|
Size: 6830
Comment:
|
Size: 12407
Comment:
|
| Deletions are marked like this. | Additions are marked like this. |
| Line 6: | Line 6: |
| ''Internal note to thesis supervisors: you need to make sure that each supervised thesis appears on our chair homepage. For this you need to: 1) put the final PDFs of the thesis and the presentation into /nfs/raid3/publications/theses - please follow the existing naming convention!; the PDFs will be available at [[http://ad-publications.informatik.uni-freiburg.de/theses/|http://ad-publications.informatik.uni-freiburg.de/theses/]]<PDF> 2) send a mail to Sabine with the following information: title of thesis, name of student, thesis abstract, link to thesis PDF, link to presentation PDF.'' |
|
| Line 8: | Line 14: |
| ''Note to interested students: if all projects in this list are ongoing, this means that our current capacity for supervising projects is reached. Maybe, if you come back later, there will be an offer again. You can also propose a topic on your own.'' | ''Note to interested students: if all projects in this list are ongoing, this means that our current capacity for supervising projects and theses is reached. Maybe, if you come back later, there will be an offer again. You can also propose a topic on your own.'' |
| Line 10: | Line 16: |
| TODO: project + thesis offered/supervised by Björn | '''Deep Learning for NLP (master project or thesis, ONGOING)''': We have several NLP problems (e.g., relation classification) in which it would be interesting to apply recent deep learning techniques. As a project the task would mainly be to re-implement some existing work and compare it to (simpler) baselines. As a thesis this would extend to improving performance using new ideas. Ideally, you already know about neural networks and have implemented and/or trained them. You should, however, know the basics of machine learning and in particular have practical experience.'' supervised by [[http://ad.informatik.uni-freiburg.de/staff/haussmann|Elmar Haussmann]]'' |
| Line 12: | Line 18: |
| '''Information Extraction from Noun Compounds (thesis, available)''': Noun compounds like ''US president Barack Obama'' or ''3rd century Roman writer Censorinus'' can carry a lot of valuable information. Here: ''Barack Obama is president of the US'' and ''Censorinus is a roman writer from the 3d century''. Noun compounds have been studied from a linguistic point of view, e.g. http://people.ischool.berkeley.edu/~nakov/selected_papers_list/TSLP2013.pdf , but so far they have largely been ignored in information extraction tasks (i.e. generate triples of the expressed facts like (''Barack Obama'') (''is president of'') (''the US'')). The task of this thesis will be to research current state-of-the art approaches for understanding noun compounds and implement/adapt/improve them to extract information in triple-form. Good if you know something about machine learning but not absolutely necessary. ''supervised by Elmar Haussmann'' | '''Deep Dive (project, ONGOING)''': [[http://deepdive.stanford.edu|DeepDive]] is a framework from the Stanford University that can construct a knowledge base (a database of facts like "Barack Obama" has-spouse "Michelle Obama") by analyzing large text corpora. It applies a large variety of state-of-the-art NLP techniques in distant-supervision and probabilistic inference. The project's goal would be 1) to evaluate the framework by applying it to a large text corpus 2) incorporate advanced methods or features to improve/enhance extraction quality. You should be confident in working in a Linux environment and you need practical experience in programming. ''supervised by [[http://ad.informatik.uni-freiburg.de/staff/haussmann|Elmar Haussmann]]'' |
| Line 14: | Line 20: |
| '''Semantic Search Engine Prototype utilizing OpenIE (thesis or larger project, available)''': Open Information Extraction (OpenIE) extracts triples like (''Barack Obama'') (''is president of'') (''the US'') from natural language text. Recent OpenIE systems perform well, but it is yet to be shown that these triples can be used effectively for information retrieval tasks. First systems searching on the extracted triples already exist, e.g. http://openie.cs.washington.edu/. The task will be to investigate how to best search on OpenIE triples, implement a search engine prototype similar to the existing one and asses/evaluate it. You should be proficient in programming in C++ or Java and it is good if you heard the information retrieval lecture. ''supervised by Elmar Haussmann'' | '''Movie Search (thesis, ongoing)''': Our semantic search engine [[http://broccoli.cs.uni-freiburg.de|Broccoli]] answers semantic queries. Queries can, in theory, be very complex and powerful. However, it is almost impossible for inexperienced users to build appropriate structured queries. Goal of this thesis is to apply and adapt the search paradigm behind Broccoli to a narrow domain. Think of a search for movies where you can search by all the common things, i.e. actors, the director, year, genre, but also by keyword search in the plot. However, for plot search, Broccoli features can come in handy. Looking for the movie where the charcter played by Di Caprio has drug problems? Linking the characters in the plot to knowledge-base entries can help