Differences

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

--- draft [2009/07/15 23:56]
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.
+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.
+<html><br/><hr/><br/></html>
-====== Description of Czech Companion November Demonstrator ======
+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.
-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.
+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.
-The dialog is driven by a dialog manager component by USFD (originally developed for the English Senior Companion prototype). 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 enough package within time frame that allows for integration, (c) possibility of reusing DAF states, tests and specified actions for the upcoming statistical DM by UOXF (however this is post November work).
+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.
+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.
-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.
+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.
-{{user:ptacek:companion-diagram-nov.png|}}
-===== 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 **(XXX prida nam Jarka OOV slova co najdeme - Mirek uz vyrabi list, bude PDTSC rucne oznackovane - do listopadu?)**
-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), passing-through of emotional markup (originating in DAFs)
-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)
-====== AZ PO LISTOPADU ======
-===== Syntactic Parsing (WP 5.2) =====
-features: adapted to domain (McD trained on manual PDTSC trees)
-performance indicator: accuracy (correctly induced edges, labels)
-===== Sentiment Analysis (WP 5.2) =====
-features: za tohle bych vydaval klasifikator, co rozhoduje ,jestli se rekne 'To je smutné/veselé'. Tem adjektivum rucne priradim negative/positive sentiment.
-performance indicator: f-measure
-===== Complete System Evaluation =====
-T5.2.7 tohle zminuje, nick webb to pro nas asi neudela
-performance indicator: number of tokens in user reply utterances, post-session questionare
-===== advances =====
-advances in Czech NLU (on reconstructed spoken data): 300-500vet(?) rucne anotovat pos, a-tree, t-tree, IE predicates, Named Entities, DA pro eval in-domain testy after Nov.
-pos ? analyzovat, generovat a kontrolovat 'jen' kde je rozdil ve forme?

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

Institute of Formal and Applied Linguistics Wiki

Differences