| { | |
| "paper_id": "W07-0307", | |
| "header": { | |
| "generated_with": "S2ORC 1.0.0", | |
| "date_generated": "2023-01-19T04:41:27.476036Z" | |
| }, | |
| "title": "Experiments on the France Telecom 3000 Voice Agency corpus: academic research on an industrial spoken dialog system *", | |
| "authors": [ | |
| { | |
| "first": "G\u00e9raldine", | |
| "middle": [], | |
| "last": "Damnati", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "[email protected]" | |
| }, | |
| { | |
| "first": "France", | |
| "middle": [], | |
| "last": "T\u00e9l\u00e9com", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Fr\u00e9d\u00e9ric", | |
| "middle": [], | |
| "last": "B\u00e9chet", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "[email protected]" | |
| }, | |
| { | |
| "first": "Renato", | |
| "middle": [], | |
| "last": "De Mori", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "[email protected]" | |
| } | |
| ], | |
| "year": "", | |
| "venue": null, | |
| "identifiers": {}, | |
| "abstract": "", | |
| "pdf_parse": { | |
| "paper_id": "W07-0307", | |
| "_pdf_hash": "", | |
| "abstract": [], | |
| "body_text": [ | |
| { | |
| "text": "The recent advances in speech recognition technologies, and the experience acquired in the development of WEB or Interactive Voice Response interfaces, have facilitated the integration of speech modules in robust Spoken Dialog Systems (SDS), leading to the deployment on a large scale of speech-enabled services. With these services it is possible to obtain very large corpora of human-machine interactions by collecting system logs. This new kinds of systems and dialogue corpora offer new opportunities for academic research while raising two issues: How can academic research take profit of the system logs of deployed SDS in order to build the next generation of SDS, although the dialogues collected have a dialogue flow constrained by the previous SDS generation? On the other side, what immediate benefits can academic research offer for the improvement of deployed system? This paper addresses these aspects in the framework of the deployed France Telecom 3000 Voice Agency service.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Since the deployment on a very large scale of the AT&T How May I Help You? (HMIHY) (Gorin et al., 1997) service in 2000, Spoken Dialogue Systems (SDS) handling a very large number of calls are now developed from an industrial point of view. Although a lot of the remaining problems (robustness, coverage, etc.) are still spoken language processing research problems, the conception and the deployment of such state-of-the-art systems mainly requires knowledge in user interfaces.", | |
| "cite_spans": [ | |
| { | |
| "start": 83, | |
| "end": 103, | |
| "text": "(Gorin et al., 1997)", | |
| "ref_id": "BIBREF3" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "The recent advances in speech recognition technologies, and the experience acquired in the development of WEB or Interactive Voice Response interfaces have facilitated the integration of speech modules in robust SDS.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "These new SDS can be deployed on a very large scale, like the France Telecom 3000 Voice Agency service considered in this study. With these services it is possible to obtain very large corpora of humanmachine interactions by collecting system logs. The main differences between these corpora and those collected in the framework of evaluation programs like the DARPA ATIS (Hemphill et al., 1990) or the French Technolangue MEDIA (Bonneau-Maynard et al., 2005) programs can be expressed through the following dimensions:", | |
| "cite_spans": [ | |
| { | |
| "start": 372, | |
| "end": 395, | |
| "text": "(Hemphill et al., 1990)", | |
| "ref_id": "BIBREF4" | |
| }, | |
| { | |
| "start": 429, | |
| "end": 459, | |
| "text": "(Bonneau-Maynard et al., 2005)", | |
| "ref_id": "BIBREF1" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "\u2022 Size. There are virtually no limits in the amount of speakers available or the time needed for collecting the dialogues as thousands of dialogues are automatically processed every day and the system logs are stored. Therefore Dialog processing becomes similar to Broadcast News processing: the limit is not in the amount of data available, but rather in the amount of data that can be manually annotated.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "\u2022 Speakers. Data are from real users. The speakers are not professional ones or have no reward for calling the system. Therefore their behaviors are not biased by the acquisition protocols. Spontaneous speech and speech affects can be observed.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "\u2022 Complexity. The complexity of the services widely deployed is necessarily limited in order to guarantee robustness with a high automation rate. Therefore the dialogues collected are often short dialogues.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "\u2022 Semantic model. The semantic model of such deployed system is task-oriented. The interpretation of an utterance mostly consists in the detection of application-specific entities. In an application like the France Telecom 3000 Voice Agency service this detection is performed by hand-crafted specific knowledge.