For this assignment you will evaluate the performance of OpenCalais, a commercial entity extraction service, against your hand annotations.

1. Pick five random news stories and hand-annotate them. Pick an English-language news site with many stories on the home page, or a section of such a site (business, sports, etc.) Then generate five random numbers from 1 to the number of stories on the page. Cut and paste the text of each article into a separate  file, and save as plain text (no HTML, no formatting.)

2. Detect entities by hand in each article. Paste the text of each article into an RTF or Word document and go through it, underlining every entity. Count every mention of a person, place, or organization, including alternate names (“the president”) and pronoun references. Count how many entity references appear in each document (multiple mentions of the same entity all count.)

3. Now run each document through OpenCalais. You can paste the text into the demo page here. Now compare the results to your hand-annotations to produce a confusion matrix:

• True Positives: entities marked by you and found by OC
• False Negatives: entities marked by you and not found by OC
• False Positives: entities found by OC but not marked by you

4. Turn in:

• Your hand-marked documents.
• A spreadsheet. Please turn in a spread sheet with one row per document, and four columns: total entities marked by you, true positives, false negatives, false positives.
• Make sure your spreadsheet includes totals of these four columns across all documents
• Your analysis. Report on any patterns in the failures that you see. Where is OpenCalais most accurate? Where is it least accurate? Are there predictable patterns to the errors? Are there ambiguities as to what is really an entity?

For this assignment you will analyze a social network using three different centrality algorithms, and compare the results.

1. Download and install Gephi, a free graph analysis package. It is open source and runs on any OS.

2. Download the data file lesmis.gml from the UCI Network Data Repository.  This is a network extracted from the famous French novel Les Miserables — you may also be familiar with the musical and the recent movie. Each node is a character, and there is an edge between two characters if they appear in the same chapter. Les Miserables is written in over 300 short chapters, so two characters that appear in the same chapter are very likely to meet or talk in the plot of the book. The edges are weighted, and the weight is the number of chapters those characters appear together in.

3. Open this file in Gephi, by choosing File->Open. When the dialog box comes up, set the “Graph Type” type to “Undirected.” The graph will be plotted. What do you see? Can you discern any patterns?

4. Now arrange the nodes in a nicer way, by choosing the “Force Atlas 2″ layout algorithm from the Layout menu at left and pressing the “Run” button. When things settle down, hit the “Stop” button. The graph will be arranged nicely, but it will be quite small.  You can zoom in using the mouse wheel (or two fingers on the trackpad on a mac) and pan using the right mouse button.

5. Select the “Edit” tool from the bottom of the toolbar on the left. It looks like a mouse pointer with question mark next to it:

6. Now you can click on any node to see its label, which is the name of the character it represents. This information will appear in the “Edit” menu in the upper left. Here’s the information for the character Gavroche.

Click around the various nodes in the graph. Which characters have been given the most central locations? If you are familiar with the story of Les Miserables, how does this correspond to the plot? Are the most central nodes the most important characters?

7. Make Gephi color nodes by degree. Choose the “Ranking” tab from panel at the upper left, then select the “Nodes” tab, then “Degree” from the drop-down menu. Press the “Apply” button.

Now the nodes with the highest degree will be darker. Do these high degree nodes correspond to the nodes that the layout algorithm put in the center? Are they the main characters in the story?

8. Now make Gephi compute betweenness and closeness centrality by pressing the “Run” button for the Network Diameter option under “Network Overview” in to the right of the screen.

You will get a report with some graphs. Just click “Close”. Now betweenness and closeness centrality will appear in the drop-down under “Ranking,” in the same place where you selected degree centrality earlier, and you can assign colors based on either run by clicking the “Apply” button.

Also, the numerical values for betweenness centrality and closeness centrality will now appear in the “Edit” window for each node.

Select “Betweenness Centrality” from the drop-down meny and hit “Apply.” What do you see? Which characters are marked as important? How does it differ from the characters which are marked as important by degree?

Now select “Closeness Centrality” and hit “Apply.” (Note that this metric uses a scale which is the reverse of the others — closeness measures average distance to all other nodes, so small values indicate more central nodes. You may want to swap the black and white endpoints of the color scale to get something which is comparable to the other visualizations.) How does closeness centrality differ from betweeness centrality and degree? Which characters differ between closeness and the other metrics?

