Biblioscape is designed with one goal in mind, that is to make the life of researchers easier. After 4 major releases, Biblioscape has evolved from a traditional bibliographic software into a Research Information Manager. Biblioscape will help researchers to get all kinds of information organized in a single place, and link them together to build a knowledge base. It consists of 7 modules addressing different aspects of a researcher's needs.
References module is for storing, managing, and searching for bibliographic references. Bibliographic records from different sources can be imported into a Biblioscape database with the right import filter. References are organized into folders. Several searching tools are provided to query the database.
BiblioWord is an easy to use word processor tightly integrated with the References module. It is also a bibliography maker that can convert temporary citations in a document into formatted citations and bibliography. If you prefer to use Word or WordPerfect, Biblioscape also provide tight integration with these two popular word processors.
Internet module can be used to search online bibliographic database via a web browser. With a single button click, your web search results can be captured into a Biblioscape database. Besides capturing bibliographic records, you can also use the Internet module to capture web pages.
Notes module is designed to collect any free text information that does not fit into the reference database. A note can be your ideas, comment, background information about an author, etc. Notes are organized in a tree structure. A note can be linked to other notes, references, tasks, etc.
Tasks module is a simple To Do list manager that is integrated with the References and Notes modules. It is designed to manage tasks related to your research, so you won't need a separate program for it. Tight integration with other modules of Biblioscape make it an ideal tool to manage research related tasks.
Charts module can be used to draw flow charts, organization charts, etc. A chart object can be linked to the References, Notes, Tasks, and Library modules. You can draw a chart to express ideas and procedures in your research, and use SQL to connect objects in your chart to other modules in Biblioscape.
Library module is for managing a small research library. It could be a researcher's personal library, a department library, even a small corporate library. It includes 7 sub-modules to handle different tasks in library automation. These are: Catalog, Serials, Circulation, Interlibrary Loan, Borrowers, Lenders, and Suppliers.
Biblioscape also includes a web server application, BiblioWeb. With just one button click, your bibliographic database can be published on the Web. Web users can be assigned Read or Write privileges to browse, search, even add and delete bibliographic records using a Web browser. This is the easiest way for a research group to share a common bibliographic database on the Web.
So wonderful software!!
2004年9月15日
2004年9月14日
The homework of Pattern Classification
There were only four questions of Chapter two of Pattern Classification. The first one was nearly a pure mathematical problem to prove some inequalities. After some reasoning, it was transformed to be discussion of some curve trend.
The second one was integration of some mathematical formula. There was a subsection function which should be drawn. I used the fplot function of Matlab for drawing it. The final picture was wonderful. The third one was simple. Using the proving skill of pigeonhole theorem, the conclusion could be proved easy.
The most difficult one was the final one. Lots of linear algebra knowledge was in essence not clearly in my mind. I found that I had not been very familiar with the mathematical knowledge of Advanced Mathematics, Linear Algebra, and Probability Theory. This was a serious problem. I should review them at some time.
The second one was integration of some mathematical formula. There was a subsection function which should be drawn. I used the fplot function of Matlab for drawing it. The final picture was wonderful. The third one was simple. Using the proving skill of pigeonhole theorem, the conclusion could be proved easy.
The most difficult one was the final one. Lots of linear algebra knowledge was in essence not clearly in my mind. I found that I had not been very familiar with the mathematical knowledge of Advanced Mathematics, Linear Algebra, and Probability Theory. This was a serious problem. I should review them at some time.
2004年9月13日
The result of LSI
Just now I had realized the LSI approach in my Summarization Evaluation task. But the final result was not good. I decided the K value of 5 in the LSI method. But the whole trend was not same as that of the five students.
Whether is it a prove that good summarization did not mean high coreference coefficient. The detail situation must be discussed with others.
Whether is it a prove that good summarization did not mean high coreference coefficient. The detail situation must be discussed with others.
2004年9月12日
Latent Semantic Indexing
Latent Semantic Indexing(LSI) is one method using for text categorization as improvement to VSM. The idea behind the LSI is to map each document and test vector into lower dimensional space which is associated with concepts and compare the documents in this space.
In Dr.Tliu's courseware about Information Retrieval, I saw some example about it. And after reading some papers which were downloaded from CNKI about LSI, I believed it could be used in my evaluation task. I should discuss the situation with Dr.Tliu and senior.
There were some homeworks of our curriculums. I must come back and prepare for them now.
In Dr.Tliu's courseware about Information Retrieval, I saw some example about it. And after reading some papers which were downloaded from CNKI about LSI, I believed it could be used in my evaluation task. I should discuss the situation with Dr.Tliu and senior.