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "The AT&T HMIHY corpus was the first large dialogue corpus, obtained from a deployed system, that has the above mentioned characteristics. A service like the France Telecom 3000 Voice Agency service has been developed by a user interface development lab. This new kind of systems and dialogue corpora offer new opportunities for academic research that can be summarized as follows:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "\u2022 How can academic research take profit of the system logs of deployed SDS in order to build the next generation of SDS, although the dialogues collected have a dialogue flow constrained by the previous SDS generation?", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "\u2022 On the other side, what immediate benefits can academic research offer for the improvement of deployed system, while waiting for the next SDS generation?", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "This paper addresses these aspects in the framework of the deployed FT 3000 Voice Agency service. Section 3 presents how the ASR process can be modified in order to detect and reject Out-Of-Domain utterances, leading to an improvement in the understanding performance without modifying the system. Section 4 shows how the FT 3000 corpus can be used in order to build stochastic models that are the basis of a new Spoken Language Understanding strategy, even if the current SLU system used in the FT 3000 service is not stochastic. Section 5 presents experimental results obtained on this corpus justifying the need of a tighter integration between the ASR and the SLU models.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "The France Telecom 3000 (FT3000) Voice Agency service, the first deployed vocal service at France Telecom exploiting natural language technologies, has been made available to the general public in October 2005. FT3000 service enables customers to obtain information and purchase almost 30 different services and access the management of their services. The continuous speech recognition system relies on a bigram language model. The interpretation is achieved through the Verbateam two-steps semantic analyzer. Verbateam includes a set of rules to convert the sequence of words hypothesized by the speech recognition engine into a sequence of concepts and an inference process that outputs an interpretation label from a sequence of concepts.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Description of the France Telecom 3000 Voice Agency corpus", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "Given the main functionalities of the application, two types of dialogues can be distinguished. Some users call FT 3000 to activate some services they have already purchased. For such demands, users are rerouted toward specific vocal services that are dedicated to those particular tasks. In that case, the FT3000 service can be seen as a unique automatic frontal desk that efficiently redirects users. For such dialogues the collected corpora only contain the interaction prior to rerouting. It can be observed in that case that users are rather familiar to the system and are most of the time regular users. Hence, they are more likely to use short utterances, sometimes just keywords and the interaction is fast (between one or two dialogue turns in order to be redirected to the demanded specific service). Such dialogues will be referred as transit dialogues and represent 80% of the calls to the FT3000 service. As for the 20% other dialogues, referred to as other, the whole interaction is proceeded within the FT3000 application. They concern users that are more generally asking for information about a given service or users that are willing to purchase a new service. For these dialogues, the average utterance length is higher, as well as the average number of dialogue turns. As can be observed in table 1 the fact that users are less familiar with the application in the other dialogues implies higher OOV rate and disfluency rate 1 . An important issue when designing ASR and SLU models for such applications that are dedicated to the general public is to be able to handle both naive users and familiar users. Models have to be robust enough for new users to accept the service and in the meantime they have to be efficient enough for familiar users to keep on using it. This is the reason why experimental results will be detailed on the two corpora described in this section.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Specificities of interactions", | |
| "sec_num": "2.1" | |
| }, | |
| { | |
| "text": "When dealing with real users corpora, one has to take into account the occurrence of Out-Of-Domain (OOD) utterances. Users that are familiar with a service are likely to be efficient and to strictly answer the system's prompts. New users can have more diverse reactions and typically make more comments about the system. By comments we refer to such cases when a user can either be surprised what am I supposed to say now?, irritated I've already said that or even insulting the system. A critical aspect for other dialogues is the higher rate of comments uttered by users. For the transit dialogues this phenomenon is much less frequent because users are fa-1 by disfluency we consider here false starts and filled pauses miliar to the system and they know how to be efficient and how to reach their goal. As shown in table 2, 14.3% of the other dialogues contain at least one OOD comment, representing an overall 10.6% of utterances in these dialogues.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "User behavior and OOD utterances", | |
| "sec_num": "2.2" | |
| }, | |
| { | |
| "text": "other transit # dialogues 350 467 # utterances 1288 717 # OOD comments 137 24 OOD rate (%) 10.6 3.3 dialogues with OOD (%) 14.3 3.6 Table 2 : Occurrence of Out-Of-Domain comments on the transit and other dialogues Some utterances are just comments and some contain both useful information and comments. In the next section, we propose to detect these OOD sequences and to take this phenomenon into account in the global SLU strategy.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 132, | |
| "end": 139, | |
| "text": "Table 2", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "User behavior and OOD utterances", | |
| "sec_num": "2.2" | |
