4. Technical terminology extraction

Technical documents rarely express domain concepts as single tokens. A concept such as finite impulse response filter, maximum power point tracking, or pulse-width modulation is a multiword term (MWT): a sequence of tokens that functions as one lexical unit with a single, stable meaning. Alongside MWTs, technical prose is saturated with abbreviations and acronyms whose definitions appear once, usually at first mention, and are then never repeated. A terminology extraction pipeline must recover both: the full forms that carry meaning and the short forms that carry the rest of the document.

This chapter builds a working extractor. We treat the task as three coupled subproblems — candidate generation, termhood scoring, and definition resolution — and we implement each with transparent, inspectable rules before reaching for statistical models.

Why Terminology Extraction Is Not Just Keyword Extraction

Keyword extraction ranks single tokens by frequency or TF-IDF. Terminology extraction differs in three ways that matter for technical documents:

• Boundary ambiguity. In low pass filter design, the term is low pass filter; design is a context word. In filter design, the term is filter design. Frequency alone cannot separate these.
• Non-compositionality. Bus in CAN bus has nothing to do with public transport. The term's meaning is not the sum of its parts.
• Definitional structure. Abbreviations are introduced by explicit patterns — Central Processing Unit (CPU), CPU (Central Processing Unit), CPU, hereinafter Central Processing Unit — and these patterns are the highest-precision signal available.

Candidate Generation: From Tokens to Term Candidates

The first stage over-generates deliberately. We extract every n-gram (n = 2…5) that matches a part-of-speech pattern typical of technical terms. The dominant patterns in English technical writing are:

POS tagging with a domain-adapted tagger is sufficient here. The key discipline is to not filter yet: a candidate that looks implausible in isolation may be a valid term in context. Filtering happens in the scoring stage, where we have more evidence.

Termhood Scoring: Deciding What Is Actually a Term

We score each candidate with three complementary measures. Let c be a candidate, f(c) its corpus frequency, and f(w) the frequency of its component words.

1. C-value rewards candidates that are frequent and not merely substrings of longer frequent terms. For a candidate c appearing in longer candidates t:

C-value(c) = log2(|c|) * f(c)                      if c is not nested
C-value(c) = log2(|c|) * (f(c) - mean(f(t)))       if c is nested in longer terms t

The nesting penalty is what prevents filter from outranking finite impulse response filter when the longer term is the real concept.

2. PMI (pointwise mutual information) measures how tightly the components co-occur relative to chance:

PMI(w1, w2) = log2( P(w1, w2) / (P(w1) * P(w2)) )

High PMI indicates a fixed expression. For a three-word term, use the minimum PMI over adjacent pairs — the weakest link governs.

3. Domain relevance compares a candidate's frequency in the target corpus against a general-purpose reference corpus. A term like impedance matching is rare in general English but frequent in electronics; the ratio is a strong termhood signal.

Combine the three scores with weights tuned on a small labeled set. In practice, a weighted sum with C-value dominant (weight ≈ 0.5), PMI ≈ 0.3, and domain relevance ≈ 0.2 works well as a starting point. Report the top-N candidates for human review rather than trusting a single threshold.

Resolving Abbreviations and Acronym Definitions

Abbreviation resolution is a pattern-matching problem with a verification step. The canonical approach (Schwartz & Hearst) is:

1. Find candidate pairs. Scan for Long Form (SHORT) and SHORT (Long Form) using a regex that allows 2–6 tokens in the long form and 2–10 characters in the short form.
2. Verify character alignment. The short form's characters must appear in order within the long form, starting at a word boundary. For Central Processing Unit → CPU, the alignment is C-P-U at word starts. For maximum power point tracking → MPPT, same rule.
3. Handle exceptions. Some acronyms embed a letter from mid-word (Light Amplification by Stimulated Emission of Radiation → LASER). Allow a small edit distance or a relaxed alignment that permits one skipped character.

Once a definition is found, propagate it: every subsequent occurrence of the short form in the document inherits the long form. This is what makes abbreviation resolution valuable — it converts a document full of opaque tokens into one where every mention is grounded.

Worked Example: A Small Technical Corpus

Consider three sentences from a signal-processing document:

S1: The finite impulse response filter was designed using a Kaiser window.
S2: A finite impulse response filter has a linear phase response.
S3: The FIR filter coefficients were quantized to 16 bits.

Candidate generation yields finite impulse response, impulse response filter, finite impulse response filter, linear phase response, Kaiser window, FIR filter, and others. C-value penalizes finite impulse response because it is nested inside the longer finite impulse response filter, which appears twice. PMI confirms finite impulse response as a tight unit. The definition detector finds no parenthetical pattern in S1–S3, but if S1 read finite impulse response (FIR) filter, the alignment check would link FIR to finite impulse response, and S3's FIR filter would resolve to the full term.

Evaluation and Failure Modes

Evaluate the extractor against a gold standard of terms and definitions annotated by a domain expert. Report precision, recall, and F1 separately for MWT extraction and for abbreviation resolution — they fail differently and should not be averaged into one number.

Common failure modes and their fixes:

• Over-generation of generic phrases. the proposed method and experimental results score high on frequency. Filter with a stopword list of academic boilerplate and a domain-relevance floor.
• Missed discontinuous terms. filter ... with linear phase splits the term. Allow a bounded window (e.g., 5 tokens) when matching multiword patterns.
• False abbreviation links. the (see) appendix can trigger a spurious match. Require the short form to be 2–10 uppercase characters and the long form to contain at least two words.
• Nested term confusion. Both impulse response and finite impulse response filter are valid terms at different granularities. Keep both, but rank the longer one higher when it is the concept actually used in the document's claims.

What to Carry Forward

The extractor you build here becomes the vocabulary layer for everything that follows: document indexing, cross-reference resolution, and the construction of domain glossaries. The three-stage structure — generate candidates broadly, score them with complementary measures, then resolve definitions by pattern and alignment — is robust because each stage compensates for the weaknesses of the others. C-value handles nesting, PMI handles cohesion, domain relevance handles specificity, and definition resolution handles the abbreviation problem that frequency-based methods cannot see at all.