Intent Generation for Goal-Oriented Dialogue Systems based on Annotations

Intent Generation for Goal-Oriented Dialogue Systems based on Annotations

Umutcan Şimşek 0000-0001-6459-474XUniversity of InnsbruckTechnikerstrasse 21aInnsbruckAustria6020  and  Dieter Fensel University of InnsbruckTechnikerstrasse 21aInnsbruckAustria6020

Goal-oriented dialogue systems typically communicate with a backend (e.g. database, Web API) to complete certain tasks to reach a goal. The intents that a dialogue system can recognize are mostly included to the system by the developer statically. For an open dialogue system that can work on more than a small set of well curated data and APIs, this manual intent creation will not scalable. In this paper, we introduce a straightforward methodology for intent creation based on semantic annotation of data and services on the web. With this method, the Natural Language Understanding (NLU) module of a goal-oriented dialogue system can adapt to newly introduced APIs without requiring heavy developer involvement. We were able to extract intents and necessary slots to be filled from annotations. We were also able to create a set of initial training sentences for classifying user utterances into the generated intents. We demonstrate our approach on the NLU module of a state-of-the art dialogue system development framework.

intent generation, goal-oriented dialog systems, semantic annotations,, actions
copyright: rightsretaineddoi: 0.0/000_0isbn: 123-4567-24-567/08/06conference: First International Workshop on Chatbot in conjunction with ICWSM 2018; June 2018; Stanford, Californiajournalyear: 2018article: Xprice: XX.XXccs: Information systems RESTful web servicesccs: Information systems Web data description languagesccs: Information systems Ontologiesccs: Applied computing Annotationccs: Computing methodologies Discourse, dialogue and pragmatics

1. Introduction

Unlike more conversation oriented, human-human interaction mimicking chatbots, goal-oriented dialogue systems typically converse with humans based on defined tasks (Jurafsky and Martin, 2017). From a natural language understanding point of view, these tasks are connected to the domain specific intents that can be identified from user utterances. Goal-oriented dialogue systems generally work with well curated back-ends either by means of querying a database or sending requests to an API that is coupled with the dialogue system. This situation naturally makes it harder to adapt dialogue systems to different back-end systems.

After almost 30 years of its invention, the web is finally becoming more machine-oriented. This is thanks to the increasing amount of semantic annotations published on the web. The de-facto standard vocabulary for publishing semantic annotations is the vocabulary, supported by the initiative consists of Bing, Google, Yahoo! and Yandex search engines. The vocabulary contains types and properties for various domains to describe entities on the web. This semantically annotated data can be consumed by agents like dialogue systems.

A goal-oriented dialogue system should be as generic as possible to be able to operate in a heterogeneous environment like the web. In other words, it should be decoupled from the back-end. This requires not only the data but also the web services to be described semantically, so the tasks and consequently the intents that the dialogue system supports can be extracted from the web service descriptions. We consider open and flexible goal-oriented dialogue systems that can utilize different web services with the minimal human intervention is a natural step towards completing complex tasks (e.g. e-commerce) in the more machine-oriented web (Şimşek and Fensel, 2018).

As a first step in this direction, we propose a straightforward approach for extracting intents from lightweight semantic web services described with vocabulary. In this paper we will briefly demonstrate how this vocabulary can be used to annotate Web APIs and how these annotations can be beneficial for generating intents for goal-oriented dialogue systems.

The remainder of this paper structured as follows: Section 2 gives a short review of the existing efforts in this direction. Section 3 exemplifies the usage of vocabulary for annotating Web APIs. Section 4 explains our approach in detail and Section 5 demonstrates a use case with a state-of-the art dialogue system development framework. We conclude the paper in Section 6 with a summary and identified future directions.