| }, | |
| { | |
| "text": "The general purpose of the proposed strategy is to detect OOD utterances in a first step, before entering the Spoken Language Understanding (SLU) module. Indeed standard Language Models (LMs) applied to OOD utterances are likely to generate erroneous speech recognition outputs and more generally highly noisy word lattices from which it might not be relevant and probably harmful to apply SLU modules.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Handling Out-Of-Domain utterances", | |
| "sec_num": "3" | |
| }, | |
| { | |
| "text": "Furthermore, when designing a general interaction model which aims at predicting dialogue states as proposed in this paper, OOD utterances are as harmful for state prediction as can be an out-ofvocabulary word for the prediction of the next word with an n-gram LM. This is why we propose a new composite LM that integrates two sub-LMs: one LM for transcribing indomain phrases, and one LM for detecting and deleting OOD phrases. Finally the different SLU strategies proposed in this paper are applied only to the portions of signal labeled as in-domain utterances.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Handling Out-Of-Domain utterances", | |
| "sec_num": "3" | |
| }, | |
| { | |
| "text": "As a starting point, the comments have been manually annotated in the training data in order to easily separate OOD comment segments from in-domain ones. A specific bigram language model is trained for these comment segments. The comment LM was designed from a 765 words lexicon and trained on 1712 comment sequences. This comment LM, called LM OOD has been integrated in the general bigram LM G . Comment sequences have been parsed in the training corpus and replaced by a OOD tag. This tag is added to the general LM vocabulary and bigram probabilities P ( OOD |w) and P (w| OOD ) are trained along with other bigram probabilities (following the principle of a priori word classes). During the decoding process, the general bigram LM probabilities and the LM OOD bigram probabilities are combined.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Composite Language Model for decoding spontaneous speech", | |
| "sec_num": "3.1" | |
| }, | |
| { | |
| "text": "Given this composite LM, a decision strategy is applied to select those utterances for which the word lattice will be processed by the SLU component. This decision is made upon the one-best speech recognition hypotheses and can be described as follows:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Decision strategy", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "1. If the one-best ASR output is a single OOD tag, the utterance is simply rejected.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Decision strategy", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "2. Else, if the one-best ASR output contains an OOD tag along with other words, those words are processed directly by the SLU component, following the argument that the word lattice for this utterance is likely to contain noisy information.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Decision strategy", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "3. Else (i.e. no OOD tag in the one-best ASR output), the word-lattice is transmitted to further SLU components.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Decision strategy", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "It will be shown in the experimental section that this pre-filtering step, in order to decide whether a word lattice is worth being processed by the higherlevel SLU components, is an efficient way of preventing concepts and interpretation hypothesis to be decoded from an uninformative utterance.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Decision strategy", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "The models presented are trained on a corpus collected thanks to the FT3000 service. It contains real dialogues from the deployed service. The results presented are obtained on the test corpus described in section 2.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Experimental setup and evaluation", | |
| "sec_num": "3.3" | |
| }, | |
| { | |
| "text": "The results were evaluated according to 3 criteria: the Word Error Rate (WER), the Concept Error Rate (CER) and the Interpretation Error Rate (IER). The CER is related to the correct translation of an utterance into a string of basic concepts. The IER is related to the global interpretation of an utterance in the context of the dialogue service considered. Therefore this last measure is the most significant one as it is directly linked to the performance of the dialogue system. Table 3 presents the IER results obtained with the strategy strat1 with 2 different LMs for obtainin\u011d W : LM G which is the general word bigram model; and LM G + OOD which is the LM with the OOD comment model. As one can see, a very significant improvement, 3.7% absolute, is achieved on the other dialogues, which are the ones containing most of the comments. For the transit dialogues a small improvement (0.2%) is also obtained.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 483, | |
| "end": 490, | |
| "text": "Table 3", | |
| "ref_id": "TABREF2" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Experimental setup and evaluation", | |
| "sec_num": "3.3" | |
| }, | |
| { | |
| "text": "The SLU component of the FT3000 service considered in this study contains two stages:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Building stochastic SLU strategies 4.1 The FT3000 SLU module", | |
| "sec_num": "4" | |
| }, | |
| { | |
| "text": "1. the first one translates a string of words W = w 1 , . . . , w n into a string of elementary concepts C = c 1 , . . . , c l by means of hand-written regular grammars;", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Building stochastic SLU strategies 4.1 The FT3000 SLU module", | |