9. Which centrality algorithm would you prefer to use to understand the structure of Les Miserables? Why? How would you validate your choice if you didn’t already know the story? That is the situation a journalist is in when they analyze unknown data.

Turn in: your answers to the questions in steps 3, 6, 7, 8 and 9, plus screenshots for the graph plotted with degree, betweenness centrality, and closeness centrality. (To take a screenshot: on Windows, use the Snipping Tool. On Mac, press ⌘ Cmd + ⇧ Shift + 4. If you’re on Linux, you get to tell me)

What I am interested in here is how the values computed by the different algorithms correspond to the plot of Les Miserables (if you are familiar with it), and how they compare to each other. Telling me that “Jean Valjean has a closeness centrality of X” is not a high-enough level interpretation — your couldn’t publish that in a finished story, because your readers won’t know what that means.

Due before class on Wednesday, Nov 28

For this assignment you will design an information filtering algorithm. You will not implement it, but you will explain your design criteria and provide a filtering algorithm in sufficient technical detail to convince me that it might actually work — including psuedocode.

1. Decide who your users are. Journalists? Professionals? General consumers? Someone else?

2. Decide what you will filter. You can choose:

• Social network updates, like the Facebook news feed or Twitter trending topics
• A news organization recommendation engine
• The whole web, like Google News
• something else

3. List all available information that you have available as input to your algorithm. If you want to filter Facebook or Twitter, you may pretend that you are the company running the service, and have access to all posts and user data — from every user. You also also assume you have a web crawler or a firehose of every RSS feed or whatever you like, but you must be specific and realistic about what data you are operating with.

4. Argue for the design factors that you would like to influence the filtering, in terms of what is desirable to the user, what is desirable to the publisher (e.g. Facebook or a news organization), and what is desirable socially. Explain as concretely as possible how each of these (probably conflicting) goals might be achieved through in software. You can imagine that you have a UI that supports certain types of interactions (e.g. likes, votes, ratings) or encourages users to act in certain ways (e.g. following) that generate data for you.

5. Write psuedo-code for a function that produces a “top stories” list. This function will be called whenever the user loads your page or opens your app, so it must be fast and frequently updated. You can assume that there are background processes operating on your servers if you like. Your psuedo-code does not have to be executable, but it must be specific and unambiguous, such that I could actually go and implement it. You can assume that you have libraries for classic text analysis and machine learning algorithms. So, you don’t have to spell out algorithms like TF-IDF or item-based collaborative filtering, or anything else you can dig up in the research literature, but simply say how you’re going to use such building blocks. If you use an algorithm we haven’t discussed in class, be sure to provide a reference to it.

6. Write up steps 1-5. The result should be no more than three pages. You must be specific and plausible. You must be clear about what you are trying to accomplish, what your algorithm is, and why you believe your algorithm meets your design goals (though of course it’s impossible to know for sure without testing; but I want something that looks good enough to be worth trying.)

Due before class, October 10

This assignment is designed to help you develop a feel for the way topic modeling works, the connection to the human meanings of documents, and one way of handling the time dimension. First you will analyze a corpus of AP news articles. Then you’ll look at the State of the Union speeches, and report on how the subjects have shifted over time in relation to historical events.

Note: This assignment requires reading the documents, not just running algorithms on them. I am asking you to tell me how well these algorithms capture the meaning of the documents, and you can only determine meaning by reading the documents. So when you see a word scored highly within a document set, go read the documents that contain that word. When a topic ranks high for a document, go read the documents that contain that topic.

1. Load the data. To begin with you’ll try LDA on a homogeneous document set of short clean articles: this collection of AP wire stories. Get this CSV loaded as a list of strings, one per document.

2) Generate TF-IDF vectors. Use the gensim package to generate tf-idf weighted document vectors. Check out the gensim example code here. You will need to go through the file twice: once to generate the dictionary (the code snippet starting with “collect statistics about all tokens”) and then again to convert each document to what gensim calls the bag-of-words representation, which is un-normalized term frequency (the code snippet starting with “class MyCorpus(object)”

Note that there is implicitly another step here, which is to tokenize the document text into individual word features — not as straightforward in practice as it seems at first, but the example code just does the simplest, stupidest thing, which is to lowercase the string and split on spaces. You may want to use a better stopword list, such as this one.

