• master_thesis_proposal


Mapping open-domain ASR output to dialogue systems grammars


Build a dialogue systems module which, given a grammar defining the strings that the system can understand, takes input from an open online ASR (e.g. Google) and maps it onto the (phonetically) nearest string generated by the grammar.

This thesis project will be carried out within Talkamatic's EU PF7 Alfred project.

Problem description

Open ASR is available over the internet, but the results are hard to use with dialogue systems with limited language understanding capabilities. Often, ASR output contains errors caused by the ASR not knowing the vocabulary of the domain which the system can deal with. The task of this project is to come up with innovative and practically useful ways of mapping ASR output to the nearest sentence (or sentences) produced by a grammar.

As a resource, the student will have a Wizard-of-Oz corpus collected in EU FP7 project Alfred, containing ASR output and transcribed speech (to be mapped to nearest in-grammar sentence).

Some ideas towards possible solutions (there may well be other, better ideas!):

  • See it as a machine translation problem?
  • Store memory of human corrections, cached as FSTs for quick application?
  • Text simplification algorithms using Integer Linear Programming

Recommended skills

Python, GF, machine translation/ILP/other method.


  • Staffan Larsson, Christos Koniaris (FLoV, GU)
  • External supervisor from Talkamatic AB.

Sub-corpus topic modeling and Swedish litterature

The goal of the Master thesis will be to: i) use/process a large Swedish text collection ii) experiment and apply topic modeling and consequently sub-corpus topic modeling (according the description by Tangherlini & Leonard, 2013) iii) adapt or create a visual, web based environment to explore the results (this will be done in various ways, preferably as a) network graphs (Smith et al., 2014); se for instance figure 1 and integrated them in b) a web based exploratory environment, such as a dashboard; se figure 2)

For more see the link below:


Analytics tools for dialogue systems


To improve the performance of dialogue systems, analyses of interactions are an important source of knowledge. When dialogue systems are deployed, the interaction can be logged for later analysis. Talkamatic AB are about to begin collection of logs resulting from dialogue system interactions, which will be available to the student.


The task is to build a toolbox for analysing logs of dialogue system interactions and presenting results from such analyses. Examples of relevant analysis features include:

  • speech recognition error % (requires transcribing user utterances)
  • task completion time
  • number of turns
  • dialogue complexity (e.g. number of subdialogues, degree of subdialogue embedding)
  • number of grounding subdialogues
  • estimated task success (measured by observing behaviour after system response)

Such basic analysis dimensions can then be used to detect and diagnose problems with systems which need fixing.


Staffan Larsson, Dialogue Technology Lab, and Talkamatic AB.

A poor man's approach to integrating POS-tagging and parsing.

A poor man's approach to integrating POS-tagging and parsing.

In the by now traditional NLP processing setup, part-of-speech tagging and syntactic parsing are separate, ordered tasks. A sentence is first POS-tagged, after which the results are used as the input for parsing. This model is convenient because it allows one to use a more efficient technique for the "simpler" task of POS-tagging and it helps to keep search space in the expensive parsing task down.

On the downside, however, we note that a POS-tagger is missing out on possibly beneficial syntactic information -- POS-tagging precedes parsing and therefore syntactic information cannot be used to choose between alternative tag sequences. In turn, we can expect parsing to suffer from a resulting decreased accuracy in POS-tagging.

Indeed, in PCFG-based parsing, parsing and POS-tagging has long been one and the same processing step. More recently, in the data-driven dependency parsing literature, algorithms for combined parsing and POS-tagging have been proposed, and they have been shown to lead to improved results.

In this project, you will investigate a simpler, more general approach to integrating POS-tagging and parsing, by letting the POS-tagger and the parser entertain multiple hypothesis about the analysis of a sentence, from which the most best analysis can then be chosen. This way one can achieve a free flow of information between the two processes -- hopefully improving accuracy -- without having to radically change the NLP setup (POS-tagging still precedes parsing). Existing tools can be used with very little alteration, which makes it a poor man's solution.

As part of the project, you will investigate, design and implement different ways of realizing the setup sketched above, and present experiments showing the impact of your choices on analysis accuracy and efficiency.

This MA project combines theoretical aspects (literature study and design of a realization of the system outlined above), implementation, and empirical study (evaluation of the system).

Programming skills and an NLP background are a prerequisite. Knowledge of statistical methods is a big plus, as is affinity with the linguistic side of processing, as this will allow you to do more insightful error analysis. Since the development material will be Swedish text, some passive knowledge of the Swedish language is assumed.

The project would be supervised by Gerlof Bouma and Richard Johansson, Yvonne Adesam or possibly others at Språkbanken.

Situated learning agents (2016)

Situated agents must be able to interact with the physical environment that are located in with their conversational partner. Such an agent receives information both from its conversational partner and the physical world which it must integrate appropriately. Furthermore, since both the world and the language are changeable from one context to another it must be able to adapt to such changes or to learn from new information. Embodied and situated language processing is trying to solve challenges in natural language processing such as word sense disambiguation and interpretation of words in discourse as well as it gives us new insights about human cognition, knowledge, meaning and its representation. Research in vision relies on information represented in natural language, for example in the form ontologies, as this captures how humans partition and reason about the world. On the other hand, gestures and sign language are languages that are expressed and interpreted as visual information.