2. Related Work

Dialogue systems benefited from semantic technologies, especially ontologies for a long time to represent domain knowledge in a powerful way (Milward and Beveridge, 2003). The semantic technologies are also essential elements of question-answering systems that work over linked data (Unger et al., 2014). As mentioned in the introduction, for goal-oriented dialogues systems that go beyond question answering, semantic annotation of the data is not enough. In order to decouple a dialogue system from the Web APIs, the services should be annotated too. The functional and behavioural description of Web APIs can potentially guide generation of dialogue flows as also explained in our recent work (Şimşek et al., 2018). Benefiting from semantic web service descriptions for a dialogue has been explored in the literature by the SmartWeb (Sonntag et al., 2007) project. They use a rule-based semantic parser (Engel, 2005) to convert user utterances to a set of ontology instances. These instances are then used for querying web services described with OWL-S (Martin et al., 2007). From a natural language understanding point of view, there is no need to identify the user intent, since it is a question answering system and the only intent is to obtain some information. As for a system that can handle multiple tasks, using the backend as a driver for the natural language understanding has been proposed. A previous work demonstrated that the intents can be modelled from a backend perspective (Berg et al., 2011). The work in (Berg, 2015) adopts such a backend driven approach, but there is no semantic descriptions involved, meaning the intents that represent different tasks should be handcrafted by the dialogue system developer.

If we consider a dialogue system as a whole, from language understanding to response generation, there has been a increasing amount of work towards using machine learning for creating such end-to-end dialogue systems. Such systems have big advantages as they do not require any apriori knowledge, which allows them to scale well. Such systems give promising results especially for chatting oriented dialogue systems. There is also promising development regarding such dialogue systems in a more goal oriented setting (Bordes et al., 2017) (Constantin et al., 2018).

Given the developments in the machine learning for dialogue systems, approaching the NLU challenge as a classification challenge is appropriate (Pappu and Rudnicky, 2013). For industrial applications, it is very common to use dialogue system development frameworks. These frameworks also benefit from machine learning approaches for identifying the user intention based on the utterance. The classification is done with supervision, therefore the classifiers require annotated natural language statements for training. As our final goal is to generate dialogue systems based on semantically annotated data and web services, we first start with enabling NLU modules to classify user intents with as little human intervention as possible. Increasing adoption of vocabulary gives as a strong motivation to use the vocabulary for data and web service annotations. We utilize semantic annotations in following ways: (a) to extract the intent and slots to be filled based on the high level description of a resource of a Web API (b) creating training sentences based on domain specific data stored in a knowledge graph. Although it has some limitations at the moment, we argue the work described in the following section will be a good step towards automatically generating goal-oriented dialogue systems simply by crawling a website annotated with and actions.

3. Web API Annotation with actions have been added to the core vocabulary in 2014. The main idea is to be able describe not only static entities on the web, but also the actions that can be taken on them. These actions in principle may be used as a lightweight semantic web services vocabulary for describing Web APIs. We analyzed actions in the scope of lightweight semantic web services. The details of this analysis is outside of the scope of this paper and we refer the reader to our work in (Şimşek et al., 2018). In this section we show what such an API annotating may look like and what kind of implications does it have for generating intents for goal-oriented dialogue systems.

Figure 1 shows an example action annotation to describe a resource to search for hotel room offers in Feratel API111Feratel is a destination management solution provider and offers an API for e-commerce of touristic services such as accommodation.. An action describes a resource of an API and a high level operation that can be applied to that resource. We can formalize an action as a quintuple


where is the type of an action, is the set of all type values of the object of the action, is the set of all type values of the result of the action, and and set of input and output parameters. Figure 1 is missing the target property, which is not relevant for a dialogue system in terms of intent generation, but it only concerns the invocation of an action over a resource as an HTTP request. An HTTP request to the described resource may need values for the input parameters represented with <property name>-input properties and the API returns a set of entities as response possibly with a schema:potentialAction attached to them. These entities must contain at least the values for the properties described with the <property name>-output properties.

The API descriptions created with actions have interesting implications for intent generation for dialogue systems. One thing should be known that does not have a strong formal semantics (Patel-Schneider, 2014). This means that the meaning of concepts is conveyed with natural language. (e.g. name and the description of the concept). We try to use such semantic information to first extract a specific intent for the operation an action represents (e.g. searching a hotel room), then by using the semantics embedded in the action annotation we will try to generate training sentences for the extracted intent. Next section explains this process in detail.