Once you have your Corpus object, tell gensim to generate tf-idf scores for you like so.

3) Do LSI topic modeling. You can apply LSI to the tf-idf vectors, like so. You will have to supply the number of topics to generate. Figuring out a good number is part of the assignment. Print out the resulting topics, each topic as a lists of word coefficients. Now, sample ten topics randomly (not the first ten, a random ten!) from your set for closer analysis. Try to annotate each of these ten topics with a short descriptive name or phrase that captures what it is “about.” You will have to refer to the original documents that contain high proportions of that topic, and you will likely find that some topics have no clear concept.

Turn in: your annotated topics plus a comment on how well you feel each “topic” captured a real human concept.

4) Now do LDA topic modeling. Repeat the exercise of step 3 but with LDA instead, again trying to annotate ten randomly sampled topics. What is different? Did it better capture the meaning of the documents? If so, “better” how?

Turn in: your annotated topics plus a comment on how LDA differed from LSI.

5) Now apply LDA to the State of the Union. Download the source data file state-of-the-union.csv. This is a standard CSV file with one speech per row. There are two columns: the year of the speech, and the text of the speech. You will write a Python program that reads this file and turns it into TF-IDF document vectors, then prints out some information. Here is how to read a CSV in Python. You may need to add the line

csv.field_size_limit(1000000000)

to the top of your program to be able to read this large file. Also, this file is probably too big to open in Excel or to read with Pandas.

The file is a csv with columns year, text. Note: there are some years where there was more than one speech! Design your data structures accordingly.

6) Determine how speeches have changed over the 20th century. We’ll use a very simple algorithm:

• Generate TF-IDF vectors for the entire corpus
• Group speeches by decade
• Sum TF-IDF vectors for all speeches in each decade to get one “summary” vector per decade
• Print out top 10 most highly ranked words in each decade vector

Turn in: an analysis of how the topics of the State of the Union have changed over the decades of the 20th century. What patterns do you see? Can you connect the terms to major historical events? (wars, the great depression, assassinations, the civil rights movement, Watergate…)

взять займ безработному. For this assignment you will evaluate the performance of OpenCalais, a commercial entity extraction service, against your hand annotations.

1. Pick five random news stories and hand-annotate them. Pick an English-language news site with many stories on the home page, or a section of such a site (business, sports, etc.) Then generate five random numbers from 1 to the number of stories on the page. Cut and paste the text of each article into a separate  file, and save as plain text (no HTML, no formatting.)

2. Detect entities by hand in each article. Paste the text of each article into an RTF or Word document and go through it, underlining every entity. Count every mention of a person, place, or organization, including alternate names (“the president”) and pronoun references. Count how many entity references appear in each document (multiple mentions of the same entity all count.)

3. Now run each document through OpenCalais. You can paste the text into the demo page here. Now count:

• Your hand count of entity references
• Correct entities: entities found by OC and marked by you
• Incorrect entities: entities found by OC but not marked by you
• Missed entities: entities marked by you but not found by OC

4. Turn in:

• Your hand-marked documents.
• A spreadsheet. Please turn in a spread sheet with one row per document, and four columns: hand-labelled, correct, incorrect, missed
• Make sure your spreadsheet includes totals of these four columns across all documents
• Your analysis. Report on any patterns in the failures that your see. Where is OpenCalais most accurate? Where is it least accurate? Are there predictable patterns to the errors? Are there ambiguities as to what is really an entity?

For this exercise, you will break into three groups. Two will analyze data, and the third will do legal research. The data for this assignment is available here.

All groups: How can we encode our ideas about racial fairness into a quantitative metric? That is the fundamental question underlying this assignment and you must answer it. Don’t build models that give you answers to useless questions; each model you build must embody some justifiable concept of fairness. Many such metrics have been proposed; part of the assignment is researching and evaluating them. I’ve proposed some models that might be interesting, but I’ll be just has happy — perhaps happier — to have you tell me why these particular model formulations will not yield an interesting result. Also, you have to tell me what your modeling results mean. Is there bias? In what way, how significant is it, and what are alternate explanations? Uninterpreted results will not get a passing grade.

Group 1: Analyze the stop and frisk data using a multi-level linear model

This group will analyze the data along the lines of Gelman 2006. However, that paper used Bayesian estimation whereas this team will used standard linear regression. Use RStudio with the lme4 package, as described in this tutorial.

