As I begin working on my project involving Taiyō magazine, I thought I’d document what I’m doing so others can see the process of cleaning the data I’ve gotten, and then experimenting with it. This is the first part in that series: first steps with data, cleaning it, and getting it ready for analysis. If I have the Taiyō data in “plain text,” what’s there to clean? Oh, you have no idea.
I just got off the phone with a researcher this morning who is interested in looking at sentiment analysis on a corpus of fiction, specifically by having some native speakers of Japanese (I think) tag adjectives as positive or negative, then look at the overall shape of the corpus with those tags in mind.
A while back, I wrote a paper about geoparsing and sentiment analysis for a class, describing a project I worked on. Talking to this researcher made me think back to this project – which I’m actually currently trying to rewrite in Python and then make work on some Japanese, rather than Victorian English, texts – and my own definition of sentiment analysis for humanistic inquiry.*
How is my definition of sentiment analysis different? How about I start with the methodology? What I did was look for salient adjectives, which I searched for by looking at most “salient” nouns (not necessarily the most frequent, but I need to refine my heuristics) and then the adjectives that appeared next to them. I also used Wordnet to look for words related to these adjectives and nouns to expand my search beyond just those specific words to ones with similar meaning that I might have missed (in particular, I looked at hypernyms (broader terms) and synonyms of nouns, and synonyms of adjectives).
My method of sentiment analysis ends up looking more like automatic summarization than a positive-negative sentiment analysis we more frequently encounter, even in humanistic work such as Matt Jockers’s recent research. I argue, of course, that my method is somewhat more meaningful. I consider all adjectives to be sentiment words, because they carry subjective judgment (even something that’s kind of green might be described by someone else as also kind of blue). And I’m more interested in the character of subjective judgment than whether it should be able to be considered ‘objectively’ as positive or negative (something I don’t think is really possible in humanistic inquiry, and even in business applications). In other words, if we have to pick out the most representative feelings of people about what they’re experiencing, what are they feeling about that experience?
After all, can you really say that weather is good or bad, that there being a lot of farm fields is good or bad? I looked at 19th-century British women’s travel narratives of “exotic” places, and I found that their sentiment was often just observations about trains and the landscape and the people. They didn’t talk about whether they were feeling positively or negatively about those things; rather, they gave us their subjective judgment of what those things were like.
My take on sentiment analysis, then, is clearly that we need to introduce human judgment to the end of the process, perhaps gathering these representative phrases and adjectives (I lean toward phrases or even whole sentences) and then deciding what we can about them. I don’t even think a human interlocutor could put down a verdict of positive or negative on these observations and judgments – sentiments – that the women had about their experiences and environments. If not even a human could do it, and humans write and train the algorithms, how can the computer do it?
Is there even a point? Does it matter if it’s possible or not? We should be looking for something else entirely.
(I really need to get cracking on this project. Stay tuned for the revised methodology and heuristics, because I hope to write more and share code here as I go along.)
* I’m also trying to write a more extensive and revised paper on this, meant for the new incarnation of LLC.
I recently wrote a small script to perform a couple of functions for pre-processing Aozora Bunko texts (text files of public domain, modern Japanese literature and non-fiction) to be used with Western-oriented text analysis tools, such as Voyant, other TAPoR tools, and MALLET. Whereas Japanese text analysis software focuses largely on linguistics (tagging parts of speech, lemmatizing, etc.), Western tools open up possibilities for visualization, concordances, topic modeling, and other various modes of analysis.
Why do these Aozora texts need to be processed? Well, a couple of issues.
- They contain ruby, which are basically glosses of Chinese characters that give their pronunciation. These can be straightforward pronunciation help, or actually different words that give added meaning and context. While I have my issues with removing ruby, it’s impossible to do straightforward tool-based analysis without removing it, and many people who want to do this kind of analysis want it to be removed.
- The Aozora files are not exactly plain text: they’re HTML. The HTML tags and Aozora metadata (telling where the text came from, for example) need to be removed before analysis can be performed.