a lot. For this thesis you will deal with many aspects that make the thesis far more researchy than technical. What is a good set of features to support? How can you achive a good user experience? Where to get good data about movies (structured and plots)? Does it make sense to take general knowledge data (like that from Broccoli) and link it to the data about the movies? It may be necessary at some point to add new operations to Broccoli. It would be good if you were experienced in (or willing to pick up) multiple programming languages. Between data integration, user interfaces and Brococli server code, it really makes sense to pick a suitable languge for each job. ''supervised by [[http://ad.informatik.uni-freiburg.de/staff/buchhold|Björn Buchhold]]'' '''Efficient code for (de)compression (project, ongoing)''': Our semantic search engine [[http://broccoli.cs.uni-freiburg.de|Broccoli]] and its successor (under development) work with compressed lists a lot. Current implementations are fast but far from optimal and have never been compared to other available compression schemes. Goal of the project is to tune the current (C++) implementation for speed, to evaluate the literature for promising alternatives and to compare them. ''supervised by [[http://ad.informatik.uni-freiburg.de/staff/buchhold|Björn Buchhold]]'' |
| Line 17: | Line 25: |
| '''Distributional Semantics for a Semantic Search Engine Prototype (project, available)''': Distributional semantics is a hot topic in natural language processing right now. It allows representing words (and phrases) as vectors in a high-dimensional space such that words with similar meaning have low distance, e.g. "Friday" and "Saturday", "king" and "queen", or "Freiburg" and "Stuttgart". Tools to generate these vectors are freely available, e.g. from Google: https://code.google.com/p/word2vec/. The task will be to generate word vectors, assess their quality, and (depending on quality) implement a prototype search engine utilizing them, e.g. for resolving synonyms. Good if you heard the information retrieval lecture, but not absolutely necessary. ''supervised by Elmar Haussmann'' | ## '''Value Recognition in Full-Text (thesis, available)''': Text documents contain various values of different types. Dates, weights, lengths, heights and many more can be expressed obviously like ''13kg'', less obivously, e.g., ''ten tons'', and highly obfuscated, e.g., ''for the first half of the year''. For our semantic search engine [[http://broccoli.cs.uni-freiburg.de|Broccoli]], we do not make use of this inforamtion, yet. The taks of the thesis is to produce a ([[http://uima.apache.org/|UIMA]]) component that extracts such values from text documents, that serve as input for further steps to enable semantic search that matches such values. Additionally the findings should be evaluated. Facts which kinds of values are truely hard to extract and which are relatively easy, is very valuable knowledge as well. - ''supervised by [[http://ad.informatik.uni-freiburg.de/staff/buchhold|Björn Buchhold]]'' |
| Line 19: | Line 27: |
| ## '''Setting up a browser for DBPedia with Broccoli (project, available)''' We derived a simplified version of the [[http://www.freebase.com/|Freebase]] ontology and set up a web application to browse the data based on our Broccoli search engine, called Freebase Easy: http://freebase-easy.cs.uni-freiburg.de . The goal of this project would be to do something similar with the [[http://dbpedia.org/|DBPedia]] ontology. This means modifying the freely downloadable dataset so that it 1) can be processed by our pipeline and 2) has a form that allows for a comfortable browsing with our user interface (e.g. providing unique and proper human readable names for instances and types) ... ''supervised by [[http://ad.informatik.uni-freiburg.de/staff/baeurlef|Florian Bäurle]]'' | |
| Line 20: | Line 29: |
| TODO: project + thesis offered/supervised by Florian | ## '''A keyword query translator for Broccoli (bachelor thesis, ONGOING)''' Our Broccoli search engine has its own special query language (SPARQL-like trees). Though the userinterface guides users in the incremental query creation process, many users got accustomed to simple keyword queries like they are used, e.g., when searching with Google. To make the interface more attractive for these users and for the general convenience of a quick query creation, it would be nice to have a mechanism that translates normal keyword queries into equivalent structured Broccoli queries. We already implemented an experimental mechanism that can make such translations (e.g., just try entering "mafia films by Francis Coppola" into the input field of the userinterface) as a proof of concept. The work of this thesis would be to implement a better, more powerful query translator ... ''supervised by [[http://ad.informatik.uni-freiburg.de/staff/baeurlef|Florian Bäurle]]'' |
| Line 22: | Line 31: |
| TODO: project + thesis offered/supervised by Claudius | ## '''Usability study for the Broccoli user interface (bachelor thesis, available)''' We developed a special user interface for the prototype of our semantic full-text search engine Broccoli: http://broccoli.cs.uni-freiburg.de . An open taks is yet to make a thorough user study to evaluate the usability of the interface, compare it to other search interfaces and identify weak points that could be improved ... ''supervised by [[http://ad.informatik.uni-freiburg.de/staff/baeurlef|Florian Bäurle]]'' |