Figure 1. A partial JSON-LD representation of a schema:SearchAction created to annotate Feratel API

4. Intent Generation

Before we go deeper in the intent generation, let us first define the term intent. An intent is the desire to complete an action. In goal-oriented dialogue systems, user utterances may carry intents regarding a certain task. An intent contains an act, an object and modifiers for that act. The act carries a communicative function for the intent. Acts can represent generic functions such as information seeking (Bunt et al., 2010) but they can also be domain specific as they are in goal-oriented dialogue systems. An object of the intent is on what the act is carried out. An act can have certain modifiers that are commonly known as frames or slots in dialogue system literature. To explain the utterance elements with an example, let’s take the following sentence: ”What are the events in Innsbruck?”. Here the generic act is information seeking carried on event objects. The act is modified or filtered with an additional information, namely the location of the object. Since a goal-oriented system typically communicates with a backend, we can use a more specific act such as searching, which is a more backend specific version of information seeking.

Currently, many state of the art NLUs use machine learning approaches for classification of intents. Therefore alongside aforementioned elements, namely act, object type and modifiers, we define one more element, which is the set of training sentences. We formalize an intent as a quadruple:


where is the act, is the set of all object types of the intent, S is the set of training sentences and M is the set of modifiers or slots of the intent. A modifier is formalized as a triple , where is the name of the modifier, is the expected value type of the modifier and is a boolean value whether the modifier is required for the intent.

In the rest of this section we will describe how we generate intents by processing annotations. First, we describe how to collect and store the annotations in a knowledge graph to be used as the domain specific information for the generation of intents and training sentences. Then we briefly explain out straightforward methodology for generating intents. At last we show how we benefit from lexical databases and word vectors for automatically generating training sentences for the generated intents.

4.1. Annotation Collection

In order to reach our goal of generating intents, we collect the annotated data and web service descriptions from a website. The collection can happen automatically through a wrapper that generates annotations from a backend (e.g. relational database, Web API) or directly by crawling the website. Then we store these annotations of data and services in our knowledge graph. This way we benefit from reasoning capabilities of the graph database solution. offers its own semantics built on RDFS. This means some RDFS entailment rules (e.g. inheritance) can be applied to the annotations given the fact that the vocabulary is also stored in the knowledge graph. Moreover, defines its own semantics, rather informally. For instance for inverse properties, they define an inverseOf property, whose semantics is explained with a natural language description. Such custom semantics can be introduced to the knowledge graph in terms of rules. The annotations stored here later can be used for different purposes. The action annotations are used for extracting intents and data annotations to help NLU for recognizing entities in user utterances and for training machine learning models for generating training sentences for the intents. In the next sections we will explain this processes in detail.

4.2. Extracting Intents from Web API Descriptions

When dealing with goal-oriented dialogue systems, the intents the dialogue system supports are directly related to the capabilities of the back-end system. If the dialogue system talks with Web APIs to complete certain tasks, then the intents are created by the dialogue system developer based on the API resources and possible operations defined on those resources. We argue that this intent generation process can be automated with the help of machine readable API descriptions, in our case with

for all  do
     i = getName()
     i = getObjectTypes()
     for all  do
         add(getModifier(), )
     end for
     add(i, I)
end for
Algorithm 1 Algorithm for extracting intents

The pseudocode in Algorithm 1 summarizes the intent generation process. We iterate over the set of action annotations and extract the name of the action as well as the types of the object values. Then we create modifiers for the intent based on the input parameters of the action. The newly created intent with the act, object type and modifiers then added to the set of intents. For example, for the action represented in Figure 1, the algorithm would lead to the intent shown in Equation 3.


The NLU modules typically need entity definitions to identify entities in the user utterances that are needed to query a database system or make an API call. These entities are usually introduced to the system by the dialogue system developer. We use semantic technologies and load the entities to the NLU based on the actions. We collect the supported entities that the NLU module needs by generating SPARQL queries to run against the knowledge graph based on the members of the set . Figure 2 shows the SPARQL query to populate the entities that NLU module needs for recognizing the values for modifier. The vocabulary allows the data to be represented in several different ways. For instance, a hotel room can be connected to a hotel with schema:containedInPlace property or inversely, with schema:containsPlace property. Since the data is stored in a knowledge graph with an inference engine, the query would return all schema:amenityFeature values given that the knowledge graph contains the entailment rule for schema:inverseOf property. Moreover, the query can be optimized by querying only the named graph where the annotations for a certain website stored.

Note that at the moment the intent is generated, the set of training sentences is an empty set. We explain the creation of training sentences for an intent in Section 4.3.