| "sec_num": "4" | |
| }, | |
| { | |
| "text": "2. the second stage is made of a set of about 1600 inference rules that take as input a string of concepts C and output a global interpretation \u03b3 of a message. These rules are ordered and the first match obtained by processing the concept string is kept as the output interpretation.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Building stochastic SLU strategies 4.1 The FT3000 SLU module", | |
| "sec_num": "4" | |
| }, | |
| { | |
| "text": "These message interpretations are expressed by an attribute/value pair representing a function in the vocal service. The models used in these two stages are manually defined by the service designers and are not stochastic. We are going now to present how we can use a corpus obtained with such models in order to define an SLU strategy based on stochastic processes.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Building stochastic SLU strategies 4.1 The FT3000 SLU module", | |
| "sec_num": "4" | |
| }, | |
| { | |
| "text": "The actual FT3000 system includes semantic knowledge represented by hand-written rules. These rules can also be expressed in a logic form. For this reason, some basic concepts are now described with the purpose of showing how logic knowledge has been integrated in a first probabilistic model and how it can be used in a future version in which optimal policies can be applied.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Semantic knowledge representation", | |
| "sec_num": "4.2" | |
| }, | |
| { | |
| "text": "The semantic knowledge of an application is a knowledge base (KB) containing a set of logic formulas. Formulas return truth and are constructed using constants which represent objects and may be typed, variables, functions which are mappings from tuples of objects to objects and predicates which represent relations among objects. An interpretation specifies which objects, functions and relations in the domain are represented by which symbol. Basic inference problem is to determine whether KB |= F which means that KB entails a formula F .", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Semantic knowledge representation", | |
| "sec_num": "4.2" | |
| }, | |
| { | |
| "text": "In SLU, interpretations are carried on by binding variables and instantiating objects based on ASR results and inferences performed in the KB. Hypotheses about functions and instantiated objects are written into a Short Term Memory (STM).", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Semantic knowledge representation", | |
| "sec_num": "4.2" | |
| }, | |
| { | |
| "text": "A user goal is represented by a conjunction of predicates. As dialogue progresses, some predicates are grounded by the detection of predicate tags, property tags and values. Such a detection is made by the interpretation component. Other predicates are grounded as a result of inference. A user goal G is asserted when all the atoms of its conjunction are grounded and asserted true.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Semantic knowledge representation", | |
| "sec_num": "4.2" | |
| }, | |
| { | |
| "text": "Grouping the predicates whose conjunction is the premise for asserting a goal G i is a process that goes through a sequence of states:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Semantic knowledge representation", | |
| "sec_num": "4.2" | |
| }, | |
| { | |
| "text": "S 1 (G i ), S 2 (G i ), . . .", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Semantic knowledge representation", | |
| "sec_num": "4.2" | |
| }, | |
| { | |
| "text": "Let \u0393 i k be the content of the STM used for asserting the predicates grounded at the k-th turn of a dialogue. These predicates are part of the premise for asserting the i-th goal.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Semantic knowledge representation", | |
| "sec_num": "4.2" | |
| }, | |
| { | |
| "text": "Let G i be an instance of the i-th goal asserted after grounding all the predicates in the premise.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Semantic knowledge representation", | |
| "sec_num": "4.2" | |
| }, | |
| { | |
| "text": "\u0393 i k can be represented by a composition from a partial hypothesis \u0393 i k \u2212 1 available at turn k \u2212 1, the machine action a k\u22121 performed at turn k \u2212 1 and the semantic interpretation \u03b3 i k i.e.:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Semantic knowledge representation", | |
| "sec_num": "4.2" | |
| }, | |
| { | |
| "text": "\u0393 i k = \u03c7 \u03b3 i k , a k\u22121 , \u0393 i k\u22121 S k (G i )", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Semantic knowledge representation", | |
| "sec_num": "4.2" | |
| }, | |
| { | |
| "text": "is an information state that can lead to a user's goal G i and \u0393 i k is part of the premise for asserting G i at turn k.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Semantic knowledge representation", | |
| "sec_num": "4.2" | |
| }, | |
| { | |
| "text": "State probability can be written as follows:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Semantic knowledge representation", | |
| "sec_num": "4.2" | |
| }, | |
| { | |
| "text": "P (S k (G i )|Y k ) = P G i |\u0393 i k P \u0393 i k |Y k (1)", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Semantic knowledge representation", | |
| "sec_num": "4.2" | |
| }, | |
| { | |
| "text": "where P G i |\u0393 i k is the probability that G i is the type of goal that corresponds to the user interaction given the grounding predicates in \u0393 i k . Y k is the acoustic features of the user's utterance at turn k.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Semantic knowledge representation", | |