| Line 24: | Line 33: |
| TODO: project + thesis offered/supervised by Sabine | '''Question Answering using Semantic Full-Text Search (thesis, available)''': Use our semantic full-text search engine Broccoli for the FACTOID and LIST questions from one of the TREC Questions Answering benchmarks, e.g. http://trec.nist.gov/data/qa/2006_qadata/QA2006_testset.xml . What is the performance of manually constructed queries, and what can be done to improve it? Good if you heard the information retrieval lecture before, but not absolutely necessary ... ''supervised by [[http://ad.informatik.uni-freiburg.de/staff/bast|Hannah Bast]] and [Elmar or Björn]'' |
| Line 26: | Line 35: |
| '''Question Answering using Semantic Full-Text Search (thesis, available)''': Use our semantic full-text search engine Broccoli for the FACTOID and LIST questions from one of the TREC Questions Answering benchmarks, e.g. http://trec.nist.gov/data/qa/2006_qadata/QA2006_testset.xml . What is the performance of manually constructed queries, and what can be done to improve it? Good if you heard the information retrieval lecture before, but not absolutely necessary ... ''supervised by Hannah Bast and TBA'' | '''An easy-to-use web app for the !CompleteSearch engine (project, available)''': The !CompleteSearch engine complex search capabilities (including prefix search and faceted search) on semi-structured data (text and databases) with very fast query times. Our current software is powerful and flexible, but has quite a learning curve before it can be used. The goal of this project would be to set up a web app, where one can upload any CSV dataset, and then have a convenient search (with meaningful default settings), without having to set up anything oneself. This would be an extremely useful web application ... ''supervised by [[http://ad.informatik.uni-freiburg.de/staff/bast|Hannah Bast]]'' |
| Line 28: | Line 37: |
| '''Improved Transfer Patterns Routing (thesis or larger project, available)''': Public transit routing at Google currently implements the transfer patterns algorithm from this paper: http://ad-publications.informatik.uni-freiburg.de/ESA_transferpatterns_BCEGHRV_2010.pdf . The evaluation in this paper is outdated, and we know of several possible major improvements in the meantime. In particular, we know how to accelerate the pre-processing time by a factor of about 10. The goal of this thesis would we to implement a much improved version of transfer patterns and properly benchmark it ... ''supervised by Hannah Bast and Sabine Storandt'' | ''' PDF to TeX Parser (project or theses, available)''': Per default, you can't rearrange or edit texts in PDF files. As long as you don't have the TeX sources of a PDF, there is no chance to edit the PDF: While creating a PDF from TeX file(s) is a quite common feature, the other way around - recovering the TeX source(s) from given PDF(s) - isn't. The goal of this project is to implement a parser that produces TeX sources from given PDF files to be able to edit the sources and to recompile them to a new PDF ... '' also available as master thesis if combined with the topic "Formula recognition in PDF files" (see below)'' ... ''supervised by [[http://ad.informatik.uni-freiburg.de/staff/korzen|Claudius Korzen]]'' |
| Line 30: | Line 39: |
| '''An easy-to-use web app for the !CompleteSearch engine (project, available)''': The !CompleteSearch engine complex search capabilities (including prefix search and faceted search) on semi-structured data (text and databases) with very fast query times. Our current software is powerful and flexible, but has quite a learning curve before it can be used. The goal of this project would be to set up a web app, where one can upload any CSV dataset, and then have a convenient search (with meaningful default settings), without having to set up anything oneself. This would be an extremely useful web application ... ''supervised by Hannah Bast'' | ''' Formula recognition in PDF files (project or theses, available)''': Per default, text and formulas in a PDF file are "drawn" into the PDF without any semantic markups. This makes it difficult to understand the meaning of a formula on extracting text / formulas from PDF files. The goal of this project is to implement a parser, that identifies and "understands" the formulas in a given PDF by translating the formulas back to their correspondent TeX commands. ... ''also available as master thesis if combined with the topic "PDF to TeX Parser" (see above)'' ... ''supervised by [[http://ad.informatik.uni-freiburg.de/staff/korzen|Claudius Korzen]]'' |
| Line 35: | Line 44: |
M.Sc. thesis Max Lotstein (Elmar) B.Sc. thesis Manuel Ruder (Elmar) M.sc. thesis Niklas Meinzer (Sabine) B.Sc. thesis Christian Ehrenfeld (Hannah) |
Internal note for our chair: each of us should offer / provide 12 months of supervision every year. Each thesis or project "counts" just like it counts for the students, where #months = #ECTS / 4. That is, B.Sc. project = 6 ECTS = 1.5 months, M.Sc. project = 16 ECTS = 4 months, B.Sc. thesis = 12 ECTS = 3 months, M.Sc. thesis = 25 ECTS = 6 months.