Figure 2. The SPARQL query generated to retrieve hotel amenities from the knowledge graph

4.3. Automated Training Sentence Creation

v = verbs
n = nouns
m = modifiers
for all  do
     add(v, relevantWords(getSynonyms()))
     add(n, relevantWords(getSynonyms()))
     add(m, )
      = buildSentences(loadCFG(), v, n, m)
end for
Algorithm 2 Algorithm for generating training sentences

The vocabulary does not have strong formal semantics(Patel-Schneider, 2014), therefore to generate training sentences automatically the only thing we can rely on is the semantics hidden in the natural language descriptions of the types and properties. For the sake of simplicity, we used a small context free grammar (CFG) to build our initial training sentences. Algorithm 2 shows our steps to generate training sentences for each intent extracted from action descriptions, based on the act, object types and modifiers.

One challenge we encountered while creating training sentences is to put the words in a context. As we utilize the verbs and nouns appearing in the action annotations and consequently in the intents, we tried to find the relevant words. Our first attempt was to find synonyms to create varieties of training sentences of an intent. However certain words can have many different meanings as well as synonyms. This was a big issue while finding the synonyms from the WordNet lexical database 222 In order to only select the context relevant synonyms, we applied the Dice coefficient (Dice, 1945) as a string similarity measurament to description of types and the WordNet synoynms of the act () and object types () in order to find the relevant synoynms. The Dice coefficient splits the strings into bigrams. As shown in Equation 4, is the bigrams that occur in the both strings, and represents the total number of bigrams in both strings.


Making such a filtering helps us to avoid situations like creating a training sentence such as ”I am looking for a case” for ”searching for an event” intent. Alongside the synonyms, we also tried the incorporate similar verbs. For instance in the context of a goal-oriented dialogue system, ”search” and ”find” can be classified in the same intent from a backend perspective. Therefore we tried different similarity metrics. WordNet similarity, specifically Wu-Palmer () metric (Wu and Palmer, 1994), which uses a zero to one scale for similarity scores . Additionally, we tried to find similar words based on generic and domain specific usage. For that we used two word vector embeddings. One of them is the ConceptNet Numberbatch (Speer et al., 2017). These vectors are trained based on existing ConceptNet 5.5 knowledge graph333 data, word2vec embeddings (Mikolov et al., 2013) trained with 300B words from Google News Dataset and GloVe embeddings (Pennington et al., 2014) trained with 840B words from CommonCrawl444 The hybrid ConceptNet Numberbatch pre-trained vectors perform better than the aforementioned vectors in different evaluations (Speer et al., 2017).

procedure relevantWords(, )
      set of relevant words from ConceptNet Vectors
      set of relevant words from Domain Specific Vectors
      = conceptnetVec.mostSimilar()
      = dsVec.mostSimilar()
     for all  do
         for all  do
              if conceptnetVec.similarity() ¿  then
              end if
         end for
     end for
end procedure
Algorithm 3 Algorithm for finding relevant words

The vector embeddings consider the context of the words while learning vectors. By applying vector operations, information like word similarities, analogies can be obtained. In order to take domain specific factors into account, we trained additional vector embeddings as Fasttext vectors (Bojanowski et al., 2017) with descriptions of entities in the knowledge graph with the type and its supertypes. For example, for an intent to search hotels, we trained the vectors with description of hotels and its supertypes like lodging businesses. This gave us around 2.5M words. Typically vector embeddings work well with billions words in terms of training data, but we use this vector only as a supporting factor to include some verbs that the more generic ConceptNet Numberbatch embbeddings may be missing. Fasttext uses ngrams as the smallest units instead of words, this actually gives us the advantage of calculating similarity scores for words that do not exist in the vocabulary, which is more likely in a small vocabulary like 2.5M. In Algorithm 3, we show how similarity calculations from generic vector embeddings (conceptnetVec) and domain specific vector embeddings (dsVec) are incorporated. The relevantWords function takes all the synonyms of a word as parameter. We first find the similar words in conceptnetVec filtered by the part-of-speech tag. Then, we compare the similar words from dsVec with the ones found in conceptnetVec and include them into our relevant words list. If the similarity score exceeds the threshold. The threshold value 0.5 has been picked as an initial value after some manual observations. This threshold value can be picked more intelligently in the future. Table 1 shows some example word pairs and their similarity scores according to different calculations. When we compare the verbs search and find, the vector similarity () gives a better score then Wu-Palmer similarity ().