| "sec_num": "4.2" | |
| }, | |
| { | |
| "text": "Probabilities of states can be used to define a belief of the dialogue system.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Semantic knowledge representation", | |
| "sec_num": "4.2" | |
| }, | |
| { | |
| "text": "A first model allowing multiple dialog state sequence hypothesis is proposed in (Damnati et al., 2007) . In this model each dialog state correspond to a system state in the dialog automaton. In order to deal with flexible dialog strategies and following previous work (Williams and Young, 2007) , a new model based on a Partially Observable Markov Decision Process (POMDP) is currently studied.", | |
| "cite_spans": [ | |
| { | |
| "start": 80, | |
| "end": 102, | |
| "text": "(Damnati et al., 2007)", | |
| "ref_id": "BIBREF2" | |
| }, | |
| { | |
| "start": 268, | |
| "end": 294, | |
| "text": "(Williams and Young, 2007)", | |
| "ref_id": "BIBREF10" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Semantic knowledge representation", | |
| "sec_num": "4.2" | |
| }, | |
| { | |
| "text": "If no dialog history is taken into account, P \u0393 i k |Y comes down to P \u03b3 i k |Y , \u03b3 i k being a semantic attribute/value pair produced by the Verbateam interpretation rules.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Semantic knowledge representation", | |
| "sec_num": "4.2" | |
| }, | |
| { | |
| "text": "The integration of this semantic decoding process in the ASR process is presented in the next section.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Semantic knowledge representation", | |
| "sec_num": "4.2" | |
| }, | |
| { | |
| "text": "With the stochastic models proposed in section 4, different strategies can be built and optimized. We are interested here in the integration of the ASR and SLU processes. As already shown by previous studies (Wang et al., 2005) , the traditional sequential approach that first looks for the best sequence of words \u0174 before looking for the best interpretation\u03b3 of an utterance is sub-optimal. Performing SLU on a word lattice output by the ASR module is an efficient way of integrating the search for the best sequence of words and the best interpretation. However there are real-time issues in processing word lattices in SDS, and therefore they are mainly used in research systems rather than deployed systems.", | |
| "cite_spans": [ | |
| { | |
| "start": 208, | |
| "end": 227, | |
| "text": "(Wang et al., 2005)", | |
| "ref_id": "BIBREF9" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Optimizing the ASR and SLU processes", | |
| "sec_num": "5" | |
| }, | |
| { | |
| "text": "In section 3 a strategy is proposed for selecting the utterances for which a word lattice is going to be produced. We are going now to evaluate the gain in performance that can be obtained thanks to an integrated approach on these selected utterances.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Optimizing the ASR and SLU processes", | |
| "sec_num": "5" | |
| }, | |
| { | |
| "text": "Two strategies are going to be evaluated. The first one (strat1) is fully sequential: the best sequence of word\u0174 is first obtained wit\u0125", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Sequential vs. integrated strategies", | |
| "sec_num": "5.1" | |
| }, | |
| { | |
| "text": "W = argmax W P (W |Y )", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Sequential vs. integrated strategies", | |
| "sec_num": "5.1" | |
| }, | |
| { | |
| "text": "Then the best sequence of concepts\u0108 is obtained with\u0108 = argmax", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Sequential vs. integrated strategies", | |
| "sec_num": "5.1" | |
| }, | |
| { | |
| "text": "C P (C|\u0174 )", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Sequential vs. integrated strategies", | |
| "sec_num": "5.1" | |
| }, | |
| { | |
| "text": "Finally the interpretation rules are applied to\u0108 in order to obtain the best interpretation\u03b3. The second strategy (strat2) is fully integrated:\u03b3 is obtained by searching at the same time for\u0174 and C and\u03b3. In this case we have:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Sequential vs. integrated strategies", | |
| "sec_num": "5.1" | |
| }, | |
| { | |
| "text": "\u03b3 = argmax W,C,\u03b3 P (\u03b3|C)P (C|W )P (W |Y )", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Sequential vs. integrated strategies", | |
| "sec_num": "5.1" | |
| }, | |
| { | |
| "text": "The stochastic models proposed are implemented with a Finite State Machine (FSM) paradigm thanks to the AT&T FSM toolkit (Mohri et al., 2002) .", | |
| "cite_spans": [ | |
| { | |
| "start": 121, | |
| "end": 141, | |
| "text": "(Mohri et al., 2002)", | |
| "ref_id": "BIBREF5" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Sequential vs. integrated strategies", | |
| "sec_num": "5.1" | |
| }, | |
| { | |
| "text": "Following the approach described in (Raymond et al., 2006) , the SLU first stage is implemented by means of a word-to-concept transducer that translates a word lattice into a concept lattice. This concept lattice is rescored with a Language Model on the concepts (also encoded as FSMs with the AT&T GRM toolkit (Allauzen et al., 2003) ).", | |
| "cite_spans": [ | |
| { | |
| "start": 36, | |
| "end": 58, | |
| "text": "(Raymond et al., 2006)", | |
| "ref_id": "BIBREF6" | |
| }, | |
| { | |
| "start": 311, | |
| "end": 334, | |
| "text": "(Allauzen et al., 2003)", | |
| "ref_id": "BIBREF0" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Sequential vs. integrated strategies", | |
| "sec_num": "5.1" | |
| }, | |
| { | |
| "text": "The rule database of the SLU second stage is encoded as a transducer that takes as input concepts and output semantic interpretations \u03b3. By applying this transducer to an FSM representing a concept lattice, we directly obtain a lattice of interpretations.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Sequential vs. integrated strategies", | |
