[ Skip to the content ]

Institute of Formal and Applied Linguistics Wiki

[ Back to the navigation ]

You are here: start » draft

Differences

This shows you the differences between two versions of the page.

Link to this comparison view

Both sides previous revision Previous revision
Next revision 		Previous revision
draft [2009/08/31 14:28]
ptacek 		draft [2009/09/30 20:54] (current)
ptacek
Line 1: Line 1:
-====== Progress Report ======+====== Architecture Description ======
  
-[[Progress Report]] - dal jsem to na zvlastni stranku, abysme si nelezli do zeli+The Czech Companion follows the original idea of Reminiscing about the User's Photos, 
 +taking advantage of the data collected in the first phase of the project (using a Wizard-of-Oz setting). The full recorded corpora was transcribed, a manual speech reconstruction was done on 92.6% of utterances((Manual speech reconstruction is still in progress.)) and a pilot dialog acts annotation was performed on a sample of 1000 sentences.
  
 +The architecture is the same as in the English version, i.e. a set of modules communicating through the Inamode (TID) backbone. However, the set of modules is different, see Figure 1. Regarding the physical settings, the Czech version runs on two notebook computers connected by a local network. One serves as a Speech Client, running modules dealing with ASR, TTS and ECA; the other one as an NLP Server. 
  
-====== Description of Czech Companion November Demonstrator ======+The NLU pipeline, DM, and NLG modules at the NLP Server are implemented using a CU's own TectoMT platform that provides access to a single in-memory data representation through a common API. This eliminates the overhead of a repeated serialization and XML parsing that an Inamode based solution would impose otherwise.
  
-The Czech version of the Companion deals with the Reminiscing about User's Photos scenario, taking advantage of data recorded in first phase of the project. The basic architecture is same as of the English version, i.e. set of modules communicating through the Inamode Relayer (TID) backbone; however the set of modules is different (see Figure 1). Regarding the physical settings, the Czech version runs on two notebook computers connected by a local network; one can be seen as a Speech Client, running modules dealing with ASR, TTS and ECA, second as an NLP Server.+<html><br/><hr/><br/></html>
  
-The dialog is driven by a dialog manager component by USFD (originally developed for the English Senior Companion prototype), we supply the transition network (DAFs). The selection is backed by (a) appropriateness for the type of dialog we aim for (the corpus reveals frequent reoccurring topics to be handled by DAFs) , (b) availability of mature package within time frame that allows for integration, (c) possibility of reusing created DAF states, tests and specified actions for the upcoming statistical DM by UOXF (however this is post November work).+The ASR module based on Hidden-Markov models transforms input speech into text, providing a front-end between the user and the Czech demonstrator. The ASR output is smoothed into a form close to standard written text by the Speech Reconstruction module in order to bridge the gap between dis-fluent spontaneous speech and a standard grammatical sentence.
  
-Our DAFs covering selected topics contain not only Companion replies mined from the corpora, but also new human-authored assessments, remarks and glosses to provide longer system utterances in order to encourage user to tell more.+Results of the part-of-speech tagging are passed on to the Maximum Spanning Tree syntactic parsing module. A tectogrammatical representation of the utterance is constructed once the syntactic parse is available. Annotation of the meaning at tectogrammatical layer is more explicit than its syntactic parse and lends itself for information extraction. The Named Entity Recognition module then marks personal names and geographical locations. Afterwards, the dialog act classifier uses number of lexical and morphological features to assess the type of user utterance (such as question, acknowledgement, etc.) that is a useful clue for Dialog Manager decisions.
  
