After my talk at Columbia, a grad student asked me “Why do you hate CRFs?”. This is a tough question to answer because of the failed presupposition, which assumes I hate CRFs and asks me to explain why. I want to defeat the question by saying that I don’t hate CRFs. The right question is:
Why isn’t there a CRF package in LingPipe?
This is a good question because if you look at any of the recent tagging/chunking bakeoffs (CoNLL, Biocreative, etc.) you’ll see that the top-scoring systems are CRFs or similar richly featured conditional models.
The short, though highly idiomatic, answer is “horses for courses”. (Etymology hint: racing horses specialize in muddy or other conditions, as explained in Abbott and Costello’s classic “Mudder and Fodder” sketch).
To put this in context, we’re looking for high recall taggers that have a decent precision at very high recall. Our chunk-level forward-backward HMM tagger achieves 99.95% recall at 8% precision, even though it’s a whopping 15% behind the best systems in first-best performance.
There are three main problems with using CRFs for the kinds of problems in which we’re interested:
- CRFs need fairly extensive feature engineering to outperform a good HMM baseline,
- large feature sets and discriminitive training can lead to very high model complexity and variance, and
- these features and training regimes make estimation (training) and inference (decoding/tagging) computationally intensive.
1. Portability: Unfortunately, the performance of CRFs is directly attributable to their large sets of hand-tuned features. Consider the following quote from (Finkel, Dingare, Nguyen, Nissim, Manning and Sinclair 2004):
Using the set of features designed for that task in CoNLL 2003 , our system achieves an f-score of 0.76 on the BioCreative development data, a dramatic ten points lower than its f-score of 0.86 on the CoNLL newswire data. Despite the massive size of the final feature set (almost twice as many features as used for CoNLL), its final performance of 0.83 is still below its performance on the CoNLL data
The baseline features are listed in section 3 (below). Surprisingly, LingPipe’s HMM performed similarly to baseline CRFs out of the box on BioCreative data. Finkel et al.’s paper describes the wealth of features they applied to port their CRF system from CoNLL to biocreative. They also describe heuristics, such as mismatched paren balancing, pruning given various known suffixes, etc.
2. Bias-Variance Tradeoff: What you see in heavily tuned conditional models such as CRFs is a strongly attenuated probability distribution. Such models are usually much more confident in their decisions than they should be. This overconfidence stems from a lack of modeling of dependencies, and shows up in our models, too, especially in classification tasks, where we’re surprised every time the document mentions “baseball”. This is why speech recognition is so bad: the acoustic context-dependent phoneme mixture models are surprised every 1/100th of a second that the speaker’s still using a Texas accent. Interestingly, 2005’s winning Spam detection entry for TREC (Bratko and Filipic 2005) mitigated this problem in an interesting way by learning on the doc being classified. Topic-level models (such as LDA) and more highly dispersed frequency models (such as zero-inflated, hierarchal or simple over-dispersed models) are interesting approaches to this problem. There’s also a nice discussion of this attenuation problem in section 1.2.3 of a classification overview paper by Nigam, McCallum and Mitchell (2006):
"The faulty word independence assumption exacerbates the tendency of naive Bayes to produce extreme (almost 0 or 1) class probability estimates. However, classification accuracy can be quite high even when these estimates are inappropriately extreme."
In statistical terms, CRFs and other conditional models are often unbiased, but high variance. As usual, this arises from high model complexity (large numbers of parameters with free structure). That is, small changes in the training data or test samples leads to wide differences in models. Our language-model based approaches, on the other hand, tend to have lower variance and higher bias. They’re not nearly as sensitive to training data, but they tend to have built-in biases. Another way to say this is that the CRFs are much more tightly fit to their training data than the language-model based generative approaches we’ve adopted in LingPipe.
Many models, including support vector machines, boosting, perceptrons and even active learning, are explicitly designed to bias the global statistics in such a way as to emphasize the cases near the decision boundary. While this helps with first-best decisions, it tends to hurt any confidence-based decisions.
We see this same bias/variance tradeoff within the models supplied by LingPipe. Our rescoring chunkers are nearly useless for confidence based extraction due to their overly attenuated models.
3. Efficiency: The current algorithms for training CRFs are slow. As in hours if not days of CPU time. Run time is much better, but still slow. For instance, Finkel et al. (2004) report beam-based decoding speeds of roughly 10KB/second. Ben Wellner’s Carafe: CRFs for IE implementation (in Objective ML, of all things!) is reportedly very fast — around 50K words/sec, with beam-based pruning, feature caching and bigram sequence stats. So it’s clear that careful engineering of features with pruning and caching can be very useful. In fact, this result’s surprising enough it’s making me want to do more exploration of pruning in our own first-best decoders (remember, “horses for courses”).
The problem with pruning in a high recall tagger is that it explicitly eliminates hypotheses for which the model estimates will have lower likelihood than the first-best hypotheses. For high recall applications, these hypotheses are relevant.
The reason training and decoding is fairly slow is simple: lots of features. For decoding, time is pretty much wholly dependent on the number of features looked up per character or token of input. Because even our models are larger than existant CPU caches and because the models are accessed randomly, the time is mostly determined by front-side bus speed, which determines how fast data shuttles between memory and the outermost cache of the CPU.
Here’s a list of “baseline” CRF features from (Krishnan and Manning 2006):
Our baseline CRF is a sequence model in which labels for tokens directly depend only on the labels corresponding to the previous and next tokens. We use features that have been shown to be effective in NER, namely the current, previous and next words, character n-grams of the current word, Part of Speech tag of the current word and surrounding words, the shallow parse chunk of the current word, shape of the current word, the surrounding word shape sequence, the presence of a word in a left window of size 5 around the current word and the presence of a word in a left window of size 5 around the current word. This gives us a competitive baseline CRF using local information alone, whose performance is close to the best published local CRF models, for Named Entity Recognition
That is a lot of features. LingPipe’s HMM feature set includes exactly two components to predict the tag for the current word: the current word itself and the tag of the previous word. In the interest of full disclosure, we estimate the word probabilities a character at a time using language models, which is where the work is in our decoder.
Theoretically, if CRFs provide much sharper estimates and much better tuning, they could turn out to be faster than the kind of simple HMMs we use for first-best decoding. This counterintuitive otucome arises in many search settings, where more expensive but tighther pruning leads to overall performance improvements.