As above, you need to choose and justify quantitative definitions of fairness. Should you look at stop rate? Hit rate? And should you compare to the racial composition of each precinct? Or per-race crime rates? And if so, which crimes?

Each conceptual definition of fairness can be embodied in many different statistical models. Compare at least two different statistical formulations. For example you might end up modeling:

• hit rate vs race, per precinct
• hit rate vs race, per precinct, with precinct random effects

You must also choose the unit of analysis. Is precinct or census block a better unit? Why? Or you could compare the same model on different units.

Your final post must include a justification of the metric and model choices, and useful visualization of the results, and interpretation of both the regression coefficients and uncertainty measures (standard errors on regression coefficients, and modeling residuals.)

Group 2: Analyze the stop and frisk data using a Bayesian model

This group will analyze the data along the lines of Simoiu 2016, by adapting their code. For a tutorial on how to set up Bayesian modeling in R, see Bayesian Linear Mixed Models using STAN.

As with group 2, you must research and choose a definition of fairness, fit at least two different statistical formulations, and interpret your results including the uncertainty (posterior distributions and residuals.) Be sure to visualize your model fit, as in figure 7 of Simoiu.

I would be happy to see you replicate the threshold test of Simoiu. However, I want you to explain why the threshold test makes sense as a fairness metric. If it doesn’t make sense I want you to design a new model. Perhaps the assumption that each officer can make correctly calibrated estimates of the probability that someone is carrying contraband is unrealistic, and your model should be based on the idea that the estimates are biased and try to model that bias as latent variables.

Group 3: Research legal and policy precedent for statistical tests

This group will scour the legal literature to determine what sorts of statistical tests have been used, or could be used, to answer legal questions of discrimination. You will also research the related policy literature: what sorts of tests have governments, companies, schools etc. used to evaluate the presence or significance of discrimination. For an entry into the literature, you could do worse than to start with the works referenced by Big Data’s Disparate Impact.

This group will not be coding, but I expect you to ask not only what fairness metrics might be appropriate (as the other groups must also do) but 1) whether or not these metrics might hold up in court and 2) whether they have ever been used outside of court.

And one particular question I would like you to answer: would Simiou’s “threshold test” have legal or policy relevance?

Due Friday Nov 10

For this assignment you will design a hybrid filtering algorithm. You will not implement it, but you will explain your design criteria and provide a filtering algorithm in sufficient technical detail to convince me that it might actually work — including psuedocode.

1. Decide who your users are. Journalists? Professionals? General consumers? Someone else?

2. Decide what you will filter. You can choose:

• Social network updates, like the Facebook news feed or Twitter trending topics
• A news organization recommendation engine
• The whole web, like Google News
• something else

3. List all available information that you have available as input to your algorithm. If you want to filter Facebook or Twitter, you may pretend that you are the company running the service, and have access to all posts and user data — from every user. You also also assume you have a web crawler or a firehose of every RSS feed or whatever you like, but you must be specific and realistic about what data you are operating with.

4. Argue for the design factors that you would like to influence the filtering, in terms of what is desirable to the user, what is desirable to the publisher (e.g. Facebook or a news organization), and what is desirable socially. Explain as concretely as possible how each of these (probably conflicting) goals might be achieved through in software. Since this is a hybrid filter, you can also design social software that asks the user for certain types of information (e.g. likes, votes, ratings) or encourages users to act in certain ways (e.g. following) that generate data for you.

5. Write psuedo-code for a function that produces a “top stories” list. This function will be called whenever the user loads your page or opens your app, so it must be fast and frequently updated. You can assume that there are background processes operating on your servers if you like. Your psuedo-code does not have to be executable, but it must be specific and unambiguous, such that I could actually go and implement it. You can assume that you have libraries for classic text analysis and machine learning algorithms. So, you don’t have to spell out algorithms like TF-IDF or item-based collaborative filtering, or anything else you can dig up in the research literature, but simply say how you’re going to use such building blocks. If you use an algorithm we haven’t discussed in class, be sure to provide a reference to it.

6. Write up steps 1-5. The result should be no more than three pages. You must be specific and plausible. You must be clear about what you are trying to accomplish, what your algorithm is, and why you believe your algorithm meets your design goals (though of course it’s impossible to know for sure without testing; but I want something that looks good enough to be worth trying.)