| "sec_num": "5.1" | |
| }, | |
| { | |
| "text": "The SLU process is therefore made of the composition of the ASR word lattice, two transducers (word-to-concepts and concept-to-interpretations) and an FSM representing a Language Model on the concepts. The concept LM is trained on the FT3000 corpus.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Sequential vs. integrated strategies", | |
| "sec_num": "5.1" | |
| }, | |
| { | |
| "text": "This strategy push forward the approach developped at AT&T in the How May I Help You? (Gorin et al., 1997) project by using richer semantic models than call-types and named-entities models. More precisely, the 1600 Verbateam interpretation rules used in this study constitute a rich knowledge base. By integrating them into the search, thanks to the FSM paradigm, we can jointly optimize the search for the best sequence of words, basic concepts, and full semantic interpretations.", | |
| "cite_spans": [ | |
| { | |
| "start": 86, | |
| "end": 106, | |
| "text": "(Gorin et al., 1997)", | |
| "ref_id": "BIBREF3" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Sequential vs. integrated strategies", | |
| "sec_num": "5.1" | |
| }, | |
| { | |
| "text": "For the strategy strat1 only the best path is kept in the FSM corresponding to the word lattice, simulating a sequential approach. For strat2 the best interpretation\u03b3 is obtained on the whole concept lattice.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Sequential vs. integrated strategies", | |
| "sec_num": "5.1" | |
| }, | |
| { | |
| "text": "WER CER IER strat1 40.1 24.4 15.0 strat2 38.2 22.5 14.5 The comparison among the two strategies is given in table 4. As we can see a small improvement is obtained for the interpretation error rate (IER) with the integrated strategy (strat2). This gain is small; however it is interesting to look at the Oracle IER that can be obtained on an n-best list of interpretations produced by each strategy (the Oracle IER being the lowest IER that can be obtained on an n-best list of hypotheses with a perfect Oracle decision process). This comparison is given in Figure 1 . As one can see a much lower Oracle IER can be achieved with strat2. For example, with an n-best list of 5 interpretations, the lowest IER is 7.4 for strat1 and only 4.8 for strat2. This is very interesting for dialogue systems as the Dialog Manager can use dialogue context information in order to filter such n-best lists. Oracle IER size of the n-best list of interpretations sequential search (strat1) integrated search (strat2) Figure 1 : Oracle IER according to an n-best list of interpretations for strategies strat1 and strat2 Table 4 also indicates that the improvements obtained on the WER and CER dimensions don't always lead to similar improvements in IER. This is due to the fact that the improvements in WER and CER are mostly due to a significant reduction in the insertion rates of words and concepts. Because the same weight is usually given to all kinds of errors (insertions, substitutions and deletions), a decrease in the overall error rate can be misleading as interpretation strategies can deal more easily with insertions than deletions or substitutions. Therefore the reduction of the overall WER and CER measures is not a reliable indicator of an increase of performance of the whole SLU module. level 1-best Oracle hyp. WER 33.7 20.0 CER 21.2 9.7 IER 13.0 4.4 Table 5 : Error rates on words, concepts and interpretations for the 1-best hypothesis and for the Oracle hypothesis of each level These results have already been shown for WER by previous studies like (Riccardi and Gorin, 1998) IER from word Oracle 9.8 from concept Oracle 7.5 interpretation Oracle 4.4 Table 6 : IER obtained on Oracle hypotheses computed at different levels. or more recently (Wang et al., 2003) . They are illustrated by Table 5 and Table 6 . The figures shown in these tables were computed on the subset of utterances that were passed to the SLU component. Utterances for which an OOD has been detected are discarded. In Table 5 are displayed the error rates obtained on words, concepts and interpretations both on the 1-best hypothesis and on the Oracle hypothesis (the one with the lowest error rate in the lattice). These Oracle error rates were obtained by looking for the best hypothesis in the lattice obtained at the corresponding level (e.g. looking for the best sequence of concepts in the concept lattice). As for Table 6, the mentioned IER are the one obtained when applying SLU to the Oracles hypotheses computed for each level. As one can see the lowest IER (4.4) is not obtained on the hypotheses with the lowest WER (9.8) or CER (7.5).", | |
| "cite_spans": [ | |
| { | |
| "start": 2056, | |
| "end": 2082, | |
| "text": "(Riccardi and Gorin, 1998)", | |
| "ref_id": "BIBREF7" | |
| }, | |
| { | |
| "start": 2249, | |
| "end": 2268, | |
| "text": "(Wang et al., 2003)", | |
| "ref_id": "BIBREF8" | |
| } | |
| ], | |
| "ref_spans": [ | |
| { | |
| "start": 557, | |
| "end": 565, | |
| "text": "Figure 1", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 1000, | |
| "end": 1008, | |
| "text": "Figure 1", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 1102, | |
| "end": 1109, | |
| "text": "Table 4", | |
| "ref_id": "TABREF3" | |
| }, | |
| { | |
| "start": 1854, | |
| "end": 1861, | |
| "text": "Table 5", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 2158, | |
| "end": 2165, | |
| "text": "Table 6", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 2295, | |
| "end": 2314, | |
| "text": "Table 5 and Table 6", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 2496, | |
| "end": 2503, | |