- There are no spaces between words in Japanese, but Western text analysis tools identify words by looking at where there are spaces. Without inserting spaces, it looks like each line is one big word. So I needed to insert spaces between the Japanese words.
How did I do it? My approach, because of my background and expertise, was to create a Python script that used a couple of helpful libraries, including BeautifulSoup for ruby removal based on HTML tags, and TinySegmenter for inserting spaces between words. My script requires you to have these packages installed, but it’s not a big deal to do so. You then run the script in a command line prompt. The way it works is to look for all .html files in a directory, load them and run the pre-processing, then output each processed file with the same filename, .txt ending, a plain text UTF-8 encoded file.
The first step in the script is to remove the ruby. Helpfully, the ruby is contained in several HTML tags. I had BeautifulSoup traverse the file and remove all elements contained within these tags; it removes both the tags and content.
Next, I used a very simple regular expression to remove everything in brackets – i.e. the HTML tags. This is kind of quick and dirty, and won’t work on every file in the universe, but in Aozora texts everything inside a bracket is an HTML tag, so it’s not a problem here.
Finally, I used TinySegmenter on the resulting HTML-free text to split the text into words. Luckily for me, it returns an array of words – basically, each word is a separate element in a list like [‘word1’, ‘word2’, … ‘wordn’] for n words. This makes my life easy for two reasons. First, I simply joined the array with a space between each word, creating one long string (the outputted text) with spaces between each element in the array (words). Second, it made it easy to just remove the part of the array that contains Aozora metadata before creating that string. Again, this is quick and dirty, but from examining the files I noted that the metadata always comes at the end of the file and begins with the word 底本 (‘source text’). Remove that word and everything after it, and then you have a metadata-free file.
Write this resulting text into a plain text file, and you have a non-ruby, non-HTML, metadata-free, whitespace-delimited Aozora text! Although you have to still download all the Aozora files individually and then do what you will with the resulting individual text files, it’s an easy way to pre-process this text and get it ready for tool-based (and also your-own-program-based) text analysis.
I plan to put the script on GitHub for your perusal and use (and of course modification) but for now, check it out on my Japanese Text Analysis research guide at Penn.
I’ve been looking (okay, not looking, wishing) for a Japanese tokenizer for a while now, and today I decided to sit down and do some research into what’s out there. It didn’t take long – things have improved recently.
First off – so what is tokenization? Basically, it’s separating sentences by words, or documents by sentences, or any text by some unit, to be able to chunk that text into parts and analyze them (or do other things with them). When you tokenize a document by word, like a web page, you enable searching: this is how Google finds individual words in documents. You can also find keywords from a document this way, by writing an algorithm to choose the most meaningful nouns, for example. It’s also the first step in more involved linguistic analysis like part-of-speech tagging (thing, marking individual words as nouns, verbs, and so on) and lemmatizing (paring words down to their stems, such as removing plural markers and un-conjugating verbs).
This gives you a taste of why tokenization is so fundamental and important for text analysis. It’s what lets you break up an otherwise unintelligible (to the computer) string of characters into units that the computer can attempt to analyze. It can index them, search them, categorize them, group them, visualize them, and so on. Without this, you’re stuck with “words” that are entire sentences or documents, that the computer thinks are individual units based on the fact that they’re one long string of characters.
Usually, the way you tokenize is to break up “words” based on spaces (or sentences based on punctuation rules, etc., although that doesn’t always work). (I put “words” in quotes because you can really make any kind of unit you want, the computer doesn’t understand what words are, and in the end it doesn’t matter. I’m using “words” as an example here.) However, for languages like Japanese and Chinese (and to a lesser extent Korean) that don’t use spaces to delimit all words (for example, in Korean particles are attached to nouns with no space in between, like saying “athome” instead of “at home”), you run into problems quickly. How to break up texts into words when there’s no easy way to distinguish between them?