Due before class, October 13

займ без процентов на карту. This assignment is designed to help you develop a feel for the way topic modeling works, the connection to the human meanings of documents, and common ways of handling a time dimension. You will analyze the State of the Union speeches corpus, and report on how the subjects have shifted over time in relation to historical events.

1. Download the source data file state-of-the-union.csv. This is a standard CSV file with one speech per row. There are two columns: the year of the speech, and the text of the speech. You will write a Python program that reads this file and turns it into TF-IDF document vectors, then prints out some information. Here is how to read a CSV in Python. You may need to add the line

csv.field_size_limit(1000000000)

to the top of your program to be able to read this large file.

The file is a csv with columns year, text. Note: there are some years where there was more than one speech! Design your data structures accordingly.

2) Feed the data into gensim. Now you need to load the documents into Python and feed them into the gensim package to generate tf-idf weighted document vectors. Check out the gensim example code here. You will need to go through the file twice: once to generate the dictionary (the code snippet starting with “collect statistics about all tokens”) and then again to convert each document to what gensim calls the bag-of-words representation, which is un-normalized term frequency (the code snippet starting with “class MyCorpus(object)”

Note that there is implicitly another step here, which is to tokenize the document text into individual word features — not as straightforward in practice as it seems at first, but the example code just does the simplest, stupidest thing, which is to lowercase the string and split on spaces. You may want to use a better stopword list, such as this one.

Once you have your Corpus object, tell gensim to generate tf-idf scores for you like so.

3) Do LSI topic modeling. You can apply LSI to the tf-idf vectors, like so. You will have to supply the number of topics to generate. Figuring out a good number is part of the assignment. Print out the resulting topics, each topic as a lists of word coefficients. Now, sample ten topics randomly from your set for closer analysis. Try to annotate each of these ten topics with a short descriptive name or phrase that captures what it is “about.” You will likely have to refer to the original documents that contain high proportions of that topic, and you will likely find that some topics have no clear concept.

Turn in: your annotated topics plus a comment on how well you feel each “topic” captured a real human concept.

4) Now do LDA topic modeling. Repeat the exercise of step 3 but with LDA instead, again trying to annotate ten randomly sampled topics. What is different?

Turn in: your annotated topics plus a comment on how LDA differed from LSI.

5) Now choose one of the following exercises:

a) Come up with a method to figure out how topics of speeches have changed over time. The goal is to summarize changes in the State of the Union speech in each decade of the 20th and 21st century. There are many different ways to use topic modeling to do this. Possibilities include: visualizations, grouping speeches by decade after topic modeling, and grouping speeches by decade before topic modeling. You can base your algorithm on either LSI or LDA, whichever you feel gives the most insight. Choose a method, then use your decade summarization algorithm to understand what the content of speeches was in each decade.

Turn in: a description of your decade summarization algorithm, and an analysis of how the topics of the State of the Union have changed over the decades of the 20th century. What patterns do you see? Can you connect the terms to major historical events? (wars, the great depression, assassinations, the civil rights movement, Watergate…)

or b) Analyze a different document set. Try LDA on a different document set, this collection of AP wire stories. Repeat the process of choosing the number of topics, fitting a model, and interpreting a random sample of 10 of them. Are the topics any clearer on this document set? If so, why? You may wish to look at previous LDA results on these documents, the top 20 words from 100 topics.

Turn in: your annotated topic sample, plus a description of the differences between the output on these documents vs. the State of the Union documents. Does one work better than the other? If so, define “better.”

This assignment is due Friday, September 29 at 10:00 AM. You may email me the results.

For this assignment you will analyze a social network using three different centrality algorithms, and compare the results.

1. Download and install Gephi, a free graph analysis package. It is open source and runs on any OS.

2. Download the data file lesmis.gml from the UCI Network Data Repository.  This is a network extracted from the famous French novel Les Miserables — you may also be familiar with the musical and the recent movie. Each node is a character, and there is an edge between two characters if they appear in the same chapter. Les Miserables is written in over 300 short chapters, so two characters that appear in the same chapter are very likely to meet or talk in the plot of the book. The edges are weighted, and the weight is the number of chapters those characters appear together in.

3. Open this file in Gephi, by choosing File->Open. When the dialog box comes up, set the “Graph Type” type to “Undirected.” The graph will be plotted. What do you see? Can you discern any patterns?