The masters thesis could be undertaken independently or as an extension of an existing project from the Embodied and Situated Language Processing (ESLP) course. Experience with dialogue systems and good Python programming skills is a plus.

Several projects are available subject of the approval of the potential supervisors. The main thread of the research would be how a linguistically inquisitive robot can update its representation of the world by engaging in dialogue conversation with a human. Sensory observations of a robot may be incomplete due to errors that robot's sensors or actuators introduce or simply because the robot has not explored and mapped the entire world yet. Can a robot query a human about the missing knowledge linguistically with clarification questions? Robotic view of the world is quite different from that of a human. How we can find a mapping between the representations that a robot builds using its sensors and the representations that are a result of human take on the world? The latter is challenging but necessary if robots and humans were to have a meaningful conversation.

Here are some suggested tasks:

A Lego robot, a miniature environment with blocks in a room

  • Online linguistic annotation of objects and situations that a robot discovers "Please tell me. What is this? And this?"
  • The ability to reason about the discovered objects (i.e. creating more complex propositions from simple ones) using some background knowledge "Aha, this is a chair... so I would expect to find a table here as well."
  • Extracting the ontological information used in the previous task from external text resources (e.g. Wikipedia).

Microsoft Kinect or Microsoft robot studio, a table situation with objects

  • Learning of spatial relations between objects on the table in interaction with humans (using, for example, the Attentional Vector Sum Model of Regier and Carlson)
  • Integrating and testing the effects of adding non-spatial features (the influence of dialogue and the knowledge about the objects) in the learning model.

Generating route descriptions in a complex building

  • How to generate route descriptions that provide the right kind of information so that a person finds the objects or location referred to?
  • Using a map of a complicated building (DG4) and representation of salient features in the building build a computational model that would generate such descriptions.
  • Connect that system with a dialogue system and explore the interaction of referring expressions with the structure and the content of dialogue.

Grounded meaning representations

  • Work towards a novel model of grounded meaning representations and validate them in an experiment such as Roy (2002) and others
  • How can information from vector space models be integrated with perceptual information?
  • What are good and effective models of information fusion: interaction between different dimensions of meaning, for example, how to incorporate world knowledge with perceptional meaning to deal with spatial cognition cases described in the work by Coventry and our own work

Earlier project (which this project could build on)


Simon Dobnik and other members of the Dialogue Technology Lab; for extracting ontological information also members of the Text Technology Lab

Building a sentiment lexicon for Swedish

The goal of this project is the semi-automatic construction of a sentiment lexicon for Swedish. For more information see link below.


Adding valency information to a dependency parser

The goal of this project is to improve a Swedish dependency parser by integrating a valency lexicon. For more information see link below.


Part-of-speech tagging/syntactic parsing of emergent texts

The goal of this project is to implement a part-of-speech tagger and investigate the possibilities of developing a syntactic parser that could handle emergent text, i.e. texts – or representations of texts – that are being produced (and thus frequently changed) in order to identify the syntactic location of for example pauses. For more information see link below.


Historical Text reuse (in Swedish Literature) (2016)

The goal of the Master thesis will be to apply (implement or adapt) techniques (e.g., borrowed from the field of bioinformatics) to identify lexically-similar passages (i.e. phrases, sentences, quotes, paraphrases) across collections of Swedish literary texts. Such techniques can use any suitable algorithms for that purpose, but preferably sequence alignment and present/visualize the results in a user friendly (and navigable) way.

Collocations for learners of Swedish

Collocations for learners of Swedish


Generate a list of collocations, phrasal verbs, set phrases and idioms important for learners of Swedish, linked to proficiency levels, for use in Lärka.


The currently developed application Lärka, www.spraakbanken.gu.se/larka, is intended for computer-assisted language learning of L2 Swedish. Lärka generates a number of exercises based on corpora available through Korp, one of them focusing on vocabulary. It has been mentioned on several occasions that we should include multi-word expressions into our exercise generator. This also complies with the CEFR “can-do” statements at different levels of proficiency (http://www.coe.int/t/dg4/linguistic/Source/Framework_en.pdf). It is, however, a non-trivial task to identify the items that should be included into the curriculum, and even more uncertain how the selected items can be assigned to different proficiency levels.

Problem description

The aims of this work are the following:

  • to study literature on collocations etc. in general and in the L2 context especially, paying special attention to the CEFR guidelines; to make an overview of the practices for training collocations etc. used in other applications and in (online) dictionaries/lexicons
  • to generate a list of collocations, (primarily) by automatic analysis of COCTAILL - a corpus of coursebook texts used for teaching Swedish. Study of different materials available outside COCTAILL, e.g. books written by Anna Hallström, multi-word expressions in Saldo and Lexin, may also prove to be beneficial, however, the challenge would be to define at which level these items should be introduced. To get some inspiration, have a look at English Vocabulary Profile: http://vocabulary.englishprofile.org/staticfiles/about.html (user: englishprofile, password: vocabulary)
  • (potentially) to implement one or more of the suggested exercise formats as web services + user interface in Lärka
  • evaluate/test on users (language learners, teachers, linguists, etc)

Recommended skills:

  • Python
  • interest in Lexical Semantics and Second Language Acquisition


  • Elena Volodina
  • potentially others from Språkbanken/FLOV