The question of tokenizing Japanese may be a linguistic debate. I don’t know enough about linguistics to begin to participate in it, if it is. But I’ll quickly say that you can break up Japanese based on linguistic rules and dictionary rules – understanding which character compounds are nouns, which verb conjugations go with which verb stems (as opposed to being particles in between words), then breaking up common particles into their own units. This appears to be how these tools are doing it. For my own purposes, I’m not as interested in linguistic patterns as I am in noun and verb usage (the meaning rather than the kind) so linguistic nitpicking won’t be my area anyway.
Moving on to the tools. I put them through the wringer: Higuchi Ichiyō’s Ame no yoru, the first two lines, from Aozora bunko.
One, kuromoji, is the tokenizer behind Solr and Lucene. It does a fairly good job, although with Ichiyō’s uncommon word usage and conjugation, it faltered and couldn’t figure out that 高やか is one word; rather it divided it into 高 や か. It gives the base form, reading, and pronunciation, but nothing else. However, in the version that ships with Solr/Lucene, it lemmatizes. Would that ever make me happy. (That’s, again, reducing a word to its base form, making it easy to count all instances of both “people” and “person” for example, if you’re just after meaning.) I would kill for this feature to be integrated with the below tool.
The other, U-Tokenizer, did significantly better, but its major drawback is that it’s done in the form of an HTTP request, meaning that you can’t put in entire documents (well, maybe you could? how much can you pass in an HTTP request?). If it were downloadable code with an API, I would be very happy (kuromoji is downloadable and has a command line interface). U-Tokenizer figured out that 高やか is one word, and also provides a list of “keywords,” which as far as I can tell is a bunch of salient nouns. I used it for a very short piece of text, so I can’t comment on how many keywords it would come up with for an entire document. The documentation on this is sparse, and it’s not open source, so it’s impossible to know what it’s doing. Still, it’s a fantastic tool, and also seems to work decently for Chinese and Korean.
Each of these tools has its strengths, and both are quite usable for modern and contemporary Japanese. (I really was cruel to feed them Ichiyō.) However, there is a major trial involved in using them with freely-available corpora like Aozora bunko. Guess what? Preprocessing ruby.
Aozora texts contain ruby marked up within the documents. I have my issues with stripping out ruby from documents that heavily use them (like Meiji writers, for example) because they add so much meaning to the text, but let’s say for argument’s sake that we’re not interested in the ruby. Now, it’s time to cut it all out. If I were a regular expressions wizard (or even had basic competency with them) I could probably strip this out easily, but it’s still time consuming. Download text, strip out ruby and other metadata, save as plain text. (Aozora texts are XHTML, NOT “plain text” as they’re often touted to be.) Repeat. For topic modeling using a tool like MALLET, you’re going to want to have hundreds of documents at the end of it. For example, you might be downloading all Meiji novels from Aozora and dividing them into chunks or chapters. Even the complete works of Natsume Sōseki aren’t enough without cutting them down into chapters or even paragraphs to make enough documents to use a topic modeling tool effectively. Possibly, run all these through a part-of-speech tagger like KH Coder. This is going to take a significant amount of time.
Then again, preprocessing is an essential and extremely time-consuming part of almost any text analysis project. I went through a moderate amount of work just removing Project Gutenberg metadata and dividing into chapters a set of travel narratives that I downloaded in plain text, thankfully not in HTML or XML. It made for easy processing. With something that’s not already real plain text, with a lot of metadata, and with a lot of ruby, it’s going to take much more time and effort, which is more typical of a project like this. The digital humanities are a lot of manual labor, despite the glamorous image and the idea that computers can do a lot of manual labor for us. They are a little finicky with what they’ll accept. (Granted, I’ll be using a computer script to strip out the XHTML and ruby tags, but it’s going to take work for me to write it in the first place.)
In conclusion? Text analysis, despite exciting available tools, is still hard and time consuming. There is a lot of potential here, but I also see myself going through some trials to get to the fun part, the experimentation. Still, stay tuned, especially for some follow-up posts on these tools and KH Coder as I become more familiar with them. And, I promise to stop being difficult and giving them Ichiyō’s Meiji-style bungo.