4. Now arrange the nodes in a nicer way, by choosing the “Force Atlas 2″ layout algorithm from the Layout menu at left and pressing the “Run” button. When things settle down, hit the “Stop” button. The graph will be arranged nicely, but it will be quite small.  You can zoom in using the mouse wheel (or two fingers on the trackpad on a mac) and pan using the right mouse button.

5. Select the “Edit” tool from the bottom of the toolbar on the left. It looks like a mouse pointer with question mark next to it:

6. Now you can click on any node to see its label, which is the name of the character it represents. This information will appear in the “Edit” menu in the upper left. Here’s the information for the character Gavroche.

Click around the various nodes in the graph. Which characters have been given the most central locations? If you are familiar with the story of Les Miserables, how does this correspond to the plot? Are the most central nodes the most important characters?

7. Make Gephi color nodes by degree. Choose the “Ranking” tab from panel at the upper left, then select the “Nodes” tab, then “Degree” from the drop-down menu. Press the “Apply” button.

Now the nodes with the highest degree will be darker. Do these high degree nodes correspond to the nodes that the layout algorithm put in the center? Are they the main characters in the story?

8. Now make Gephi compute betweenness and closeness centrality by pressing the “Run” button for the Network Diameter option under “Network Overview” in to the right of the screen.

You will get a report with some graphs. Just click “Close”. Now betweenness and closeness centrality will appear in the drop-down under “Ranking,” in the same place where you selected degree centrality earlier, and you can assign colors based on either run by clicking the “Apply” button.

Also, the numerical values for betweenness centrality and closeness centrality will now appear in the “Edit” window for each node.

Select “Betweenness Centrality” from the drop-down meny and hit “Apply.” What do you see? Which characters are marked as important? How does it differ from the characters which are marked as important by degree?

Now select “Closeness Centrality” and hit “Apply.” (Note that this metric uses a scale which is the reverse of the others — closeness measures average distance to all other nodes, so small values indicate more central nodes. You may want to swap the black and white endpoints of the color scale to get something which is comparable to the other visualizations.) How does closeness centrality differ from betweeness centrality and degree? Which characters differ between closeness and the other metrics?

9. Which centrality algorithm would you prefer to use to understand the structure of Les Miserables? Why? How would you validate your choice if you didn’t already know the story? That is the situation a journalist is in when they analyze unknown data.

10. Turn in: your answers to the questions in steps 3, 6, 7, 8 and 9, plus screenshots for the graph plotted with degree, betweenness centrality, and closeness centrality. (To take a screenshot: on Windows, use the Snipping Tool. On Mac, press ⌘ Cmd + ⇧ Shift + 4. If you’re on Linux, you get to tell me)

What I am interested in here is how the values computed by the different algorithms correspond to the plot of Les Miserables (if you are familiar with it), and how they compare to each other. Telling me that “Jean Valjean has a closeness centrality of X” is not a high-enough level interpretation — your couldn’t publish that in a finished story, because your readers won’t know what that means.

Due before class on Friday, December 2 

For this assignment you will evaluate the performance of OpenCalais, a commercial entity extraction service, against your hand annotations.

1. Pick five random news stories and hand-annotate them. Pick an English-language news site with many stories on the home page, or a section of such a site (business, sports, etc.) Then generate five random numbers from 1 to the number of stories on the page. Cut and paste the text of each article into a separate  file, and save as plain text (no HTML, no formatting.)

2. Detect entities by hand in each article. Paste the text of each article into an RTF or Word document and go through it, underlining every entity. Count every mention of a person, place, or organization, including alternate names (“the president”) and pronoun references. Count how many entities appear in each document.

3. Now run each document through OpenCalais. You can paste the text into the demo page here. Now count:

• Correct entities: entities found by OC and marked by you
• Incorrect entities: entities found by OC but not marked by you
• Missed entities: entities marked by you but not found by OC

4. Turn in:

• Your hand-marked documents.
• A spreadsheet. Please turn in a spread sheet with one row per document, and four columns: hand-labelled, correct, incorrect, missed
• Make sure your spreadsheet includes totals of these four columns across all documents
• Your analysis. Report on any patterns in the failures that your see. Where is OpenCalais most accurate? Where is it least accurate? Are there predictable patterns to the errors? Are there ambiguities as to what is really an entity?