structure, dysfluency annotation • How are they made – Hand correction of the output of an errorful automatic process • 3 million words – 1 million words tagged with predicate/argument structure for extraction semantic knowledge 12 Penn Treebank Cont.’d • Corpora – Wall Street Journal – ATIS (Air Travel Information System) – Brown Corpus – IBM Manual Sentences – Library of America Texts: Mark Twain, Henry Adams, Herman Melville ... – MUC-3 Messages • Example: ( (S (NP-SBJ Rally 's) (VP operates and franchises (NP (NP (QP about 160) fast-food restaurants) (PP-LOC throughout (NP the U.S)))) Seeking/VBG to/TO block/VB [ the/DT investors/NNS ] from/IN buying/VBG [ more/JJR shares/NNS ] ./. 13 DSO • Word Sense Corpus – Contains sentences in which about 192,800 word occurrences have been tagged with WordNet senses – Taken from the Brown corpus and the Wall Street Journal corpus – 121 nouns and 70 verbs 14 Hansard • Official records (Hansards) of the 36th Canadian Parliament, both in English of French • 1.3 million pairs of aligned sentences of English and French – Example • Comme il est 14 h 30, la Chambre s'ajourne jusqu'\xe0 lundi prochain, \xe0 11 heures, conform\xe9ment au paragraphe 24(1) du R\xe8glement. • It being 2.30 p.m., the House stands adjourned until Monday next at 11 a.m., pursuant to Standing Order 24(1). • Useful for Machine Translation 15 Morphological Tools • PC-KIMMO – A two-level morphological parser • Porter Stemmer • Penn Treebank Tokenizer – Seperate document into words – “dog?” -> “dog ?” 16 Porter Stemmer • Simple algorithm, use a set of cascaded rewrite rules – Example • Ational->ATE (relational->relate) • Stem: – The main morpheme of the word, supplying the main meaning • Fast • Used very widely in Information Retrieval – Run stemmer on keywords and the words in the documents 17 Part-Of-Speech(POS) Taggers • • • • Part-Of-Speech: noun, verb, pronoun, etc. Brill’s Tagger HMM Tagger MXPOST 18 Brill’s Tagger • • • • Transformation-Based Learning(TBL) tagger /projects/nlp/brill-pos-tagger First labels every word with its most-likely tag Then Use Learned TBL Rules to correct mistakes – Example: • Change NN to VB when the previous tag is TO 19 HMM Tagger • Also called Maximum Likelihood Tagger • Xerox PARC's HMM tagger: ftp://parcftp.xerox.com/pub/tagger/ • Choose the tag sequence with the maximum possibility given the words seen. 20 MXPOST: Maximum Entropy POS Tagger • Maximum Entropy Model is a framework integrating many information sources(called features) for classification • Each candidate tag is a class • Given features of the word(the around words, the morphological feature, and around tags, etc.), decide which class it belongs. 21 Syntactic Parsers • Collin’s Parser • XTAG • MXPOST: Maximum Entropy Parser 22 Collin’s Parser • Context-free Grammar • Use frequencies to solve ambiguities • Got some idea of this parser – Web-based Chart parser 23 XTAG • An on-going project to develop a wide-coverage grammar for English • using a lexicalized Tree Adjoining Grammar (TAG) formalism – Context sensitive grammar • consists of a parser, an X-windows grammar development interface and a morphological analyzer. • /projects/nlp/xtag/ 24 XTAG Cont’d 25 Semantic Knowledge Bases and Semantic Parser • • • • Analyze what does it say WordNet Penn Treebank Web-based Semantic Parser 26 WordNet • Respresents lexical relations • Useful in word sense disambiguation 27 Penn Treebank Predicate: fool(Kris) 28 Semantic Parser • A web-based chart parser enriched with semantic constraints • Example: – Input: My dog has fleas. – Output: has(my(dog),fleas) • 29 Speech Tools • ISIP • EPOS • CSLU Toolkit 30 ISIP • ISIP(Institute for Signal and Information Processing) public domain speech recognition system • Open research software • Online courses, tutorials, dictionaries, databases • Build your own speech recognition system 31 EPOS • a language independent rule-driven Text-toSpeech (TTS) system • supports several main speech generation algorithms 32 CSLU Toolkit • Basic framework and tools for people to build, investigate and use interactive language systems • speech recognition, natural language understanding, speech synthesis and facial animation technologies • Easy to use , spread from higher education into homes 33 Thanks! 34