1 / 28

Computational Linguistics

Computational Linguistics. Ling 200 Spring 2006. Speech and language processing. Computational Linguistics use of computers to facilitate linguistic research Natural Language Processing computer-natural language interface applications. Combines disciplines. Linguistics

velvet
Download Presentation

Computational Linguistics

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Computational Linguistics Ling 200 Spring 2006

  2. Speech and language processing • Computational Linguistics • use of computers to facilitate linguistic research • Natural Language Processing • computer-natural language interface applications

  3. Combines disciplines • Linguistics • e.g. grammar engineering • Electrical Engineering • e.g. speech recognition • Computer science • e.g. machine translation • Psychology • e.g. cognitive modeling

  4. 2 Minute question (part 1) List the specific language related skills HAL exhibits. In other words, list the different abilities the computer (HAL) must have to display human-like language?

  5. Today’s goals • Convey: • some areas of research • some of the difficulties involved • some development strategies • Provide examples of particular technologies as illustration

  6. Computerized natural language • speech recognition • language understanding • language generation • speech synthesis

  7. Other areas of interest • searching • understanding search request • finding relevant documents • ordering by degree of relevance • information extraction • retrieving information from documents • data mining • discovering patterns and relationships in data

  8. ...and still more topics • machine translation • http://babelfish.altavista.com • http://www.google.com/translate • summarization • grammar checking • spell checking

  9. Commonly used tools • formal rule systems • computational search algorithms • formal logic • probability theory • machine learning techniques

  10. Speech Recognition Demo Software Used: iListen from MacSpeech

  11. What is Speech Recognition? • Definition: Speech recognition turns acoustic input into strings of phonemes and then finds the best matching word in a database. • Can be built for open domain use, theoretically recognizing all possible strings of words • e.g. dictation systems • Can also be built for a particular domain, recognizing small, finite sets of utterances • e.g. automated call-centers.

  12. Speech RecognitionAcoustic Model • First, the continuous speech signal is broken up into short segments. • Segments are analyzed into features, which you can think of as quantitative versions of the phonetic features you learned in class. • By comparing segments against internally stored phonological model, well matched phonemes are proposed for each segment • End up with a list of most likely phoneme sequences.

  13. Speech RecognitionLanguage Model • Sequences of phonemes are verified by comparing with a database of words and their likelihoods (in real time), and only actual words and phrases are accepted • [rɛkənajspič]  • [rɛkənajspič] • ‘recognizespeech’ • [rɛkənajspiš]  • [rɛkənajspiš] • ??‘recognizespeesh’ *Fast speech: [z] -> [s] / _[s]

  14. Problems Acoustic Model • Recognizing different voice qualities as the same basic sounds. • You can think of this as choosing the correct phoneme. • Phonemes sound different (allophones), depending on their environments. • word position: /p/ --> [ph] / #_ • assimilation: /z/ --> [s] / _C [-voice] • deletion: [s] --> ø / _[s] • “Three cats sit.” • Speech signal is continuous and full of non-speech noise.

  15. ProblemsAmbiguity • Same or very similar sequence of phonemes can correspond to multiple words or phrases • Homophones • Words • [dir] ‘deer’ ‘dear’ • Phrases (remember there is no pause to separate word boundaries) • [rɛkənajspič] ‘recognize speech’ • [rɛkənajspič] ‘wreck a nice beach’

  16. Potential FixLanguage Model • Weight word/phrase interpretations (statistical language modeling) • Lexical: Consider how often a word actually occurs. • [dir] ‘deer’ (50) ‘dear’ (215) • Choose most frequent, in this case ‘dear’ • Condition on context: Consider how often a word occurs within a particular context. • I just shot a [dir]. (shot, a, dear) 1 (shot, a, deer) 10 • In this case, ‘deer’ occurs more frequently in this environment, so we choose ‘deer’ as our interpretation.

  17. DemoTraining Data Matters • Word and context frequencies are not just pulled from thin air. • Frequencies are calculated (training) • From some collection of text (a corpus). • Speech recognizers often train on a user’s emails and documents, to better match the user’s lexical choice and phrase patterns. • This training data helps decipher homophonous strings (strings that are acoustically ambiguous).

  18. Demo 2Training Data Matters • I will attempt to utter the following phrase and iListen should transcribe my speech. • It’s hard to… • [rɛkənajspič]‘recognize speech’ • [rɛkənajspič]‘wreck a nice beach’

  19. Demo 3 Linguist • What if software is trained for a Computational Linguist? • Trained on 3 Wikipedia articles about various topics in Computational Linguistics • Which interpretation should we expect, based on words and phrases likely to be present in computational linguistics documents? • Results:Is hard to recognize speech New set the state • but is so bad and found a 544 is no sound better, even so it is etc is not really that bad so at and his exist listening to 89, Nancy of

  20. Demo 4 Beach Bum • What if software is trained for a Beach Bum? • Trained on 3 Wikipedia articles on beach topics. • Which interpretation should we expect, based on frequent words and phrases likely to be found in beach-related documents? • Results:It’s hard to wreck nice beach and

  21. Language understanding • morphology • syntax • semantics • pragmatics • discourse

  22. "I made her duck.” • I cooked waterfowl for her • I cooked waterfowl belonging to her • I created the (plaster?) duck she owns • I caused her to quickly lower her head or body • I waved my magic wand and turned her into undifferentiated waterfowl

  23. Language generation “I'm sorry, Dave, I'm afraid I can't do that” • pragmatics: • politeness • indirect speech • morphology: • contractions • discourse: • reference (“that”)

  24. Who/what is ELIZA?

  25. Dialogue systems - issues • HAL has complete understanding - How close are we to this? • Eliza had no semantic understanding and only minimal syntactic knowledge • dialogue systems: effective in limited domains like travel

  26. Dialogue systems: demo [David] • Chatbot website: • http://daden.co.uk/chatbots/

  27. 2 minute question (part 2) • Do you think that HAL quality computer communication is a reasonable expectation? • Why or why not?

More Related