| "text": "Table 5", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "error", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "This paper presents a study on the FT3000 corpus collected from real users on a deployed general public application. Two problematics are addressed: How can such a corpus be helpful to carry on research on advanced SLU methods eventhough it has been collected from a more simple rule-based dialogue system? How can academic research translate into short-term improvements for deployed services? This paper proposes a strategy for integrating advanced SLU components in deployed services. This strategy consists in selecting the utterances for which the advanced SLU components are going to be applied. Section 3 presents such a strategy that consists in filtering Out-Of-Domain utterances during the ASR first pass, leading to significant improvement in the understanding performance.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Conclusion", | |
| "sec_num": "6" | |
| }, | |
| { | |
| "text": "For the SLU process applied to in-domain utterances, an integrated approach is proposed that looks simultaneously for the best sequence of words, concepts and interpretations from the ASR word lattices. Experiments presented in section 5 on real data show the advantage of the integrated approach towards the sequential approach. Finally, section 4 proposes a unified framework that enables to define a dialogue state prediction model that can be applied and trained on a corpus collected through an already deployed service.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Conclusion", | |
| "sec_num": "6" | |
| } | |
| ], | |
| "back_matter": [], | |
| "bib_entries": { | |
| "BIBREF0": { | |
| "ref_id": "b0", | |
| "title": "Generalized algorithms for constructing statistical language models", | |
| "authors": [ | |
| { | |
| "first": "Cyril", | |
| "middle": [], | |
| "last": "Allauzen", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Mehryar", | |
| "middle": [], | |
| "last": "Mohri", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Brian", | |
| "middle": [], | |
| "last": "Roark", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2003, | |
| "venue": "41st Annual Meeting of the Association for Computational Linguistics (ACL'03)", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Cyril Allauzen, Mehryar Mohri, and Brian Roark. 2003. Generalized algorithms for constructing statistical lan- guage models. In 41st Annual Meeting of the Asso- ciation for Computational Linguistics (ACL'03), Sap- poro, Japan.", | |
| "links": null | |
| }, | |
| "BIBREF1": { | |
| "ref_id": "b1", | |
| "title": "Semantic annotation of the french media dialog corpus", | |
| "authors": [ | |
| { | |
| "first": "Helene", | |
| "middle": [], | |
| "last": "Bonneau-Maynard", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Sophie", | |
| "middle": [], | |
| "last": "Rosset", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2005, | |
| "venue": "Proceedings of the European Conference on Speech Communication and Technology (Eurospeech), Lisboa", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Helene Bonneau-Maynard, Sophie Rosset, Christelle Ay- ache, Anne Kuhn, and Djamel Mostefa. 2005. Se- mantic annotation of the french media dialog corpus. In Proceedings of the European Conference on Speech Communication and Technology (Eurospeech), Lis- boa, Portugal.", | |
| "links": null | |
| }, | |
| "BIBREF2": { | |
| "ref_id": "b2", | |
| "title": "Spoken Language Understanding strategies on the France Telecom 3000 voice agency corpus", | |
| "authors": [ | |
| { | |
| "first": "Geraldine", | |
| "middle": [], | |
| "last": "Damnati", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Frederic", | |
| "middle": [], | |
| "last": "Bechet", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Renato", | |
| "middle": [ | |
| "De" | |
| ], | |
| "last": "Mori", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2007, | |
| "venue": "Proceedings of the International Conference on Acoustics, Speech and Signal Processing", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Geraldine Damnati, Frederic Bechet, and Renato De Mori. 2007. Spoken Language Understanding strategies on the France Telecom 3000 voice agency corpus. In Proceedings of the International Con- ference on Acoustics, Speech and Signal Processing (ICASSP), Honolulu, USA.", | |
| "links": null | |
| }, | |
| "BIBREF3": { | |
| "ref_id": "b3", | |
| "title": "How May I Help You ?", | |
| "authors": [ | |
| { | |
| "first": "A", | |
| "middle": [ | |
| "L" | |
| ], | |
| "last": "Gorin", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "G", | |
| "middle": [], | |
| "last": "Riccardi", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J", | |
| "middle": [ | |
| "H" | |
| ], | |
| "last": "Wright", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1997, | |
| "venue": "Speech Communication", | |
| "volume": "23", | |
| "issue": "", | |
| "pages": "113--127", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "A. L. Gorin, G. Riccardi, and J.H. Wright. 1997. How May I Help You ? In Speech Communication, vol- ume 23, pages 113-127.", | |
| "links": null | |
| }, | |
| "BIBREF4": { | |
| "ref_id": "b4", | |
| "title": "The ATIS spoken language systems pilot corpus", | |
| "authors": [ | |
| { | |
| "first": "Charles", | |
| "middle": [ | |
| "T" | |
| ], | |
| "last": "Hemphill", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "John", | |
| "middle": [ | |
| "J" | |
| ], | |
| "last": "Godfrey", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "George", | |
| "middle": [ | |
| "R" | |
| ], | |
| "last": "Doddington", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1990, | |