In this section, we presented a method for creating initial training sentence for extracted intents from action annotations. We utilized the act and object types of an intent. Then we enriched our word pool with synonyms and similar words we obtained from different sources. This allows us to build not only sentences like ”search a hotel room”, but also ”find a hotel room”, even though search and find are not synonyms but they have a certain semantic relationship. In the next section we will show how the generated training sentences and intents look when loaded to a state-of-the-art NLU .

Word Pair
search-find 0.67 0.33
search-need 0.51 0.33
search-look_around 0.57 0.33
Table 1. Similarity score comparisons between vector embeddings and WordNet (Wu-Palmer)

5. Use Case: Generating Intents for DialogFlow

Since the last couple of years, the frameworks that enable dialogue system development gained popularity. There are now many examples of such frameworks like DialogFlow555, Wit.ai666 and relatively new Amazon Lex777 The advantange of these frameworks is that they provide a lot of out-of-the-box features to ease many aspects of dialogue system development such as natural language understanding and dialogue management. They also offer mechanisms to integrate the bots with several applications like Facebook Messenger or Slack, which helped to gain a vast industrial adoption in a rather short period of time. Therefore in this section, we decided to demonstrate our approach on the NLU modules of one of these frameworks, namely DialogFlow. The other frameworks in the same direction also operate in a similar way, therefore our implementation should work with other frameworks as well as custom developed systems with similar principles with minimal effort.

In Figure 1, we showed a schema:SearchAction created for a real hotel booking API. Based on that action, we extracted the intent as shown in Equation 3. In this section, we will present how does the extracted intent look in DialogFlow. In DialogFlow, a basic intent has a name, training sentences and slots that are needed to be filled to fullfil the intent. Given the intent , we choose a naming scheme such as , where . That means, for the intent in Equation 3, search.HotelRoom-LodgingReservation is generated as the intent name. The slots are also generated based on extracted from schema:SearchAction. For example, in the case of the intent in Equation 3, the name of the slot for check-in time is generated as object.LodgingReservation.checkinTime, which additionally indicates the JSON path of the input property, which is useful for the backend logic that fulfills the intent. As for the training sentences, the Algorithm 2 uses the act and name of the object types as well as a sample of possible values for each modifier from the knowledge-graph. If value cannot be found in the knowledge graph, then a placeholder is created instead. For the intent in Equation 3 a list of possible training sentences is shown in Listing 1.

1I search for a hotel room
2We look for a hotel room
3search for a hotel room,
4We search for a lodging reservation
5We want to search a hotel room
6find a hotel room with free wifi
7find a hotel room with wellness
8find a hotel room on 27.05.2018
9looking for an accommodation
10I want to discover a lodging reservation
11rout up a hotel room
12We are seeking  a hotel room
13I explore a lodging reservation
14I am looking around for a hotel room for 4
15I want to search for a hotel room from 27.05.2018 to 30.05.2018
16I want to look for a hotel room
17We are looking  a hotel room
18We want to explore a hotel room
Listing 1: A sample of generated training sentences

The sentences are built automatically based on the results of the algorithms described in Section 4.3. The small context free grammar is used for creating possible grammatical variations for the generation. At the moment, the prepositions required for inserting the modifiers are statically specified in the grammar (e.g. Lines 5-8). While the verbs ”search, look for, explore” are retrieved from WordNet, find, rout up, look around” are retrieved from the vector embeddings. Similarly, word accommodation is retrieved from the vector embeddings as it is related to the phrase ”hotel room”. There are some generated sentence that do not make too much sense, such as ”I want to discover a lodging reservation”, but these do not make too much of a difference while classifying user utterances by intents.

Figure 3. Generated intents, slots and training sentences loaded to DialogFlow

Figure 3 shows the generated intent and training sentences added to DialogFlow. For implementation specific reasons, we created a mapping of certain data types from (e.g. schema:Number, schema:Date) to built-in datatypes of DialogFlow. The intents are loaded to DialogFlow via the API programmatically.

In this section, we demonstrated our approach with a state-of-the-art dialogue system development framework. The NLU module requires annotated training sentences for user utterance classification, therefore we provided automatically generated natural language sentences to train the machine learning model in the background.