There were some homeworks of our curriculums. I must come back and prepare for them now.
2004年9月11日
Summarization Evaluation1
After some discussion with Yhb and Dr.Tliu, I began to realize the first Summarization Evaluation method. It was based on the thoughts of Documents Classification.
Just now I had finished the programs and analysis the trend of the data. But the effect was bad. There were 100 groups human evaluation scores, but only three had the same trend with my results.
When I came back to my bedroom, I began to think for the reasons. The evaluation method was based on the document vectors using tf*idf values and calculating the similarity between the vectors. This method was considering only the words frequency without more information. But the four summarization systems were not only considering the words frequency. So my methods should be updated.
But how to update? The basic features that I could bethink of were used in the current summarization systems. If I use some of them, the final evaluating system couldt ake sides some systems. So I must use some other methods. What was the correct methods? This was a big problem for me currently.
Just now I had finished the programs and analysis the trend of the data. But the effect was bad. There were 100 groups human evaluation scores, but only three had the same trend with my results.
When I came back to my bedroom, I began to think for the reasons. The evaluation method was based on the document vectors using tf*idf values and calculating the similarity between the vectors. This method was considering only the words frequency without more information. But the four summarization systems were not only considering the words frequency. So my methods should be updated.
But how to update? The basic features that I could bethink of were used in the current summarization systems. If I use some of them, the final evaluating system couldt ake sides some systems. So I must use some other methods. What was the correct methods? This was a big problem for me currently.
2004年9月10日
Teacher's Day!
This is teacher's day! I wish all teachers will have more wonderful life and health.
This morning, we sent some little presents to our best teachers. And when our six students sent our present to our Graduate English Teacher, she was at home alone. She was a Buddhist and told lots of stories to us. In her words, I comprehended one thing. Innovation was one thing appearing in your mind at some moment by some stimulate. But this was based on your ample knowledge accumulation.
Teacher is wonderful career. Blessing them!
This morning, we sent some little presents to our best teachers. And when our six students sent our present to our Graduate English Teacher, she was at home alone. She was a Buddhist and told lots of stories to us. In her words, I comprehended one thing. Innovation was one thing appearing in your mind at some moment by some stimulate. But this was based on your ample knowledge accumulation.
