Dictionary-Based Text Analysis

A dictionary method applies a fixed, human-specified word list and needs no labeled examples — you decide in advance which words count. Supervised classification instead learns the mapping from text features to categories from hand-labeled training documents, so it can pick up context-dependent and non-obvious cues but requires costly labeled data. Dictionaries are more transparent and portable in principle but brittle across domains; classifiers are more accurate within the domain they are trained on but can fail to generalize. They are often combined, with a dictionary providing a baseline and a classifier improving on it. See supervised-text-classification.

You can, but you must validate it on your own data first. Lexicons such as the Lexicoder Sentiment Dictionary or LIWC were built and validated on specific kinds of text, and their word lists may misfire in a different domain, register, or time period — financial 'liability', legislative 'amendment', or domain jargon can carry meanings the dictionary did not anticipate. The responsible workflow is to hand-code a sample, compare the dictionary scores to those codes, inspect disagreements, and revise or reject the dictionary accordingly.

By default they do not. Simple term counting treats 'good' and 'not good' identically, and cannot tell 'tax relief' from 'tax burden'. Some implementations add negation-handling windows, multi-word expressions, or wildcard stems to mitigate this, but these are partial fixes that themselves need validation. When context, negation, and word order are central to the concept, dictionary methods are the wrong tool, and supervised classification or model-based approaches that use richer features are preferable.