6. Conclusion

As the web evolves to be a more machine-readable platform, the task of consumption of the content uncovers new challenges. Goal-oriented dialogue systems are prominent agents to consume this machine readable content. However, creating a dialogue system that can adapt to new data and services is not trivial, since every web service requires different workflows for completing certain tasks. This reflects to dialogue systems in terms of dialogues, which means that the dialogue systems should be able to guide the user with an appropriate dialogue according to the web service workflow. Towards this direction, we argue that the first step is that a dialogue system understands what user wants to do. To that end, we developed and demonstrated an approach for generating intents based on the lightweight semantic web service descriptions with vocabulary. To achieve our goal, we also utilized a knowledge graph to benefit from the domain specific knowledge. Given the fact that many state-of-the-art dialogue system development frameworks use supervised machine learning for NLU, we also introduced a method for generating training sentences for the intents. At the moment, we can generate variety of sentences by utilizing lexical databases and word vector embeddings, however our method still has its limitations. The main limitation is the generation of meaningful sentences with modifier values. We see this problem as a text imputation or sentence completion problem and currently investigating solutions with LSTM neural networks for the future work, as well as better ways to incorporate sparse domain specific text for training the ML models. Another limitation is the selection of threshold value for word similarity. We plan to find a better way to find this threshold value based on the statistical analysis of the data.

We see the work explained in this paper as a contribution also towards automated response generation, which is one of the next steps for automated generation of dialogue systems based on web service descriptions. Although we showed some results of this approach, the real evaluation would be to see (a) if the generated sentences make sense linguistically (b) how a dialogue system with automatically generated intents perform in comparison to a dialogue system with manually created intents. In the future work, we will address the aforementioned limitations and evaluate the dialogue system with automatically generated intents from natural understanding point of view888The demo bot can be found here: with experiments involving real users in real use cases like e-tourism.


The authors would like to thank the members of the team (Richard Dvorsky, Thibault Gerrier, Roland Gritzer, Philipp Häusle, Omar Holzknecht, Elias Kärle and Dennis Sommer) for fruitful discussions and implementation support.