Teacher is wonderful career. Blessing them!
2004年9月9日
《模式分类》第一章学习体会
刚才学习完毕《模式分类》的第一章《绪论》，现在不得不说这本书是到目前为止我见到过的关于机器学习最好的教材之一。第一章《绪论》中列出了许多模式分类同时也是机器学习中存在的问题。现列举并讨论如下：
模式分类的最终目的和处理方法就是，首先将模型分为几类，然后对感知到的数据进行处理，以滤除干扰（由采样引起而非模型引起）。然后，选择出与感知数据最接近的模型类别。
在模式分类的过程中往往需要看绿一个分类的“总体代价”函数。我们真正的任务是要确定一种决策（decision），使该函数最小。这是决策理论的中心任务，而模式分类可能是其中最重要的一个子领域。
模式分类需要抽取待分类模式的各种特征，特征选择的多少和如何选择特征是一个非常重要的问题。以往我的观点和做法是将能够抽取到的所有特征都送给决策树使用，这样得到的分类器的封闭测试的效果非常好，但是往往“推广能力”（就是开放测试时正确区分模式的能力，亦称“泛化能力”）会较差（这种现象叫做“过度拟合”）。同时特征选取过多可能会给将来在非常高维德空间中进行分类操作埋下了“祸根”。“推广能力”和“复杂度”需要进行折中，二者互为矛盾，但是折中到什么程度需要进一步的讨论。
从根本上说分类判决任务实面向特定任务或特定代价的。例如，假如我们的目的是销售鱼子（酱）的话，我们很可能试图按照鱼的性别进行分类，把雄的和雌的分开。或者，我们想把受损的鱼筛选出（以制备猫食）等等。不同的判决任务将需要不同的特征，其判别边界也与鱼分类问题很不相同。因此，建造一个通用的，能够精确的执行各种各样的分类任务的人工模式分类器将是一个极端困难的任务。这使得我们对人类能在各种模式分类任务中迅速和灵活的切换更加增加了几分赞美和敬佩之心。
模式分类的目标是寻找这样一种分类器，使得对于同一类别的样本尽可能认为一致，对于不同类的样本的区分尽可能明显。其中如何选择特征是至关重要的一步。在选择或设计特征的过程中，很明显，我们希望发现那些容易提取、对不相关变形保持不变、对噪声不敏感，以及对区分不同类别模式很有效的特征集。
本章小结中提到，模式识别的进展至少从以下三重意义上传达出积极的信息：（1）问题一定可以解决，因为人和生物体的识别能力是最好的“存在性证明”；（2）解决其中很多问题的数学理论已被发展起来；（3）还存在许多吸引人的未解决问题为进一步的研究发展提供了丰富的机遇。对比于自然语言处理，我认为同样可以得到三条积极信息：（1）人的自然语言的处理能力是自然语言处理中问题可解的强有力证明；（2）自然语言处理借助其他学科的成果和语言学特有的规律已经取得了长足的发展，但是现在自然语言处理的能力还非常有限，公认的理论基础建立的还不很完备；（3）自然语言处理中仍然存在大量的未解决的问题使得该领域中存在无数的机遇。可以说自然语言处理是一个不断摸索的领域，其中伴随着各种理论和方法的诞生必将推动其发展。
本书中最难部分是第九章《独立于算法的机器学习》。许多很微妙，而又至关重要的具有理论和实践意义的结论将被讨论。这其中包括偏差-方差关系、自由度问题、设计“简单”分类器的必要性，以及计算复杂度等问题。在某种意义上，只有懂得了本章的结论，才可能透彻的理解和更好的运用其他章节的知识。
如此好书，难求也。唯以踏实学习之！
模式分类的最终目的和处理方法就是，首先将模型分为几类，然后对感知到的数据进行处理，以滤除干扰（由采样引起而非模型引起）。然后，选择出与感知数据最接近的模型类别。
在模式分类的过程中往往需要看绿一个分类的“总体代价”函数。我们真正的任务是要确定一种决策（decision），使该函数最小。这是决策理论的中心任务，而模式分类可能是其中最重要的一个子领域。
模式分类需要抽取待分类模式的各种特征，特征选择的多少和如何选择特征是一个非常重要的问题。以往我的观点和做法是将能够抽取到的所有特征都送给决策树使用，这样得到的分类器的封闭测试的效果非常好，但是往往“推广能力”（就是开放测试时正确区分模式的能力，亦称“泛化能力”）会较差（这种现象叫做“过度拟合”）。同时特征选取过多可能会给将来在非常高维德空间中进行分类操作埋下了“祸根”。“推广能力”和“复杂度”需要进行折中，二者互为矛盾，但是折中到什么程度需要进一步的讨论。
从根本上说分类判决任务实面向特定任务或特定代价的。例如，假如我们的目的是销售鱼子（酱）的话，我们很可能试图按照鱼的性别进行分类，把雄的和雌的分开。或者，我们想把受损的鱼筛选出（以制备猫食）等等。不同的判决任务将需要不同的特征，其判别边界也与鱼分类问题很不相同。因此，建造一个通用的，能够精确的执行各种各样的分类任务的人工模式分类器将是一个极端困难的任务。这使得我们对人类能在各种模式分类任务中迅速和灵活的切换更加增加了几分赞美和敬佩之心。
模式分类的目标是寻找这样一种分类器，使得对于同一类别的样本尽可能认为一致，对于不同类的样本的区分尽可能明显。其中如何选择特征是至关重要的一步。在选择或设计特征的过程中，很明显，我们希望发现那些容易提取、对不相关变形保持不变、对噪声不敏感，以及对区分不同类别模式很有效的特征集。
本章小结中提到，模式识别的进展至少从以下三重意义上传达出积极的信息：（1）问题一定可以解决，因为人和生物体的识别能力是最好的“存在性证明”；（2）解决其中很多问题的数学理论已被发展起来；（3）还存在许多吸引人的未解决问题为进一步的研究发展提供了丰富的机遇。对比于自然语言处理，我认为同样可以得到三条积极信息：（1）人的自然语言的处理能力是自然语言处理中问题可解的强有力证明；（2）自然语言处理借助其他学科的成果和语言学特有的规律已经取得了长足的发展，但是现在自然语言处理的能力还非常有限，公认的理论基础建立的还不很完备；（3）自然语言处理中仍然存在大量的未解决的问题使得该领域中存在无数的机遇。可以说自然语言处理是一个不断摸索的领域，其中伴随着各种理论和方法的诞生必将推动其发展。
本书中最难部分是第九章《独立于算法的机器学习》。许多很微妙，而又至关重要的具有理论和实践意义的结论将被讨论。这其中包括偏差-方差关系、自由度问题、设计“简单”分类器的必要性，以及计算复杂度等问题。在某种意义上，只有懂得了本章的结论，才可能透彻的理解和更好的运用其他章节的知识。
如此好书，难求也。唯以踏实学习之！
订阅：
博文 (Atom)