-For a sample dialogue, see the Scenario Brief below.  +The dialog is driven by a Dialog Manager component by USFD (originally developed for the English Senior Companion prototype). 
- +CU has supplied the transition networks covering following topics: retired_person, husband, child, wife, wedding and Christmas. 
-{{user:ptacek:czech_companion_diagram.png|}} +Dialogue Manager provides information about the appropriate communicative function along with the sentence that is to be generated to the NLG module. The TTS module integrated with the TID avatar transforms system responses from the text form into the speech and visual (face expressions, gestures) representation. As such, it provides an interface between the demonstrator and the user.
- +
- +
-===== Automatic Speech Recognition (WP 5.1)===== +
-features: improved language models, real-time speaker adaptation +
-performance indicator: WER  +
- +
- +
- +
- +
- +
-===== Speech Reconstruction (WP 5.2) ===== +
-features: omit filler phrases, remove irrelevant speech events, handle false starts, repetitions, and corrections, polish word ordering performance indicator: BLEU score between actual output and manually reconstructed sentences from corpora (T5.2.1), baseline: Moses with default settings  +
- +
- +
- +
- +
- +
- +
- +
- +
-===== Morphology Analyzer and POS tagging (WP 5.2) ===== +
-features: coverage of photo-pal domain, domain adapted tagger +
-performance indicator: OOV rate, accuracy  +
- +
- +
- +
- +
-===== Syntactic Parsing (WP 5.2) ===== +
-features: induce dependencies and labels +
-performance indicator: accuracy (correctly induced edges, labels) +
- +
- +
- +
- +
- +
-===== Semantic Parsing (WP 5.2) ===== +
-features: assignment of semantic roles (69 roles), coordinations, argument structure, partial ellipsis resolution, pronominal anaphora resolution, post parsing detection of ungrammatical edges (caused by long user utterances) +
-performance indicator: accuracy (correctly induced edges, labels) +
- +
- +
-===== Information Extraction (WP 5.2) ===== +
-features: template based identification of predicates +
-covering predicates from  before-mentioned set of DAFs. +
-performance indicator: accuracy +
- +
- +
- +
-===== Named Entities Recognition (WP 5.2) ===== +
-features: detect person names, geographical locations, organization names +
-performance indicator: f-measure +
- +
- +
- +
- +
-===== Dialog Act Tagging (WP 5.2) ===== +
-features: domain tailored tag-set (variation of DAMSL-SWBD) +
-performance indicator: accuracy +
- +
- +
- +
- +
- +
- +
- +
- +
- +
-===== Dialog Manager (WP 5.3) ===== +
-features: integrated DAF-based dialog manager from previous English prototype, +
-manual creation of DAFs covering following topics: Person_retired, Person_in_productive_age, Child, Husband, Wife, Wedding, Christmas, Death, Handling_stalled_dialog (most frequent topics in corpora), using customized DAF Editor provided by USFD. +
-performance indicator: acceptability - manual evaluation of actions selected by DM +
- +
- +
- +
- +
- +
- +
- +
-===== Natural Language Generation (WP 5.4) ===== +
-features: adding of functional words, morphological adjustments, generating paraphrases for hard-coded utterances, underspecified input (dott format). +
-performance indicator: BLEU score +
- +
- +
- +
-===== Emotional TTS (WP 5.5) ===== +
-features: emotions will be expressed implicitly, through the usage of communicative functions; new female voice database was recorded for this purposes +
-performance indicator: listening tests  +
- +
- +
- +
- +
-===== Emotional Avatar Integration (WP 5.5) ===== +
-features: new Czech female voice with affective features will be integrated with the TID avatar +
-performance indicator: subjective evaluation of the naturalness and the ability to convey emotions (small-scale, given the time constraint) +
- +
- +
- +
-====== Scenario Brief ====== +
- +
-C1 Dobrý den, jak se jmenujete? (intro-daf-based) +
-H1 Marie. +
- +
-C2 Těší mě. Já jsem Pavla a chtěla bych Vás blíž poznat. Jsem ráda, že mi ukážete Vaše fotografie. +
- +
-C2 Co je tohle? (photo_maintenance) +
-H2 To je moje rodina, rodiče, sestra, její manžel, moje dvě děti a její dvě děti. +
-H2 Byli u mě na Vánoce na návštěvě. +
- +
-C3 Trávíte Vánoce takhle pohromadě? +
-H3 Teď už ne. +
-H3 Teď už všechny moje dětičky vylétaly z hnízda. +
- +
-C4 Jak se jmenuje váš manžel? +
-H4 Manžel už nežije, jmenoval se Karel Šístek. +
- +
-C5 Máte ráda Vánoce? +
-H5 Teď už ne. +
-H5 Měla jsem je ráda, když byly děti malé. +
-H5 Teď už to ale není ono. +
- +
-C6 Takže teď trávíte Vánoce sama? (ask_on_detail-daf-based) +
-H6 Nejsem úplně sama, chodím každé Vánoce k dceři, ale už to není, jak to bývalo, když byly doma malé děti. +
- +
-C7 Jak děti vyrostou, tak má člověk najednou příliš mnoho času. +
-C7 Kdo je kdo na té fotce? +
-H7 U stromečku je můj synovec, ta holčička vedle je moje dcera, chlapeček s hračkou je můj syn, vedle je sestra, její manžel a nad ním je můj otec. +
- +
-C8 To jste ale velká rodina. +
-C8 Jak se jmenuje vaše sestra? (companion commenting last utterance  bundled with follow-up question - to achieve longer responses) +
-H8 Sestra se jmenuje Marie Havlová, její manžel je Vladimír, ten bohužel nedávno zemřel. +
- +
-C9 To je smutné. +
-C9 Znala jste ho dobře? +
-H9 Velmi dobře, trávili jsme obě rodiny společně hodně času. +
- +
-C10 Máte už nějaká vnoučata? +
-H10 Mám vnučku Natálku. +
-H10 Ona je takové živé děťátko, taková neposeda. +
- +
-C11 Takže Natálka je neposeda. repeat-rephrase +
-C11 Máte ještě nějaké další vnučky nebo vnuky? +
-H11 Čekáme ještě další přírustek. +
- +
-C12 Hlídáte je někdy? +
-H12 Hlídám dost často. +
-H12 Já mám děti ráda a jsem s nimi moc ráda. +
- +
-C13 To je fajn. +
-C13 S dětmi je legrace. +
-H13 Ano. +
- +
-C Podíváme se na další fotku?+
  
 +The Knowledge Base consists of objects (persons, events, photos) that model the information acquainted in the course of dialog. Those objects also provide a very basic reasoning (e.g. accounting for the link between date of birth and age properties). Each object's property is able to store multiple values with a varying level of confidence((Provided either by ASR module or from lexical clues contained in respective utterance.)), and values restricted to a defined time span.
  

[ Back to the navigation ] [ Back to the content ]