Internal note to thesis supervisors: you need to make sure that each supervised thesis appears on our chair homepage. For this you need to: 1) put the final PDFs of the thesis and the presentation into /nfs/raid3/publications/theses - please follow the existing naming convention!; the PDFs will be available at http://ad-publications.informatik.uni-freiburg.de/theses/<PDF> 2) send a mail to Sabine with the following information: title of thesis, name of student, thesis abstract, link to thesis PDF, link to presentation PDF.
List of available and ongoing topics for current B.Sc. and M.Sc. projects and theses
Note to interested students: if all projects in this list are ongoing, this means that our current capacity for supervising projects and theses is reached. Maybe, if you come back later, there will be an offer again. You can also propose a topic on your own.
Deep Learning for NLP (master project or thesis, ONGOING): We have several NLP problems (e.g., relation classification) in which it would be interesting to apply recent deep learning techniques. As a project the task would mainly be to re-implement some existing work and compare it to (simpler) baselines. As a thesis this would extend to improving performance using new ideas. Ideally, you already know about neural networks and have implemented and/or trained them. You should, however, know the basics of machine learning and in particular have practical experience. supervised by Elmar Haussmann
Deep Dive (project, ONGOING): DeepDive is a framework from the Stanford University that can construct a knowledge base (a database of facts like "Barack Obama" has-spouse "Michelle Obama") by analyzing large text corpora. It applies a large variety of state-of-the-art NLP techniques in distant-supervision and probabilistic inference. The project's goal would be 1) to evaluate the framework by applying it to a large text corpus 2) incorporate advanced methods or features to improve/enhance extraction quality. You should be confident in working in a Linux environment and you need practical experience in programming. supervised by Elmar Haussmann
Movie Search (thesis, ongoing): Our semantic search engine Broccoli answers semantic queries. Queries can, in theory, be very complex and powerful. However, it is almost impossible for inexperienced users to build appropriate structured queries. Goal of this thesis is to apply and adapt the search paradigm behind Broccoli to a narrow domain. Think of a search for movies where you can search by all the common things, i.e. actors, the director, year, genre, but also by keyword search in the plot. However, for plot search, Broccoli features can come in handy. Looking for the movie where the charcter played by Di Caprio has drug problems? Linking the characters in the plot to knowledge-base entries can help a lot. For this thesis you will deal with many aspects that make the thesis far more researchy than technical. What is a good set of features to support? How can you achive a good user experience? Where to get good data about movies (structured and plots)? Does it make sense to take general knowledge data (like that from Broccoli) and link it to the data about the movies? It may be necessary at some point to add new operations to Broccoli. It would be good if you were experienced in (or willing to pick up) multiple programming languages. Between data integration, user interfaces and Brococli server code, it really makes sense to pick a suitable languge for each job. supervised by Björn Buchhold
Efficient code for (de)compression (project, ongoing): Our semantic search engine Broccoli and its successor (under development) work with compressed lists a lot. Current implementations are fast but far from optimal and have never been compared to other available compression schemes. Goal of the project is to tune the current (C++) implementation for speed, to evaluate the literature for promising alternatives and to compare them. supervised by Björn Buchhold