  • (1)
  • Berg et al. (2011) Markus Berg, Bernhard Thalheim, and Antje Düsterhöft. 2011. Dialog acts from the processing perspective in task oriented dialog systems. In Proceedings of the 15th Workshop on the Semantics and Pragmatics of Dialogue (SemDial 2011). Citeseer, Los Angeles, CA, 176–177.
  • Berg (2015) Markus M. Berg. 2015. NADIA: A Simplified Approach Towards the Development of Natural Dialogue Systems. In Natural Language Processing and Information Systems, Chris Biemann, Siegfried Handschuh, André Freitas, Farid Meziane, and Elisabeth Métais (Eds.). Springer International Publishing, Cham, 144–150.
  • Bojanowski et al. (2017) Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. 2017. Enriching Word Vectors with Subword Information. Transactions of the Association for Computational Linguistics 5 (2017), 135–146.
  • Bordes et al. (2017) Antoine Bordes, Y-Lan Boureau, and Jason Weston. 2017. Learning End-to-End Goal-Oriented Dialog. In The 5th International Conference on Learning Representations (ICLR 2017). 1–15.
  • Bunt et al. (2010) Harry Bunt, Jan Alexandersson, Jean Carletta, Jae-Woong Choe, Alex Chengyu Fang, Koiti Hasida, Kiyong Lee, Volha Petukhova, Andrei Popescu-Belis, Laurent Romary, Claudia Soria, and David Traum. 2010. Towards an ISO standard for dialogue act annotation. In The Proceedings of Seventh International Conference on Language Resources and Evaluation (LREC). European Language Resources Association, Valletta, Malta, 2548–2555.
  • Constantin et al. (2018) Stefan Constantin, Jan Niehues, and Alex Waibel. 2018. An End-to-End Goal-Oriented Dialog System with a Generative Natural Language Response Generation. CoRR abs/1803.02279 (2018). arXiv:1803.02279
  • Dice (1945) Lee R. Dice. 1945. Measures of the Amount of Ecologic Association Between Species. Ecology 26, 3 (1945), 297–302.
  • Engel (2005) Ralf Engel. 2005. Robust and efficient semantic parsing of free word order languages in spoken dialogue systems. In In Proceedings of 9th Conference on Speech Communication and technology.
  • Jurafsky and Martin (2017) Dan Jurafsky and James H. Martin. 2017. Dialog Systems and Chatbots. In Speech and Language Processing (3 ed.). Unpublished Draft, Chapter 28, 418–440.
  • Martin et al. (2007) David Martin, Mark Burstein, Drew McDermott, Sheila McIlraith, Massimo Paolucci, Katia Sycara, Deborah L. McGuinness, Evren Sirin, and Naveen Srinivasan. 2007. Bringing Semantics to Web Services with OWL-S. World Wide Web 10, 3 (Aug. 2007), 243–277.
  • Mikolov et al. (2013) Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. 2013. Efficient Estimation of Word Representations in Vector Space. CoRR abs/1301.3781 (2013). arXiv:1301.3781
  • Milward and Beveridge (2003) David Milward and Martin Beveridge. 2003. Ontology-Based Dialogue Systems. In Proceedings of the 3rd IJCAI Workshop on Knowledge and Reasoning in Practical Dialogue Systems. Acapulco, Mexico, 9–18.
  • Pappu and Rudnicky (2013) Aasish Pappu and AI Rudnicky. 2013. Predicting Tasks in Goal-Oriented Spoken Dialog Systems using Semantic Knowledge Bases. In Proceedings of the 14th Annual Meeting of the Special Interest Group on Discourse and Dialogue. Metz, France, 242–250.
  • Patel-Schneider (2014) Peter F Patel-Schneider. 2014. Analyzing Springer International Publishing, 261–276.
  • Pennington et al. (2014) Jeffrey Pennington, Richard Socher, and Christopher D. Manning. 2014. GloVe: Global Vectors for Word Representation. In Empirical Methods in Natural Language Processing (EMNLP). 1532–1543.
  • Şimşek and Fensel (2018) Umutcan Şimşek and Dieter Fensel. 2018. Now We Are Talking! Flexible and Open Goal-Oriented Dialogue Systems for Accessing Touristic Services. e-Review of Tourism Research 9, ENTER 2018 Research Notes (2018).
  • Şimşek et al. (2018) Umutcan Şimşek, Elias Kärle, and Dieter Fensel. 2018. Machine Readable Web APIs with Actions. In the Proceedings of 14th International Conference on Semantic Systems (To Appear). Vienna.
  • Sonntag et al. (2007) Daniel Sonntag, Ralf Engel, Gerd Herzog, Alexander Pfalzgraf, Norbert Pfleger, Massimo Romanelli, and Norbert Reithinger. 2007. SmartWeb Handheld - Multimodal Interaction with Ontological Knowledge Bases and Semantic Web Services. In Artifical Intelligence for Human Computing: ICMI 2006 and IJCAI 2007 International Workshops, Banff, Canada, November 3, 2006, Hyderabad, India, January 6, 2007, Revised Selected and Invited Papers (LNAI 4451), Thomas S. Huang, Anton Nijholt, Pantic Pantic, and Alex Pentland (Eds.). Springer Berlin Heidelberg, 272–295.
  • Speer et al. (2017) Robert Speer, Joshua Chin, and Catherine Havasi. 2017. ConceptNet 5.5: An Open Multilingual Graph of General Knowledge. In The Proceedings of AAAI Conference on Artificial Intelligence 2017. 4444–4451.
  • Unger et al. (2014) Christina Unger, André Freitas, and Philipp Cimiano. 2014. An Introduction to Question Answering over Linked Data. Springer International Publishing, Cham, 100–140.
  • Wu and Palmer (1994) Zhibiao Wu and Martha Palmer. 1994. Verb Semantics and Lexical Selection. CoRR abs/cmp-lg/9406033 (1994). arXiv:cmp-lg/9406033
Comments 0
Request Comment
You are adding the first comment!
How to quickly get a good reply:
  • Give credit where it’s due by listing out the positive aspects of a paper before getting into which changes should be made.
  • Be specific in your critique, and provide supporting evidence with appropriate references to substantiate general statements.
  • Your comment should inspire ideas to flow and help the author improves the paper.

The better we are at sharing our knowledge with each other, the faster we move forward.
The feedback must be of minimum 40 characters and the title a minimum of 5 characters
Add comment
Loading ...
This is a comment super asjknd jkasnjk adsnkj
The feedback must be of minumum 40 characters
The feedback must be of minumum 40 characters

You are asking your first question!
How to quickly get a good answer:
  • Keep your question short and to the point
  • Check for grammar or spelling errors.
  • Phrase it like a question
Test description