| "venue": "Proceedings of the workshop on Speech and Natural Language", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "96--101", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Charles T. Hemphill, John J. Godfrey, and George R. Doddington. 1990. The ATIS spoken language sys- tems pilot corpus. In Proceedings of the workshop on Speech and Natural Language, pages 96-101, Hidden Valley, Pennsylvania.", | |
| "links": null | |
| }, | |
| "BIBREF5": { | |
| "ref_id": "b5", | |
| "title": "Weighted finite-state transducers in speech recognition", | |
| "authors": [ | |
| { | |
| "first": "Mehryar", | |
| "middle": [], | |
| "last": "Mohri", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Fernando", | |
| "middle": [], | |
| "last": "Pereira", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Michael", | |
| "middle": [], | |
| "last": "Riley", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2002, | |
| "venue": "Computer, Speech and Language", | |
| "volume": "16", | |
| "issue": "1", | |
| "pages": "69--88", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Mehryar Mohri, Fernando Pereira, and Michael Ri- ley. 2002. Weighted finite-state transducers in speech recognition. Computer, Speech and Language, 16(1):69-88.", | |
| "links": null | |
| }, | |
| "BIBREF6": { | |
| "ref_id": "b6", | |
| "title": "On the use of finite state transducers for semantic interpretation", | |
| "authors": [ | |
| { | |
| "first": "Christian", | |
| "middle": [], | |
| "last": "Raymond", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Frederic", | |
| "middle": [], | |
| "last": "Bechet", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Renato", | |
| "middle": [ | |
| "De" | |
| ], | |
| "last": "Mori", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Geraldine", | |
| "middle": [], | |
| "last": "Damnati", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2006, | |
| "venue": "Speech Communication", | |
| "volume": "48", | |
| "issue": "", | |
| "pages": "288--304", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Christian Raymond, Frederic Bechet, Renato De Mori, and Geraldine Damnati. 2006. On the use of finite state transducers for semantic interpretation. Speech Communication, 48,3-4:288-304.", | |
| "links": null | |
| }, | |
| "BIBREF7": { | |
| "ref_id": "b7", | |
| "title": "Language models for speech recognition and understanding", | |
| "authors": [ | |
| { | |
| "first": "Giuseppe", | |
| "middle": [], | |
| "last": "Riccardi", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "L", | |
| "middle": [], | |
| "last": "Allen", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Gorin", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1998, | |
| "venue": "Proceedings of the International Conference on Spoken Langage Processing (ICSLP)", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Giuseppe Riccardi and Allen L. Gorin. 1998. Language models for speech recognition and understanding. In Proceedings of the International Conference on Spo- ken Langage Processing (ICSLP), Sidney, Australia.", | |
| "links": null | |
| }, | |
| "BIBREF8": { | |
| "ref_id": "b8", | |
| "title": "Is word error rate a good indicator for spoken language understanding accuracy?", | |
| "authors": [ | |
| { | |
| "first": "Ye-Yi", | |
| "middle": [], | |
| "last": "Wang", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Acero", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "C", | |
| "middle": [], | |
| "last": "Chelba", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2003, | |
| "venue": "Automatic Speech Recognition and Understanding workshop -ASRU'03", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Ye-Yi Wang, A. Acero, and C. Chelba. 2003. Is word error rate a good indicator for spoken language under- standing accuracy? In Automatic Speech Recognition and Understanding workshop -ASRU'03, St. Thomas, US-Virgin Islands.", | |
| "links": null | |
| }, | |
| "BIBREF9": { | |
| "ref_id": "b9", | |
| "title": "Spoken language understanding", | |
| "authors": [ | |
| { | |
| "first": "Ye-Yi", | |
| "middle": [], | |
| "last": "Wang", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Li", | |
| "middle": [], | |
| "last": "Deng", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Alex", | |
| "middle": [], | |
| "last": "Acero", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2005, | |
| "venue": "Signal Processing Magazine", | |
| "volume": "22", | |
| "issue": "", | |
| "pages": "16--31", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Ye-Yi Wang, Li Deng, and Alex Acero. 2005. Spoken language understanding. In Signal Processing Maga- zine, IEEE, volume 22, pages 16-31.", | |
| "links": null | |
| }, | |
| "BIBREF10": { | |
| "ref_id": "b10", | |
| "title": "Partially observable markov decision processes for spoken dialog systems", | |
| "authors": [ | |
| { | |
| "first": "Jason", | |
| "middle": [ | |
| "D" | |
| ], | |
| "last": "Williams", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Steve", | |
| "middle": [], | |
| "last": "Young", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2007, | |
| "venue": "Computer, Speech and Language", | |
| "volume": "21", | |
| "issue": "", | |
| "pages": "393--422", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Jason D. Williams and Steve Young. 2007. Partially ob- servable markov decision processes for spoken dialog systems. Computer, Speech and Language, 21:393- 422.", | |
| "links": null | |
| } | |
| }, | |
| "ref_entries": { | |
| "TABREF2": { | |
| "content": "<table/>", | |
| "html": null, | |
| "type_str": "table", | |
| "text": "", | |
| "num": null | |
| }, | |
| "TABREF3": { | |
| "content": "<table><tr><td>: Word Error Rate (WER), Concept Error</td></tr><tr><td>Rate (CER) and Interpretation Error Rate (IER) ac-</td></tr><tr><td>cording to the SLU strategy</td></tr></table>", | |
| "html": null, | |
| "type_str": "table", | |
| "text": "", | |
| "num": null | |
| } | |
| } | |
| } | |
| } |