Question Answering using Semantic Full-Text Search (thesis, available): Use our semantic full-text search engine Broccoli for the FACTOID and LIST questions from one of the TREC Questions Answering benchmarks, e.g. http://trec.nist.gov/data/qa/2006_qadata/QA2006_testset.xml . What is the performance of manually constructed queries, and what can be done to improve it? Good if you heard the information retrieval lecture before, but not absolutely necessary ... supervised by Hannah Bast and [Elmar or Björn]
An easy-to-use web app for the CompleteSearch engine (project, available): The CompleteSearch engine complex search capabilities (including prefix search and faceted search) on semi-structured data (text and databases) with very fast query times. Our current software is powerful and flexible, but has quite a learning curve before it can be used. The goal of this project would be to set up a web app, where one can upload any CSV dataset, and then have a convenient search (with meaningful default settings), without having to set up anything oneself. This would be an extremely useful web application ... supervised by Hannah Bast
PDF to TeX Parser (project or theses, available): Per default, you can't rearrange or edit texts in PDF files. As long as you don't have the TeX sources of a PDF, there is no chance to edit the PDF: While creating a PDF from TeX file(s) is a quite common feature, the other way around - recovering the TeX source(s) from given PDF(s) - isn't. The goal of this project is to implement a parser that produces TeX sources from given PDF files to be able to edit the sources and to recompile them to a new PDF ... also available as master thesis if combined with the topic "Formula recognition in PDF files" (see below) ... supervised by Claudius Korzen
Formula recognition in PDF files (project or theses, available): Per default, text and formulas in a PDF file are "drawn" into the PDF without any semantic markups. This makes it difficult to understand the meaning of a formula on extracting text / formulas from PDF files. The goal of this project is to implement a parser, that identifies and "understands" the formulas in a given PDF by translating the formulas back to their correspondent TeX commands. ... also available as master thesis if combined with the topic "PDF to TeX Parser" (see above) ... supervised by Claudius Korzen
Completed B.Sc. or M.Sc. projects and theses
[Die Liste ist noch unvollständig, insbesondere fehlen gerade noch die ganzen Projekte. Titel der Arbeit sollte da auch noch stehen und ein Link zur jeweiligen Webseite bzw. Arbeit und Präsentation. Und das Anfangs- und Enddatum.]
M.Sc. thesis Max Lotstein (Elmar)
B.Sc. thesis Manuel Ruder (Elmar)
M.sc. thesis Niklas Meinzer (Sabine)
B.Sc. thesis Christian Ehrenfeld (Hannah)
B.Sc. thesis Philipp Bausch (Elmar)
M.Sc. thesis Eugen Sawin (Hannah)
M.Sc. thesis Patrick Brosi (Hannah + Sabine)
B.Sc. thesis Marius Bethge (Björn)
M.Sc. thesis Cynthia Jimenez (Sabine)
M.Sc. thesis Jonas Sternisko (Hannah)
M.Sc. thesis Ragavan Natarajan (Florian)
B.Sc. thesis Benjamin Meier (Claudius)
M.Sc. thesis Susanne Eichel (Hannah)
B.Sc. thesis Axel Lehmann (Hannah)
B.Sc. thesis Adrian Batzill (Hannah)
B.Sc. thesis Anton Stepan (Björn)
M.Sc. thesis Mirko Brodesser (Hannah + Sabine)
M.Sc. thesis Manuel Braun (Hannah + Sabine)
B.Sc. thesis Robin Schirrmeister (Hannah)
B.Sc. thesis Simon Skilevic (Hannah)
M.Sc. thesis Ilinca Tudose (Hannah + Elmar)
B.Sc. thesis Christiane Schaffer (Florian)
M.Sc. thesis Dirk Kienle (Hannah)
B.Sc. thesis Ina Baumgarten (Hannah + Björn)
B.Sc. thesis Niklas Meinzer (Hannah + Björn)
M.Sc. thesis Claudius Korzen (Hannah)
M.Sc. thesis Florian Bäurle (Hannah)
M.Sc. thesis Elmar Haußmann (Hannah)
M.Sc. thesis Oliver Mitevski (Hannah + Marjan)
M.Sc. thesis Björn Buchhold (Hannah)
Diploma thesis Johannes Schwenk (Hannah)
B.Sc. thesis Mirko Brodesser (Hannah + Marjan)